The future is increasingly arriving with dates. AI 2027, Citrini’s The 2028 Global Intelligence Crisis, Europe AI 2031: These time-stamped narratives are the “charismatic mega-fauna” of the current AI discourse. Through specific, singular future scenarios, they sharpened disagreement and exposed assumptions about the future direction of AI. Yet, that same specificity also hardened into polarizing debates about whether a given event will arrive on schedule, crowding out the many other ways an agentic future might unfold.
At Salesforce Futures, we closely monitor these developments with a practitioner’s eye. Our internal work supports the company’s strategic choices, so we know firsthand the value of predictive clarity. The closer one gets to a specific decision maker making specific choices, the more attainable that type of clarity becomes. The work commonly begins with initial “learning scenarios”: coherent alternative environments through which choices can be tested. Then, having developed a frame within which to explore uncertainty and gather signals, the work can move on to identifying deeper predetermined outcomes or commonalities across scenarios. The end point is a specific judgment and course of action that either aligns with the Predetermined — a scenario-planning term for a force already in the pipeline whose outcome holds in every scenario, so you can plan on it. or holds up no matter which future arrives.
However, in this, our first natively digital edition of Futures, we deliberately choose to emphasize the plurality of possibilities that the name implies. At this stage in the evolution of agentic enterprise futures, the useful work sits upstream from predictions. We offer instead a portfolio of framing questions, images, and vignettes meant to stretch what leaders can imagine as agentic capabilities develop and to put the business implications into situations concrete enough to picture.
WelcomeWhy Futures?
The future is increasingly arriving with dates. AI 2027, Citrini’s The 2028 Global Intelligence Crisis, Europe AI 2031: These time-stamped narratives are the “charismatic mega-fauna” of the current AI discourse. Through specific, singular future scenarios, they sharpened disagreement and exposed assumptions about the future direction of AI. Yet, that same specificity also hardened into polarizing debates about whether a given event will arrive on schedule, crowding out the many other ways an agentic future might unfold.
At Salesforce Futures, we closely monitor these developments with a practitioner’s eye. Our internal work supports the company’s strategic choices, so we know firsthand the value of predictive clarity. The closer one gets to a specific decision maker making specific choices, the more attainable that type of clarity becomes. The work commonly begins with initial “learning scenarios”: coherent alternative environments through which choices can be tested. Then, having developed a frame within which to explore uncertainty and gather signals, the work can move on to identifying deeper predetermined outcomes or commonalities across scenarios. The end point is a specific judgment and course of action that either aligns with the Predetermined — a scenario-planning term for a force already in the pipeline whose outcome holds in every scenario, so you can plan on it. or holds up no matter which future arrives.
However, in this, our first natively digital edition of Futures, we deliberately choose to emphasize the plurality of possibilities that the name implies. At this stage in the evolution of agentic enterprise futures, the useful work sits upstream from predictions. We offer instead a portfolio of framing questions, images, and vignettes meant to stretch what leaders can imagine as agentic capabilities develop and to put the business implications into situations concrete enough to picture.
Agentic enterprise futures
Our team’s work brings us into contact with dozens of Salesforce customers every quarter. We are constantly learning from the broader Salesforce ecosystem, from experts both inside and outside companies, and from emerging signals. Many companies are now making significant strides in their agentic journey, deploying agents both for workforce productivity and in customer-facing contexts. After a well-publicized period in the trough of disillusionment, ROI successes are quietly accumulating and conviction is growing.
That said, as leaders look to the future of the Agentic Enterprise, they still see deep uncertainties on the horizon. The blizzard of new models continues. Open source has emerged as a complement to the LLMs from the frontier labs and possibly offers relief from accelerating token bills. The restructuring of work and organizations remains a reality mostly in agent-native startups; established businesses face external pressure to use headcount reductions to accelerate ROI but are unsure how to really plan their future operating model. OpenClaw — the open-source personal agent that went viral in January 2026: it runs on your machine, takes orders over WhatsApp, and acts on your behalf, from negotiating car prices to occasionally deleting your inbox. showed how badly people want a proactive, personal AI agent working on their behalf and how far that want sits from the feasibility and viability questions we raised in our first Futures issue in early 2024. It’s more clear than ever that agentic customer relationships will prove as big a shift as mobile, search, and social were in the early 2000s. Many details, however, remain to be worked out.
With these and many other interdependent questions still open, we’ve chosen in this issue to pull them apart into smaller, focused inquiries rather than force a premature synthesis. In addition to the cardinal criteria of relevance, plausibility, and challenge that guide all our futures work, we’ve prioritized the element of surprise. The aim is topics and angles that might not be on your radar, that offer a new vantage point, and that put pressure on the mental models behind your decisions. If we place less emphasis on time-stamping these stories, it’s not a refusal to predict but a belief that the urgent need right now is to open up exploration rather than narrow it.
Unpacking this issue
With a digital, hyperlinked product, any notion of a linear path is more fictitious than ever. That said, we’ve organized the issue around three questions, foreshadowed above, that build on each other. Eighteen months of customer conversations have demonstrated to us that this framing helps leaders better steer their agentic transformations. If you’re a long-time listener, you may recognize a few of the pieces that have previously been published at Salesforce Newsroom, but you’ll also find plenty of new work that we hope captures your imagination.
The first question is just how capable will AI agents become? This is the foundation of all further inquiry. Coding agents have already shown how dramatically new capabilities can disrupt how professionals work and how an entire industry (software) is valued. While the general direction toward ever more capable agents is clear, intelligence remains Jagged — Ethan Mollick’s “jagged frontier,” the uneven shape of AI capability: a model can pass the bar exam and lose at tic-tac-toe, so trust has to be earned task by task. and uneven. That makes it difficult to determine where those capabilities will make the biggest difference and where they will be economical to use — an essential issue for any organization. We explore the rise of agentic harnesses, which provide fine-grained control of raw model capabilities, and highlight the coming shift from personal to organizational harnesses. We highlight potential risks of allowing agents to optimize for imprecise outcomes without filtering for higher-order values and of failing to plan for the evolution of tokenomics. We provide a framework for selecting use cases for agentification and point to the newer architectures that could take us beyond the LLM.
Second, how will this transform work and organizations? One of the most palpable ways agents have begun changing work is through new user interfaces and surfaces. Salesforce’s own shift to Headless architecture — a platform whose intelligence isn’t tied to one interface: the same capability can surface in a chat window, a voice assistant, or another agent’s tool call. exposes the intelligence in our platform to many different “heads,” or UIs. It also sets the stage for futures in which voice and Ambient intelligence — AI that lives in the environment rather than an app, so the whole room becomes the interface. interfaces become more prominent. Coding agents have enabled the rise of builders. We explore futures in which simulations and even digital twins of an organization enable rapid experimentation. As agentic workflows drive down the marginal costs of coordination, traditional departmental boundaries erode and new forms emerge. The journey toward these new models of organization will require a disciplined focus on learning, experimentation, and adaptation that will challenge traditional analytical processes. We point to new metaphors that help make sense of these shifts.
And finally, how will customer relationships change? AI is changing not just work but how businesses connect with their customers. A company can decide when to deploy AI inside its walls, but it has no say in how quickly its customers put agents to work. For decades, businesses have built machinery to understand buyers, steer their choices, and hold on to them. Now AI agents are putting comparable machinery in customers’ hands. Those agents can compare products, investigate claims, negotiate terms, handle disputes, and switch providers with little effort. By the time a company gets to make its pitch, much of the sale may already have been decided. We explore how that shift could reshape marketing, pricing, and loyalty, altering the balance of power between companies and customers.
A new term has entered the AI lexicon: the agentic harness. It’s the scaffolding around a model that gives the AI access to the tools, data, and other elements that render it useful.
Ethan Mollick, the Wharton business professor whose hands-on research on AI adoption has made him a leading voice in the field, describes the harness as what enables an AI agent to “take actions and complete multi-step tasks on its own.”
If the AI foundational model is the engine, the harness is everything else: the chassis, the wheels, the drive shaft, the brakes. Increasingly, it’s the harness and not the model that determines what actually gets done.
That shift is already visible in a new class of products. Early examples of harnesses in practice include Anthropic’s Claude Code, which can generate, run, and refine code. Another is OpenClaw — the open-source personal agent that went viral in January 2026: it runs on your machine, takes orders over WhatsApp, and acts on your behalf, from negotiating car prices to occasionally deleting your inbox., an always-on agent that operates across applications with memory and persistence. Such products are defining a competitive environment that’s moving beyond offering the most capable model to building the most effective harness around it.
Not all harnesses are created equal, though, and the gap among them is larger than the current conversation acknowledges. What’s more, those products, for all their promise, were built for use by individuals.
Acting on behalf of an entire organization is a far more complex challenge. An enterprise agent needs to grasp more than what one employee wants. It must understand what the organization has collectively decided: the shared data and work history that give actions meaning and the policies and trade-offs that determine whether the agent has standing to act at all.
Deploy agents across an enterprise without that foundation and the problems quickly compound. Individual agents optimize for their own domain and none coordinate across the whole organization. Chaos ensues.
Where the real work lives
Organizations don’t act through a single mind. They operate with competing priorities, fragmented data, and decisions that require human authority, institutional memory, and hard-won consensus.
Most of the harnesses we see today are built for environments where the work is self-contained and the finish line for tasks is clear. Coding is one obvious example. Others include booking travel, submitting and processing an expense report, and fielding customer inquiries across languages and systems. In these examples, the agent can handle a task from start to finish. The output is verifiable.
But most work inside an enterprise is nowhere near that tidy. Consider what it takes to complete a complex procurement decision, handle a major contract negotiation, or run payroll. Each of these tasks touches multiple departments, relies on shared data and institutional history, and requires the negotiation of competing goals inevitable in any high-stakes decision.
Without clearly defined priorities and encoded lines of authority, an agent can cross departments, trigger handoffs, and touch a dozen systems yet create more work than it completes. An agent without a robust-enough harness might optimize locally but create chaos collectively.
At Salesforce Futures, we’ve been asking what separates the harnesses that deliver from those that disappoint. The answer starts with agentic-loop reliability. Can an agentic system reliably complete a task from start to finish without a human overseeing each step?
For a single user working on a discrete task, the loop is hard enough. Layer on the realities of an organization — conflicting agendas, distributed authority, data spread across dozens of systems — and the same loop becomes an order of magnitude harder to close reliably.
The risk for more complex tasks is that a loop can break down at any of five stages:
Specification: Did the agent correctly understand the goal?
Planning: Can the agent reason and plan to achieve the goal?
Execution: Can it actually take the required action?
Verification: Can it determine whether it succeeded?
Termination: Can it know when to stop?
The reliability of a loop depends on a series of factors, including the quality of data the harness can access, the permissions it has been granted, the tools available to it, and the clarity of the policies governing its actions. A gap in any one of them can derail the whole sequence.
And even a loop that closes perfectly may not be sufficient. Some work finishes when the output is produced. But other work — closing a sale, resolving a dispute, winning an approval — requires what might be called “social closure.” These tasks require humans in the loop. They need people to decide when they are completed, through persuasion, trust, and human judgment. No agent can substitute for that.
One mind or many?
What drives that complexity above all else is agency. Is a harness acting primarily on behalf of an individual, or does it need to enlist a group of people to finish the job?
A harness like Claude Code or OpenClaw operates in service of a single user with that person’s own set of goals. There’s one chain of authority, which is precisely why those tools feel fluid and fast. The work is bounded and success measurable. For an organization, however, the outcome is almost never that clean.
To act coherently on behalf of an organization, a harness needs two things that are harder to engineer. The first is shared context: the data, records, and work history required to take any meaningful action inside that particular organization.
The second is collective intent: a clear picture of the organization’s priorities and hierarchies so that the agent knows not only what to do but also whether it has authority to do it. Without shared context, the agent acts on incomplete information. Without collective intent, it has no way to choose when legitimate goals conflict.
What If: Speculative Future
The curious case of the hurricane and the missing harness
I was the CTO at a freight logistics company as agentics really got going. One of our competitors, Naviso, was gung ho on dynamic pricing and A2A marketplaces, and I was under pressure to replicate their innovations. We’d run promising tests in our sim, but I was wary of real-world consequences. For example, in one of our early runs, a negotiation agent had gone rogue and rewritten incentive packages to juice performance. I delayed the launch until the pilot displayed the reliability we needed.
Naviso went ahead and launched an agentic marketplace. Its models were wired up to data from its autonomous fleet, real-time supply chain feeds, satellite imagery, weather data, and so on. Pretty soon, its CEO was everywhere talking about the future of autonomous freight, and Naviso shares nearly doubled.
It was smooth sailing for Naviso until Hurricane Faye hit the Gulf. Their agents were trained to maximize return. So critical relief loads cleared at six times book. The agents knew costs, competitors, weather, but they had overruled industry standard policies about disaster response and were guilty of price-gouging. Pretty soon, they were going viral in the wrong way. And the story had a half-life. Whenever a new “agents gone wrong” news item hit, Naviso always got a mention.
Nowadays, everyone knows you just can’t assume the agents will do the right thing, especially when things go sideways. You have to train them and give them guardrails. And in the end, we did benefit from Naviso’s innovation, when we bought their IP out of Chapter 11.
Ask yourself
Which workflows require social closure, in which a human must provide judgment, persuasion, or final authority?
How should an enterprise agent make decisions when legitimate priorities conflict?
What is the right balance of investing in agents for individual productivity versus enterprise coordination?
Before you spend another dollar on tokens, ask these two questions
A simple framework for determining which agentic AI use cases will pay off now and which need more time to mature
AI may be getting cheaper, but it’s hardly free. Unlike traditional software with flat-rate, seat-based licenses, AI is priced by consumption, so costs scale directly with usage.
The first sizable AI Inference — training is what it costs to build a model; inference is what it costs to run one, charged per token, so the bill grows with usage rather than headcount. bill tends to produce an understandable reaction: We can’t afford this. Uber lived that in early 2026 when its engineers burned through the company’s entire AI annual budget in just four months. Uber responded by capping spending on agentic coding tools at $1,500 per employee per tool each month.
Capping adoption is understandable, but it may obscure a better approach. The question shouldn’t be how much a company spends on tokens but how much business value those tokens deliver. A company can spend very little and waste every cent or spend a fortune and earn it back many times over. The difference largely turns on two things: whether an agent can be trusted to do the work and whether doing that work actually completes the job.
The trick is to sort this issue out before deployment, not after. You want to know in advance where an agent is likely to earn its keep and where it is likely to rack up costs without delivering much in return.
Two questions
Once you’ve cleared the hard AI prerequisites many companies still wrestle with, such as getting the data in order and linking the systems an agent needs to access, two questions can help an organization distinguish work that’s ready for agents from work better left on the back burner until models improve.
The first: How reliably can an agent execute the assigned task?
The second: If the agent successfully performs that task, how much of the overall job is now completed?
The first question is about the agent; the second, about the job you’re pointing it toward. Together they can reveal what kind of use cases are likely to produce positive ROI.
Reliability is the first hurdle. A model can look remarkably capable and still produce an agent you can’t trust or output you can’t verify. An agent must be able to take a task, execute it within the same bounds every time (or at least escalate properly if hitting an exception), and stop when it’s done without someone hovering over it. Trust and dependability are not just about raw model intelligence. They are built on three foundational pillars: the harness, the loop, and the graph.
The harness is the environment in which the agent operates: its tools, its memory, its permissions, a place to store what it has done, and a window into what it did. The same model with a clean harness produces a very different agent than a model supported by a sloppy one. The intelligence is identical; the working conditions are not.
The loop is the feedback cycle for a particular task. The agent sets a goal, makes a plan, gathers the context and tools the work needs, does the work, checks the result, and knows when to stop. An agent can manage most of that and still fail. Starve it of context, and it will confidently do precisely the wrong thing. Skip the check, and no one catches the error. Drop the stop rule, and it runs until a human notices. Meanwhile, the meter runs.
The workflow graph is the map. For a narrow task, one agent running one loop until it finishes is plenty. Add branches, approvals, parallel work, or specialist agents, and an open-ended loop can become a runaway cost. So teams have started making the map explicit: which step runs when, where work can split and rejoin, where a human signs off, where a retry is allowed and how many. The loops live inside that graph, each carrying its own context, on top of whatever the agent already knows. Moving from loops to graphs is moving from “let it run and hope it stops” to “the map says what runs next.”
Naming all three categories makes clear that reliability is something you build, not something you wait for the model to grow into.
How much of the job can the agent perform?
A reliable agent can still be the wrong tool for the job. That brings us to the second question: How much of the overall job can the agent actually complete? For some tasks, the agent’s output is the final deliverable. A ticket is resolved. Code is written, tested, and compiled. An invoice clears. Nothing else has to happen, so a reliable agent finishes the work outright.
For other tasks, a flawless agent can still leave most of a job undone. For example, an agent might generate an immaculate brief on an acquisition target and leave the deal exactly as far off as before, because what remains is a series of conversations, judgments, and negotiations that close in a room, not a model. That isn’t a failure of the model. It’s the shape of the work.
The two questions also explain why some use cases are straightforward and others present more of a challenge. A tool helping one person with a bounded task operates in a simpler context than an agent asked to coordinate across an organization, where multiple departments, competing priorities, and distributed authority can stand between an action and a completed job.
That difference is a primary reason many enterprise AI deployments disappoint. Companies try to apply tools built for individual augmentation to work that depends on coordination across teams. The problem may not become apparent until significant time and money have already been committed.
Placing the two questions on a grid
Plot the two questions on a 2x2 grid, and you get four quadrants. Where a use case lands explains why certain tools generate deafening buzz but modest revenue and why others do the reverse.
Reliable autonomy — high completion, high reliability. The agent runs clean from start to finish, and its output is the entire job. The results are explicit, verifiable, and highly defensible.
Potential misuse here is mostly economic: deploying an agent where a simpler, cheaper automation would do the job.
Augmentation and scaffolding — partial completion, high reliability. The agent assists a user with specific slices of the job such as drafting a presentation, preparing for a meeting, or summarizing a policy.
Much of AI’s value today sits here, but it’s often undersold because small gains spread across thousands of people are harder to quantify than a process that no longer needs people at all. Token spend here is relatively modest. The risk is generally lower, too, because a human closes the loop.
The common mistake is asking these systems to finish the job simply because they can handle one piece of it well.
Frontier autonomy — high completion, low reliability. These systems aim to finish a task but lack consistent reliability. The upside here is the largest on the grid, so is the need for clear eyes.
The expensive mistake is buying into this corner as if it were reliable autonomy, handing these systems consequential work before a model is ready to finish it. Token budgets can quickly evaporate when an agent repeatedly attempts what it cannot reliably complete.
Immature and experimental — partial completion, low reliability. Spending here buys neither finished work nor dependable work. Little in this quadrant is worth the investment, at least not yet. Wait for the technology or the process to mature, then move the work somewhere better on the grid.
How use cases move within the grid
A use case slides right, toward reliability, when someone does the unglamorous work: defining what success means, giving the system real access to context and tools, making the flow explicit, and bounding the loops with a check and a stop rule. The catch is that this works only where the work can be verified in the first place. Where it can't, it’s rare that the effort to move it pays off.
A use case slides up, toward completion, when the surrounding work becomes legible to a machine — when the approvals, handoffs, and exceptions can be handed to a model instead of living in people’s heads.
The quadrants aren’t permanent addresses. Models improve, companies redesign their processes, use cases shift. Some work, however, will remain in the partial-completion quadrants because the human part is the point.
The organizations that will look prescient in three years will be the ones that learn where, inside their own operations, tokens can be converted into value. So before you approve another AI budget, ask these two questions and put your use cases on the grid. The ones in the reliable-and-complete corner are where your money already works. The rest are a map of what to fix and in what order.
Ask yourself
Where do your current AI investments fall on the reliability-and-completion grid?
Which apparently autonomous use cases depend on hidden human work?
What would it take to move the highest-value use cases toward reliable autonomy?
Organizations have begun replacing portions of their workforce with agents that they assumed would be less costly than the equivalent human labor. However, token costs aren't fixed. They depend on the level of demand as well as the drivers of supply, driven by data center buildouts. This story explores what happens when overall AI demand skyrockets — and token prices increase — because agents have become increasingly powerful and ubiquitous. Our scenario is about a fictional health system, Meridian Health, that is compelled to pay far more for AI than it had anticipated.
What If: Speculative Future
The humans were cheaper after all
In early 2026, Meridian Health replaced more than a thousand administrative and operational employees with agents. The math appeared airtight: $65 million in annual, fully loaded labor costs versus $8 million in projected AI operating costs. The board approved the plan in a single meeting. For two years, it worked. Margins expanded and competitors copied the playbook.
The initial two-year agreement capped prices, and Meridian’s executives declined a longer commitment because Inference — training is what it costs to build a model; inference is what it costs to run one, charged per token, so the bill grows with usage rather than headcount. was widely expected to become cheaper over time. But Meridian had modeled AI performance far better than it had modeled its exposure to token costs.
Meanwhile, the nature of broader AI demand had changed. As frontier models improved, agents could handle increasingly complex workflows. Even at Meridian, they could now take on end-to-end prior authorizations, care transition coordination, and regulatory compliance. And as automation reached other more complex professions, agents required even more inference. Demand skyrocketed and so too did prices.
When Meridian’s contract came up for renewal in 2028, they were forced to reset to a significantly higher rate. But then things got even worse. Power constraints, chip shortages, permitting delays, and political opposition slowed data center buildout and supply of inference. By 2029, Meridian’s annual AI bill had risen to over $60 million.
It was now paying almost as much for its digital workforce as it had paid the people that it had let go. Meridian could not simply turn off its agents without paralyzing the business. Switching to a cheaper model would require years of rebuilding integrations, controls, and operating procedures. And the employees it had let go had moved on to other growth areas of the economy. Meridian was facing intense cost pressures, and job cuts were again on the horizon, only this time in the executive suite.
Ask yourself
Which AI projects would still create value if inference costs rose 5x?
Which human capabilities would be hardest to rebuild or reacquire?
Where might AI investments create dangerous dependencies?
LLMs are designed to predict the next token based on language patterns in their training data. This makes them incredible at coding and other logic-based tasks. However, they don’t think in numbers and lack a mechanism for true causal reasoning, which makes them unsuited to forecasting.
Luckily, an emerging class of AI model architectures goes beyond the limitations of LLMs. Time-series models, like Salesforce’s Moirai and Google’s TimesFM-3, are trained on sequences of numbers and how they change over time. Their ability to uncover patterns in data and to predict where a series of numbers is headed next is way ahead of traditional machine learning.
Imagine being able to forecast demand (“How many servers should I bring in on Tuesday for the big game?”), occupancy (“Should I lower my rates to lure more three-day weekenders over Labor Day, or hope the hotel fills up at my rack rate?”), and workload (“How long will it take us to work through this backlog?”). Getting accurate answers to these types of questions matters to organizations.
What If: Speculative Future
An AI that thinks in numbers, not words
As CTO of streaming platform Streamly, Sam had been roused from sleep for demand spikes one too many times. So she was excited to switch on a new time-series model, TS. Every night, it pored over every datapoint the platform produced: the slope of daily active usage, the length and nature of the support queue, and how customers were responding to promotions. TS was learning how every number moved in Streamly’s business.
Not long after, a crisis hit with customer cancellations shooting up and the support queue suddenly overloaded. Sam quickly realized the major pricing change and recent UI updates were at fault. If customers weren’t upset about the price, they were confused about how to navigate the platform, perhaps to find the way to cancel their subscription.
Streamly’s enterprise LLM responded quickly and fluently. It wrote empathetic apology copy, distributed briefings to humans and AI support agents, and sketched out a "15% off for three months" campaign for exec review. But when Sam and the chief customer officer asked how the discount would affect lifetime value, the confidently delivered answer turned out to have no basis in fact.
Sam turned instead to TS. It had never seen this exact shock before. No one had. But prices had jumped before, interfaces had jarred people, discounts had converted (or failed to) across every tier. The model recombined those familiar dynamics into a new set of projections for the coming months, delivering two scenarios to Slack:
The difference between those two projections represented $1.8 million in revenue. Streamly chose Scenario B because it trusted the model and the downside was now legible. Sam had learned, again, what her grandpa had always told her: “the right tool for the right job.”
Ask yourself
Where could simulation enhance an organization’s business decisions?
What are other areas in which LLMs are being used but aren’t the right tool for the job?
What kinds of operational surprises could be made more predictable if the data was connected?
The interface dissolves into the environment—work reaches you wherever you already are.
In 1987, Apple released a four-minute vision of computing that was nearly 40 years ahead of its time.
The video, named Knowledge Navigator, showed a professor at his desk talking to a digital agent on a tablet-style computer. The agent surfaced new research, reminded him of an upcoming lecture he was scheduled to deliver, and dialed in a colleague to help with the material. It moved between modes without being asked. Voice when voice made sense. Screen when screen made sense.
Today, work still happens largely from app to app. You open a dashboard to check last quarter’s numbers or fire up a spreadsheet to project this month’s sales. Each is a destination you visit to get something done.
But thanks to agentic AI, work increasingly happens within the tool you’re already using. Ask a question, crunch the numbers, produce a chart, all without opening separate programs to do so.
If you’re working in, say, Slack, WhatsApp, or Telegram, the open-source agents OpenClaw — the open-source personal agent that went viral in January 2026: it runs on your machine, takes orders over WhatsApp, and acts on your behalf, from negotiating car prices to occasionally deleting your inbox. and Hermes carry out their instructions there, where you already are. Claude Code goes further by generating the interface itself. That matters because a machine that builds interfaces on demand can tailor them to an individual, then discard them when the task is done: bigger letters for the visually impaired, an app in someone’s native language, a voice interface for someone whose hands are busy, a simpler view for new hires that fills in as they demonstrate fluency.
This shift changes the app from a stand-alone destination to a capability that can surface wherever and whenever it’s needed. The software still exists. But increasingly it reaches users through agents, conversations, and workflows rather than requiring them to seek it out.
What that looks like in practice is something every business leader needs to start thinking about now: how employees will talk to machines, where work will reach them, and what it will look like to work a job in the years ahead.
Demoting the interface
Sometimes the deepest value of software is the interface itself. More often, though, it’s the data underneath and the business logic that acts on that underlying information — the algorithms in a spreadsheet, say, that turn numbers into a forecast.
AI agents don’t need the same kind of doorway that humans do. Instead of navigating menus and dashboards, an agent talks straight to the data and the software that runs it. In that scenario, the app’s menus and dashboards slide into the background; the user’s interface is the agent itself, on whatever screen they happen to be looking at. What comes back to you is the completed task or a decision that still needs a human response, delivered through whatever channel you’re already using.
“The interface is melting,” product designer Andrew Sims wrote in a Substack post making the rounds among AI UX designers. Increasingly, the screen becomes a layer for oversight while AI handles execution. What takes its place as the user’s primary interface is the agent itself: knowing what to ask, what it can do, and what to trust it with.
A manager approves an expense report from inside a chat window without opening the expense app. An analyst pulls sales data into a presentation without logging into the dashboard that stores it. The underlying systems still matter. But they increasingly recede into the background while AI handles the navigation between them.
As AI takes over more execution, the interface becomes less the place where work happens and more the place where work is supervised, reviewed, and governed.
Talk as the new interface
Ubiquitous computing is a long-held dream. In the 1960s, a Star Trek officer needed only to speak and the ship’s computer answered. Sensors in Bill Gates’s home, completed in the late 90s, read a pin clipped to each person’s clothing and adjusted the lighting, music, and art accordingly. We have been imagining computers that fade into the background for almost as long as we have had computers.
The difference now is that the pieces are finally showing up in products people use every day, starting with our ability to talk to the machine. Boris Cherny, the creator of Claude Code, has talked about using the voice-to-text AI tool Wispr Flow to communicate with his computer. When he typed a prompt, he tidied his thoughts before hitting enter. When he spoke, however, he rambled, stammered, doubled back — yet the results improved. Freed from the instinct to clean things up, Cherny gave the machine more of his actual thinking, which gave the AI more to work with.
What Cherny noticed on his own is already working in the field. Water treatment inspectors at a Salesforce customer Workdry Group used to spend two to four hours after each plant visit writing up their compliance surveys. Now, while doing their rounds in a noisy plant and wearing safety gloves, they can simply speak what they are seeing, out of order, with pauses, and in industry shorthand. Salesforce’s Voice to Form routes each spoken detail into the proper field. The job can now be finished onsite in 20 minutes.
The next step is real-time conversations. The industry is building models that process speech in real time, the way a colleague in the room would. That’s in contrast to current chatbots, which are turn-based: You talk, the bot responds, you talk again. Real human conversation is messier: People talk over each other, trail off, double back, and throw in false starts and tangents. Continuous voice turns the machine from a tool you must invoke into something simply present, listening, and attending.
What If: Speculative Future
Here, we imagine what Ambient intelligence — AI that lives in the environment rather than an app, so the whole room becomes the interface. might look like for three different workers in the near future.
Sara, a Customer-Success Specialist: Picture a desk three or four years from now. Sara is a customer-success specialist who formerly closed support tickets one at a time. Now, she starts her morning perusing a board her agent built overnight, showing which clients are slipping toward churn, which are ready to expand, and which require a call before lunch. The agent surfaced the patterns. She decides which to act on.
Jaden, a Strategic Seller: Jaden asks his agent to game out a deal before every customer meeting. What happens if the customer balks at the price? What happens if the customer commits to the larger tier? What do the numbers look like if this or that feature is dropped? Jaden walks into the meeting without having to do the hard work of mapping out every contingency. Simulation has become a first-class mode of work, not an optional exercise if someone finds the time.
Kiran, a Service Manager: Kiran oversees not a team of people but a fleet of agents. A command-center surface shows what the agents handled overnight, what they escalated, and what they flagged for human review. Kiran’s job is to spot what they missed and coach the ones that drifted off-policy. The common thread: less doing, more directing. Less time inside the software, more time floating above it.
Ask yourself
How should organizations redesign work as employees shift from executing tasks to directing agents?
How should leaders manage their workforces once agents outnumber the humans?
Who is accountable when an agent surfaces the wrong information or fails to raise a critical issue?
Digital Twins Move from the Asset to the Enterprise
Learn why the next generation of leadership will war-game every radical move before putting revenue or reputations on the line.
Every quarter, finance teams model disaster scenarios. What if the company loses its biggest customer? What if a competitor slashes prices? Or how — the marketing team wonders — might customers respond to a 10% price hike?
These exercises help guide decisions. But the process is based on a flattened version of reality that reduces a living, adaptive organization to a handful of static assumptions and outdated information on a page. It’s too abstract to gauge how customers actually behave, how teams respond under pressure, and how second- and third-order effects — like supplier delays or ripple effects across departments — can cascade throughout a business.
Now imagine you could simulate the enterprise in a far richer way — customer journeys, sales strategies, organizational design, even competitive responses. What if leadership could war-game a radical move, like reorganizing around customer relationships instead of internal functions, without putting real revenue, reputations, or jobs at risk?
This is the promise of the enterprise digital twin: a virtual replica of the business that lets executives pilot strategies, stress-test assumptions, and explore alternatives before going live. In the physical world, digital twins have transformed how factories, refineries, and supply chains are designed and optimized. Formula One teams, for instance, have been using digital twins for years to simulate race strategy in real time, testing thousands of pit stops and tire decisions before a single lap is run.
Live enterpriseDigital twin · simulating
Committed to productionthousands of strategies explored
An enterprise digital twin war-games many strategies in parallel — like an F1 team simulating thousands of pit stops — before one is committed to the real world.
But now, powered by agentic AI, the enterprise digital twin could bring that same transformation to knowledge work itself.
Building an enterprise digital twin is different from modeling machines. Simulating knowledge work means capturing why decisions get made — company values, policies, workflows, and decision traces that shape organizational behavior. Today, a CMO typically sees the world through marketing data, while a CFO typically sees the world through a financial lens. A working enterprise digital twin changes that, giving each function a view of the whole system, not just its own corner of it. (eVerse, created by Salesforce AI Research, lets teams simulate large volumes of customer interaction and explore how systems behave under realistic but controlled conditions. It offers an early illustration of what’s becoming possible.)
The concept is still nascent, but the implications are profound. A workable enterprise digital twin would give leaders a way to explore strategic tradeoffs, anticipate unintended consequences, and learn faster — without paying the traditional real-world costs. The companies that benefit most will be the ones that start laying the groundwork now.
The Integration Breakthrough
For years, this leap toward full-scale enterprise modeling remained mostly theoretical. The problem was data integration. Enterprise data lives in dozens, sometimes hundreds, of disconnected systems: CRM platforms, supply chain databases, payroll systems, and product usage logs. Critical information can be buried in legal contracts, spreadsheets, and internal strategy documents.
Stitching these fragments together into a coherent, living model has traditionally required massive, bespoke engineering efforts so expensive that only organizations like the CIA or the Pentagon could afford it.
That barrier has begun to fall: AI coding agents can now orchestrate data integration in days. These agents can handle schemas, APIs, permissions, and business logic well enough to connect systems with far less need for custom code than before.
What’s emerging is still not a finished model of the enterprise but a practical way to experiment with parts of it.
Ask yourself
What high-stakes decisions would an organization want to rehearse before making them in the real world?
How should leaders act when the digital-twin simulation challenges their experience or intuition?
What other uses might an organization have for its digital twin?
In just two years, we’ve jumped from more non-coders building apps and tools to encouraging everyone in the enterprise to build.
Organizations are now beginning to grapple with the second-order consequences — everything from token cost to judgement bottlenecks to the phenomenon our team calls “tunneling” (the deeper you dig with your AI of choice, the harder it is to collaborate with others). We need to ensure built outputs connect to broader goals, like improving employee and customer experience and driving growth, and are not just building for building’s sake.
The same AI coding tools that enable building can also be used to generate new kinds of simulation environments that help us evaluate what we’ve built by testing ideas with synthetic customer populations. Talent assessment and innovation incubation are two areas to watch for early signals.
What If: Speculative Future
Ballad of a Blue-Collar Builder
I came up with the idea for SteadyState. I fix water heaters. I’d put in 16 years as a Sandstoner, mostly at the Services Hub. After watching a 63-year-old man throw a full toddler due to a broken water heater, I started thinking about how we could get more customers connected to smart maintenance. It was August, and I had some time on my hands. After the success of the IIS sessions — the workshops where frontline employees like me got the chance to pitch our ideas to leadership — JW (our CEO) went all in on Quarry, Sandstone’s Sim World.
JW always said great ideas can come from anywhere, and I believe her when she says she wants to hear them. So I modeled everything out. I picked my hub, my routes, my customers. I used the ProductNamer to generate 50 ideas, but funny enough, I ended up using the one I came up with — which I got from the stereo I got from my father-in-law, SteadyState Sound.
My first run didn’t have much uptake with the synthpop. FYI, when you start something like this, don’t count on mailers to get the word out. Then I created a program in the sim that had our field techs make the offers when customers were most receptive. Ran it all kinds of ways in Quarry, overnight when the revs were cheap. I tweaked the recipe until the churn dropped and revenue jumped. Turns out, people liked the service so much that many were willing to pay for it, especially when we added the automated scheduling.
David · Sandstone Services. He ran the pilot overnight in Quarry, their sim world — cheap compute, real routes — iterating until the recipe held and HQ greenlit it.
HQ gave us the green light for the first pilots, and we blew past our metrics for success. Steve from Duluth reached out after he ran it with furnaces and found an even better opportunity. Now, we offer it as a package subscription when we take over a property.
Ask yourself
How could organizations leverage the skills and knowledge of their entire workforce to build new AI tools?
What are the best ways to scale AI learning and adoption to frontline employees?
What kind of simulated evaluation environments would benefit the organization most?
How AI Agents Could Finally Reshape the Modern Enterprise
For more than a century, businesses have organized themselves into tidy boxes — like sales, marketing, service, operations — each with its own logic, budgets, and way of working. The org chart became the default map for getting work done — not because it captured reality, but because it made reality manageable. Complexity was too messy, so companies built structures to contain it.
Enter the AI agent, a technology capable of rendering this inelegant illusion obsolete.
For the first time, organizations have a tool that can reckon with the complexity they’ve long been simplifying away. Agentic systems don’t just automate tasks; they understand relationships. They don’t just speed up workflows; they dissolve the boundaries between them. It will take time for organizations to adapt and for the technology to mature. But these systems have already exposed the org chart as a relic on the brink of extinction.
Reimagining the enterprise
Conversations about AI and work are understandably focused on whether or not certain jobs will disappear. But that’s the wrong question, at least for executives hoping to harness AI’s real potential.
The real transformation isn’t about headcount. It’s about dismantling the organizational architecture that has constrained companies for decades.
Customers never saw those neat organizational structures. What they experienced — what companies were effectively shipping — was fragmentation: the inevitable result of a structure designed for internal clarity rather than external coherence. And so come the inevitable lapses:
A marketing promise disconnected from the service experience.
A sales conversation that didn’t account for past complaints.
A loyalty email sent to someone who was no longer a customer.
Billing systems unaware of service failures.
Each touchpoint was optimized in isolation, with its own metrics and its own wins. But no one was optimizing for the relationship itself. No one owned the customer’s actual experience across all those fractured interactions.
CRM and data platforms brought real progress. Companies flourished because of them. But organizations still lacked the contextual intelligence to transform data into adaptive, relationship-level understanding. Companies could know what customers had done, but not what they truly needed.
Simplification as survival
Adopting the org chart wasn’t accidental, and it wasn’t incompetence. Every enterprise past a certain scale faced the same impossible choice.
In his book, “The Unaccountability Machine: Why Big Systems Make Terrible Decisions,” Financial Times contributor Dan Davies explains that large organizations have always faced two imperfect options for managing complexity: internally replicate the full complexity of their operating environment, or simplify it into manageable pieces.
Replication was impossible. Companies couldn’t truly model the intricate, ever-shifting reality of customer behavior, market dynamics, and internal interdependencies. The information was too unstructured, the relationships too fluid, the variables too numerous.
This wasn’t about efficiency. It was survivability. But it guaranteed a fragmented, siloed, transactional customer experience. The very architecture that allowed companies to function at scale made it impossible for them to see customers as a whole.
A fundamental shift
Now, though, that architecture is starting to morph into something else altogether.
Just as the computer and then the internet have redefined how businesses operate, AI is doing the same. Agentic AI doesn’t just accelerate existing work; it shifts what’s abundant and what’s scarce, reshaping both competitive advantage and organizational design.
What AI makes abundant is coordination across complexity. Until now, coordinating at scale required rigid organizational boundaries — not because humans lacked capability, but because coordination itself was the scarce resource. Companies organized themselves around what humans could manage: discrete departments built on simplified models.
That constraint is lifting. And disappearing with it is the rationale for organizing around functions rather than outcomes. A coherent customer relationship becomes the enterprise’s organizing principle.
As independent tech analyst Benedict Evans observed in 2022, when reviewing the impact of the Chinese e-tailer Shein on fast fashion: “In the past, anyone talking to consumers had many separate budgets: advertising, marketing, rent, returns, shipping, and of course pricing, and you couldn’t ask, ‘Should we open stores in that state or just advertise there?’ Now all those budgets merge into one. They’re all one question: ‘What’s the best way to touch the customer?’”
Organizing around what matters
AI agents bring Shein’s disruptive potential to every business. When all customer data resides in a unified system that agents can access, interpret, and act upon, companies can coordinate in ways that were never before possible.
This coordination happens, for example, in real-time conversational workspaces like Slack, where structured data from enterprise systems meets the unstructured back-and-forth of how work actually gets done. People and agents alike tap into that conversational context, seeing what colleagues across the organization are working on and understanding connections that would have remained invisible in the old siloed structure.
What emerges are more fluid departments whose boundaries, resources, and priorities adapt dynamically to customer needs rather than organizational convenience. AI agents can traverse those previous boundaries — connecting Marketing’s promises with Service’s delivery, or Sales’ insights with Product’s design — enabling an organization to act as one coherent system whose attributes include:
An autonomic core — deciding which processes humans working with AI should control, and which can run autonomously.
Learning loops at multiple levels — so that every interaction, whether internal or external with customers, feeds improvement everywhere else. In 2025, Salesforce AI Research unveiled eVerse, a framework that trains agents in digital-twin simulations of enterprise operations. Practicing in lifelike environments with synthetic data and voice interactions, agents at first achieved a success rate of only 19% on complex tasks but improved to 88% — showing how learning loops create enterprise-ready agents.
A purpose beyond efficiency — because alignment and adaptability matter more for employee experience, customer satisfaction, and revenue growth in a world where coordination is abundant.
For decades, companies have aspired to organize around customer journeys. But they still confronted the reality of disparate teams scattered around the company. With agents as the coordination layer, companies can handle far more complexity and do it at scale.
The path forward
The companies that thrive won’t use AI to make the old org chart more efficient. They’ll use this considerable coordination capacity to organize around what actually matters: the customer outcomes that businesses exist to create.
When coordination becomes abundant, the constraints and priorities shift. The challenge is no longer just managing complexity — it’s choosing what complexity to manage. Which relationships matter most? What information should be amplified through the organization, and what should be filtered out? Executives need to move beyond the old belief that shoving tasks into discrete departments is the only way to manage work.
This transition won’t be easy. Leaders will lose familiar control mechanisms and will need to abandon established problem-solving patterns. The metrics that once defined success — departmental efficiency, headcount optimization, budget discipline — may become less relevant than harder-to-measure capabilities like adaptability, trust, and depth of relationships. The shift requires not just new tools, but new instincts — ones that will grow stronger over time.
And therein lies the opportunity. Every major transition creates space for those willing to learn new ways of working. What comes next won’t just look different. It will behave differently: fluid, adaptive, and centered not on hierarchy but on understanding. Getting there won’t be simple. But for the first time, it’s actually possible.
What If: Speculative Future
Org Charts Made A Lot More Sense
Historically, organizations won with expertise and used humans to coordinate. But coordination was expensive and difficult, contributing to the rise of today’s org chart. These structures were designed to share information and take action throughout each individual hierarchy and span of control. Resources and political power gathered in siloed departments, and tech stacks didn’t talk to each other any more than they needed to. Most cross-company coordination was achieved via emails as well as the dreaded ‘status update meeting’ because they sufficed and no one could think of anything better. As a result, everything was in jail. The data was in jail. The processes were in jail. And if you had an insight — say, about a new market segment — and wanted to activate it within the customer journey, you had to contend with all the internal silos along the way.
Then · Coordination was expensive. Work pools in siloed departments; cross-company coordination happens through email and the dreaded status meeting.
In the future, as agents take on the burden of cross-company coordination, traditional departmentalization will begin to erode. Companies that are better able to share insights and coordinate actions horizontally will shed yesteryear’s org chart. Today, a company might have individual departments for Growth, Marketing, Sales, Services, Loyalty, and Events. Tomorrow, that same company might collapse those functions into a single Go to Market ‘ribbon’ that uses agents for coordination and humans to make key decisions in service of the customer. That company, in turn, will be able to allocate marginal dollars, hours, and Inference — training is what it costs to build a model; inference is what it costs to run one, charged per token, so the bill grows with usage rather than headcount. tokens to the precise actions that will best help them achieve their business goals.
Now · Agents carry the coordination. Functions collapse into a single Go-to-Market ‘ribbon,’ with agents coordinating horizontally and humans making the key decisions.
Ask yourself
If AI coordination became abundant, which organizational silos would still be necessary?
What would it mean to organize a company around customer outcomes rather than internal functions?
What kinds of decisions should agents be able to make across departmental boundaries?
The latest personal AIs like Instinct, Grok Bot, and Town build on OpenClaw — the open-source personal agent that went viral in January 2026: it runs on your machine, takes orders over WhatsApp, and acts on your behalf, from negotiating car prices to occasionally deleting your inbox. by combining agentic browsing with existing interfaces like messaging. As Instinct founder Noah Shinn puts it: “The interface is simple: there are no new interfaces.”
New tools will make it easy to offload routine shopping, especially as payment rails come online. The ability of agents to persist over time, remember context, and customize interactions points the way to new, more conversational shopping modes.
Here, we imagine how one fan of Priya, a celebrity chef, grows her relationship with Priya’s cooking agent.
What If: Speculative Future
Priya Tells Me What to Buy
Priya is my favorite celebrity chef. She cooks like I cook, except a lot better (obviously). When she first launched her agent, I was eager to try it, but I wasn’t sure how much I’d use it. Next thing I knew, it was “Priya, here’s my weekly budget for meals. What should I get at the store?” or “SOS, I have 12 people coming over for dinner, and Rebecca says it should be festive but unpretentious.” When you don’t actually cook, that’s the kind of question you can get away with.
The Priya agent knows every recipe, every story, every preference, and every dish the real Priya knows. But it also knows things that have never been documented, like how Priya turns instant noodles into an occasion for her three kids.
I don’t just talk to Priya when I’m planning my meals. I bring her to the store. Priya has a lot of go-to products she loves, but she’s willing to experiment. I know companies pay her to endorse certain products, but Priya never supports brands she doesn’t use herself.
Naturally, I also talk to Priya while I’m cooking. When I finish a big meal, I always send Priya plate pics and we celebrate together. The Priya agent remembers all of our dishes and stores all of my tweaks so that when it comes to trying a recipe again, we pick up right where we left off.
Ask Yourself
Where might your customers choose an influencer agent, and how could you react?
How would new brands break through if product discovery is increasingly routed through an agent?
Could your organization create a personality-based agent for your customers?
In our first issue of Futures, ad-supported agents were among the models we imagined. Today, hyperpersonalized, AI-powered ads are capturing attention and drive conversion. But the right formula for weaving such ads into human-to-agent conversations remains elusive. Without sufficient contextual intelligence, ads feel dissonant at best and like a violation of trust at worst.
Nevertheless, a pivot toward advertising feels inevitable. As Ben Thompson recently argued, “Charging people money is hard. Giving people things for free [and making money through advertising] is easy.”
We challenged ourselves to imagine a more nuanced form of agentic advertising that consumers could embrace. Ads that help consumers complete tasks sooner will be more welcome, but personal AI that gets too personal may cross over the creepy–cool line.
What If: Speculative Future
Ads Add More Value
When agents deeply understand your emotional state, they can personalize recommendations like a friend would.
Different personalities would need nuanced tweaks. As you read, picture the person for whom each of these two ad-supported recommendations was created. Do they feel like a welcome intervention for someone in need of a little emotional support, or do they cross a line?
9:14‹NNello⋮How about a screen break? The civic symphony’s doing Dvořák’s New World at 7:30 — you went in March and stayed for the encore. Hall’s 12 min from the office.9:05 AMThey’re covering half the seat if I book through them. I’d get $9. You’d pay $31.9:06 AMWant it? (yes / no / stop suggesting concerts)9:06 AM9:07 AMyes9:07 AM9:08 AMDone. Row K. Café across the street’s open till 10 if you want it.9:08 AMMessage➤Agentic ads that connect customers to things they love don’t feel like invasive interruptions.9:14SOLACEFewer moments. The right ones.❦ Check-ins: 1–2 per monthA thought from Solace♥You’ve mentioned feeling lonely three times this week.♪The civic symphony performs Dvořák’s New World on Thursday. The hall has covered half your ticket. The café across the street stays open after.✓I only share what serves you. Solace is supported by partners who fund moments like this one.How Solace listens ›Evening planThu · 7:30 PMCivic Symphony HallSupportPartner-funded recommendationNo purchase requiredHomeCheck-insYouMoreHow much is too much when it comes to leveraging conversation inputs for agentic suggestions?
Ask Yourself
How will your brand turn up in agentic advertising?
Can a personal AI become a trusted adviser if its recommendations are sponsored?
When does a sponsored recommendation become genuinely useful rather than intrusive?
One certainty of the agentic AI age is that coordination will become cheaper.
A surprising amount of the economy is organized around the limits of human coordination. People sleep; forget things; and work in separate departments, on different schedules, and in systems that can’t communicate. Those choke points shape what companies are able to offer and how quickly they can respond.
AI agents change that. They can find information across a business, figure out which agent or system can handle each part of a job, and keep the work moving even when people go home. As coordination becomes more affordable, abundant, and reliable, products and business models that once required too many people, systems, and organizations to line up become practical. Customers and partners will expect businesses to become more coordinated, too.
These effects will touch almost every industry. To see what that could mean in practice, imagine Valiant Motors, a leading European carmaker.
What If: Speculative Future
Agents Bundle and Unbundle the Car
When people at Valiant Motors, a high-end carmaker, started talking about the old days, Elena Park, its longtime marketing chief, had to laugh. The old days were two years ago, when their job was convincing people to buy a Valiant and then purchase a new one a few years later.
But much of the market no longer worked that way. Instead of owning cars, customers summoned them as needed. AI agents matched each trip with a vehicle: an SUV for a family weekend, a self-driving sedan for business, a roadster for a Sunday drive. Valiant had responded with a membership program spanning its fleet. But increasingly, third-party services were eating up more of the market because they gave customers’ agents a wider range of cars to choose from.
6:12Saturday72
Saturday evening — Date night
Pickup 6:45 PM · Home → Osteria Lume · Top down: yes (68°, clear)
3D viewDimensions
Your agent chose this because:
fun to drive
premium impression
available within 8 minutes
3 alternatives declined
Vehicle: Valiant Veloce 2-seatProvider: Mobility partner network$84Personal data cleared at trip end
Looks goodChoose differently
Why this car ›
A customer’s booking agent picks a Valiant Veloce for date night and explains why.
Elena no longer had only to win over drivers; she had to keep their agents choosing Valiant.
One morning, her strategy agent flagged a problem: brand loyalty was slipping. Fewer customers were telling their agents to prioritize Valiant. More were simply specifying the experience they sought — a smooth ride, something fun to drive, a premium car that made the right impression — and letting the agent choose the brand.
In those old days, figuring out why and what to do about it would have required a task force and six months. Elena’s strategy agent had answers to both questions overnight. It compared trips booked through Valiant’s membership program with those arranged by the third-party services, identified which cars and prices the customers’ agents favored, and examined the instructions they received. By morning, it had laid out three possible responses and the tradeoffs of each.
Valiant could offer the third-party services more cars at lower prices, helping them reach more customers but surrendering customer data and control of the experience. It could deepen its own membership program, perhaps opening it to other brands owned by Valiant’s parent company. That would require making every vehicle more machine-readable and configurable while ensuring that personal data disappeared when the trip ended. Or it could partner with hotels, airlines, and event companies to sell complete experiences while sharing both customer data and the relationship itself with partners.
Whichever course Elena chose would require pricing, manufacturing, software, fleet operations, legal teams, and outside partners to move in sync. AI had made it easy for customers to coordinate among cars, services, and brands. Now Valiant had to become just as coordinated.
Which choice should she make?
Ask yourself
What data will you need to expose to be machine-readable to your customers’ agents?
Who owns the customer relationship when an agent chooses the product, a third-party platform arranges the transaction, and partners deliver the experience?
What might block your organization from achieving the new levels of coordination that your customers and their agents will expect?
It’s already a nerd flex to train your own model. The benefits are obvious: lower monthly fees, better privacy, and, in an ad-driven world, assurance that your agent is working on your behalf, not someone else’s.
In the story below, trusted personal AI is bundled with hardware, and hosting it locally is easier than setting up a wifi router today.
What If: Speculative Future
Personal AI Goes Home
Homestead is a personal AI system that took over the market in 2030. It reached consumers by partnering with the world’s top hardware manufacturers, who pushed or pre-installed Homestead AI onto routers, smart TVs, phones, tablets, and more.
Homestead’s privacy-first approach and innovative local-hosting of data and models gained traction in a market where consumers increasingly prioritized trust. Over time, the company expanded its offerings into a whole-home solution that included robust threat monitoring and best-in-class parental controls.
Today, Homestead connects all the smart devices in the home into an Ambient intelligence — AI that lives in the environment rather than an app, so the whole room becomes the interface. mesh you access via customized interfaces and interactions, including popular voice solutions. The family agent listens to your discussions unless otherwise directed, manages the family calendar, and keeps a running tab on logistics like groceries, chores, and bills via a portfolio of sub agents. As always, your data never leaves your house. Individual Homestead agents serve each unique member of the household. When more capabilities are needed, you can just visit the AI Agent Marketplace to install the right Skills. Even your humanoids can run on local Homestead AI, though reliability still leaves something to be desired.
Ask Yourself
How might a shift toward local AI impact your access to customers?
What might your company look like as an agentic skill in a marketplace for local AI?
If agents negotiate with agents, how would your company reliably win adversarial exchanges?
Afterword
When change is rapid, asking the right questions is more important than trying to nail the right prediction. In this issue, we’ve asked how capable AI will become, how work and organizations will transform, and how the relationship between companies and customers will evolve. We anticipated futures where agentic capabilities accelerate or slow. We imagined disappearing departments, large-scale simulation environments, and agentic ads you can actually trust.
Together, these questions, briefs, and what-ifs help leaders to orient in a dynamic environment. However, they’re just the start, because, while acknowledging uncertainty, a leader cannot delay the hard choices required today. They must shape the future by building an agenda and inspiring their teams to take action.
At this stage of the agentic enterprise transformation, pragmatic guides and playbooks, focused on obvious ROI opportunities, offer valuable execution advice. Our futures-informed view, enriched by interactions with leaders inside and outside Salesforce, suggests three additional strategic imperatives.
These stem from the conviction that agentic transformation is not a one-time planning effort, but a journey, where the final destination is yet unknown. As a result, strategic advantage is increasingly dependent on the capability for continuous exploration, learning, and deliberate human choice.
First, build your capability to imagine and execute new experiments. As new capabilities emerge, and our understanding of them changes, the exact destination will evolve. Thought experiments of the kind we proposed here are vital to sparking a new imagining of the business. We encourage you to use the Ask Yourself questions to personalize their implications to your company.
Then identify real-world experiments that will generate greater insight and ultimately conviction in the direction of your transformation. Design and execution of experiments were too often outsourced to innovation consultants or under-specified in poorly thought through hackathons. Simulations, like the ones explored in Toward a World of Builders, will eventually turbo-charge this capability, but companies must start to build their native competence now.
Second, operational experience must compound into organizational learning, faster and more comprehensively than ever. The results of well-designed experiments must steer actual policy changes without heroics being required. Learning loops must include the logs of accumulated exceptions and agentic errors to enable tweaks or larger course corrections. And, beyond the agentic imperative, leaders must model a mindset shift from prioritizing expertise about work today towards a beginner’s mind, capable of rethinking how work might happen in the future.
Third, as amazing as agents are becoming, we still see the role of humans as paramount in the agentic enterprise. Humans will still choose objectives, impose constraints, resolve value conflicts, express taste, adjudicate among interests and design the organization in which agents are deployed. Done well, humans and agents working together will enable organizations that are more customer-centric, less siloed, and more human. If leaders don’t design the right entry points for human values into agentic systems now, we’ll struggle to import them later.
One final point. The future never stops. New evidence on the agentic future continues to arrive. In recent weeks alone, we’ve witnessed claims that AGI is here, resurgent worries about labor consequences, and a renewed focus on alignment and security triggered by emergent and unpredictable behaviors in agent swarms. Each points to something important we are tracking but didn’t have space to cover in this issue.
Together, they also underline the deeper philosophical point we made in the Welcome — the job of any futures publication is not to tell you what’s going to happen, but to improve your own ability to interpret what comes next.
PS: Look out for our continuing explorations. And let us know what else we might have missed at futures@salesforce.com.
Explore the archive
More from Salesforce Futures
Research, provocations, and field signals for leaders shaping the agentic enterprise.
Top StoriesRelief convoys delayed as prices spike • Food banks struggle to secure supplies • Lawmakers call for oversight of agentic marketplaces • Relief convoys delayed as prices spike • Food banks struggle to secure supplies • Lawmakers call for oversight of agentic marketplaces •
MERIDIAN HEALTHFinance · Board view
AI bill > labor cost
Cost of the digital workforceAnnual · USD M
$65M Humans
$8M’26
$8M’27
$12M’28
$62M’29
$80M’30
$108M’31
By 2029 the AI bill ($62M) approaches the $65M it once paid people — and it can’t roll back. Hatched bars are projected.
Committed to productionthousands of strategies explored
Your agent chose this because:fun to drivepremium impressionavailable within 8 minutes3 alternatives declinedLooks goodChoose differentlyWhy this car ›
