You’ve built agents. Now build the enterprise that puts humans and agents to work, side by side.

You went all in on agentic AI and it paid off. Your first agents are deployed, and as you iterate, redeploy, and monitor them, you’re seeing real, measurable outcomes. The promise of our first AI Playbook, Become an Agentic Enterprise: A Step-By-Step Guide — getting your first agents into production and running successfully — is now reality. 

But your success is raising new questions: 

• How do you keep the momentum going? 
• How do you turn a couple of successful agents into an impact that transforms your entire business? 

This is where scaling comes into play.

We’ve hit the ceiling of human scale

For twenty years, the constraint on almost every business was the same: how much a human team could carry. How many cases one rep could close, how many deals one AE could work, how many tickets one engineer could reasonably resolve in a day. Everything about how modern companies are organized — pods, sprints, escalation ladders, quarterly plans and reviews — was designed to squeeze the last drop of productivity out of finite human bandwidth.

That ceiling has been reached. 

Joe Inzerillo, president of enterprise and AI technology at Salesforce, puts it plainly: “We’ve hit the capacity of human scale. Now we get to use AI to redesign the underlying processes that have been holding us back.”

This is the reframe on the future of work that we believe is worth reflection and discussion amongst teams in every industry. The point of AI agents isn’t to replace people in the workforce. It’s to finally do something about the human-scale processes that have been quietly straining every function in the business. Adding more agents on top of broken workflows doesn’t scale a company. Rebuilding the workflows underneath so that humans and agents can each do what they do best? That’s what actually scales.

How to use this guide

If you’re already running agents and starting to feel the growing pains, you’re in the right place. If you’d rather jump straight to the chapter that addresses your immediate bottleneck, our guide supports that too. And, if you are not ready to scale, are still piloting, or could use a refresher on the basics to share with your teams, start with our first step-by-step guide: Become an Agentic Enterprise.

The chapters in this guide include:

You’ll start by building the crew that can operate at agentic speed, for jobs that didn’t exist three years ago. From there, you’ll learn how to redesign the processes and systems underneath your agents, so scaling doesn’t mean rebuilding from scratch every time. Then, discover how to turn early wins into a system where agents find each other, talk to each other, and hand off work without you needing to stand in the middle. And finally, get an honest look at what’s coming next, so you’re building today what you’ll need tomorrow. The result is a foundation that acts as a single source of truth, one that empowers everyone in your org to be a builder.

Ready to scale agentic AI throughout your organization? Let’s dive in.

Create a crew of builders

For three decades, Harvard Business School’s Linda A. Hill has studied what actually makes organizations innovate — research that culminated in her latest book, Genius at Scale: How Great Leaders Drive Innovation. Her conclusion, after tracking innovation leaders from Pixar to Pfizer to our own AI Research team at Salesforce, is disarmingly simple: “Leading innovation isn’t about getting individuals to follow you into the future. It’s about building an environment in which they will co-create the future with you.” That reframe matters more than ever now. In an era where any company can rent frontier intelligence off the shelf, the durable advantage isn’t the AI you deploy, it’s the people you empower.

To build the most sophisticated agent orchestration system in the world, you need humans. You need them to, of course, design it, monitor and manage it, but you also need them in the weeds. A partner that collaborates with your agentic system as AI itself rapidly evolves. As we discussed in Chapter 2: Design your human-AI workforce in our first guide, part of this requires building AI fluency across your workforce before it’s needed, creating the conditions where employees see agents as force multipliers rather than threats, and designing how humans and agents succeed together. The other part? It’s about hiring people specifically for this work, for jobs that didn’t even exist three years ago. We know this isn’t easy (we’re building an AI-first workforce too) and the majority of businesses are figuring this out in real time as well. In our recent The State of Agentic AI in the Enterprise report we found that 89% of deployed organizations are already rethinking or reworking their workforce model to formally account for AI agents — and not one said they had no plans to.

This chapter teaches you how to build the team AI needs to succeed. We’ll look at how Salesforce is filling roles by hiring directly for this work. We’ll delve into what changes when you craft an agent-native workspace instead of having those “skills” bolted on to existing roles, and what it actually takes to lead a team where everyone is expected to build with AI. 

IT lays the foundation

Probably like you, before our agentic transformation, IT owned everything end-to-end with no shared infrastructure. Now, we’re living in a different world. We give teams from every discipline the tools to build what they want. But before doing so we had to ask ourselves, “How do we enable everyone to be a builder but still protect the integrity of everything that matters for running a company at scale?”. 

Our solution: IT still owns systems of record, integrity, SOX, trust, and compliance. The complex foundations are under their control so everyone can safely build. This gives our teams a safe place to tinker without having to route what they build through IT. The teams own whatever they’re building and shifting this ownership to the people closest to the problem makes speed compound. 

Building is now a team sport 

Crafting an agentic workforce goes beyond hiring and redesigning roles. We’ve learned it also requires a shared workspace where your people can build and work together with agents as teammates. For us, that shared workspace is Slack and within it, Slack Code empowers everyone to build, right in the flow of work. 

Unlike in the past, when being a technologist was a solo endeavor and required a specialized skill set, Slack Code is open to all skill levels and lets a team build an agent collaboratively, right in a Slack code channel. Once deployed, agents built in Slack Code can simply be mentioned, like a human teammate, and work with you. It also runs on the same enterprise-grade security and permissions as the rest of Slack, so there’s always a final human say before anything goes live.

The jobs that don’t exist on your org chart yet

If you’ve deployed agents, you know the work has changed. Agents don’t fail in the way software does. Unlike software where a crashed service can get engineering involved at 3am, an agent that’s messing up doesn’t throw up any alarms. It just keeps confidently doing its job, until a customer angrily complains on social media or someone internal catches it being inaccurate. 

To combat this from the start, Salesforce built a framework for this called the Agent Development Lifecycle (ADLC). You may remember it from our first playbook in Chapter 6: Never Stop Improving.

ADLC requires ongoing human management from an agent’s first day on the job. But successful agent management depends less on traditional job titles and more on the mastery of specific jobs to be done (JTBD).

How these jobs are staffed depends on two dimensions: the size of your company and the size of the job. A nimble startup might have one or two agentic technologists covering all nine functions. A large enterprise might have dedicated teams for each. A smaller org deploying a complex, high-stakes agent might need to flex a big team to support delivery. But all functions are critical to success in ADLC.

Through our work building and deploying agents at scale, we’ve identified ten critical JTBD that every organization must address:

Building functions

• Agent architecture: Designing the technical framework, including capabilities, boundaries, and the placement of “deterministic fences” around probabilistic reasoning.
  
• Agent development: Translating architecture into working behavior by configuring topics, actions, and custom code integrations.
  
• Knowledge management: Managing the data and content strategy, such as CRM records and knowledge articles, that provide the agent with the factual grounding it needs to reason accurately.

Improvement functions

• Efficacy analysis: Monitoring quality metrics and analyzing session data to identify performance gaps and run “performance reviews” for agents.
  
• Experience management: Designing the conversational framework and natural language “job descriptions” to ensure agent interactions feel intuitive and consistent.

Driving functions

• Business ownership: Defining business outcomes, setting ROI targets, and prioritizing which JTBD is worth an agent’s time.
  
• Product management: Owning the agent’s roadmap and feature prioritization — translating business strategy and user feedback into a backlog of capabilities, defining success for each iteration, and serving as the connective tissue between business ownership and building functions.
  
• Agent management: Providing human-in-the-loop oversight, monitoring live sessions, and handling real-time operational triage and escalations.
  
• Program management: Coordinating ADLC across cross-functional teams to manage timelines, dependencies, and the calibration loop.
  
• Change and adoption: Designing the training and stakeholder communication necessary to integrate agents into new workflows and shift user behavior.

The organizations that lead in this era will be those that staff these critical functions intentionally, ensuring every JTBD has a seat at the table.

The team is the thing 

Creating a culture of builders is an exercise that only becomes more cohesive with time. As we illustrated in this chapter, attracting AI-native talent through a designated, outreach program, hiring and redesigning more senior roles to cater to specific stages of the ADLC, and giving both groups a shared place to build, like Slack, is a multi-pronged approach that we’ve seen work.

The design of your team will be unique to your business. But you still want to bring in talent that’s ready right now, show commitment to developing your existing ambitious employees, and create a working environment where experimentation and new ways of working can happen transparently, in service of what the business is trying to achieve.

Try this activity: building with purpose 

Start designing the team your AI needs to succeed by answering the below questions. 

 

Move at the speed of agents

Just three years ago, the limits on how quickly work could get done were set by a human team’s capacity. Things like: How much can one person excel across different disciplines, or how many handoffs can one team navigate before it becomes a bottleneck? But these limits are starting to vanish. We’re now entering a working world no longer bound by how deep one person’s expertise runs, or how many handoffs a team can coordinate before it stalls.

With AI, even the meaning behind the words we used to describe fast-moving teams is becoming outdated. Nimble used to mean a team that could ship in a two-week sprint. Now, with agentic capabilities, two-week sprints have been condensed into two days. At Salesforce, Robbie Birbeck, VP of business technology, puts it plainly, “The cycles are shrinking. We’re experimenting with week-long cycles, but I think almost every day starts to feel like its own cycle.”

Teams still gather for something that looks like an agile scrum or stand-up meeting, but the work they take on is reactive, built on what got shipped yesterday, not what got planned two sprints ago. If you try to plan a week or two of work in advance, you’ll find yourself living in the past, because the team accomplished more than anyone predicted the day before.

Some are thriving in this new pace of work. Others are struggling, often on the same team. Salesforce didn’t try to solve that tension universally. Instead, we identified who was already operating at this new rhythm, captured what was working, and replicated it across the company.

This chapter is about that redesign. You’ll see where Salesforce points its newfound speed, and the small-team model built to capture it: the Rapid Innovation Squad, or RIS. You’ll see the discipline that keeps its results honest, and how to build a version of RIS on your own team.

Go after green space 

The speed at which agents can deliver is astounding. However, they need well-defined problems to pursue so they can produce meaningful value. In Chapter 3 of our first guide, we explored how to find the right first use case to put AI to work. Now that we’re scaling, we can go bigger and look for green space. 

Green space is an area with problems that have limited or no solutions. It’s an arena of pure potential that reduces risk and empowers action. Andy White, SVP of Salesforce on Salesforce technology, puts the importance of staking out green space in blunt terms: “Who’s your business stakeholder that’s really going to stand up and say I want to disrupt some things? Because what we’re talking about here is disruption.”

Two examples of our first green spaces: Salesforce Help and our Engagement Agent. For Salesforce Help, our customer service program, we had case volume growing 20% year over year and needed a more sustainable model. We created Agentforce to solve customer inquiries at scale. For our Engagement Agent, in sales, 75% of leads were getting zero follow-up at all. We chose to take on the pipeline risk of an agent working leads humans hadn’t touched.

Enter the RIS — Salesforce’s agentic “test kitchen”

If green space is the playing field, our Rapid Innovation Squads (RIS) are the teams at Salesforce that are “playing to win.” The RIS is our small-team model for building software alongside agents, rather than through a traditional enterprise roadmap. The structure is deliberately small: a three-person unit built from a business leader, a product-builder mindset, and an integrated engineer. Joe Inzerillo, president of enterprise and AI technology, who has helped scale RIS from a couple of squads to nearly fifty individual ones running in parallel, describes the calculus behind that size: “A typical squad’s two to four, maybe five people. It’s humans and agents working together, but in especially close coordination because the team size is so small.” 

When a team adopts this model, business and builders become one. Handoffs disappear, and decisions happen 10x faster. And because feedback comes back in days instead of months, you know faster what’s actually worth scaling. 

How RIS patterns up-level traditional sprints

The RIS organizational mindset 

When staffing a RIS, Salesforce looks for a tinkerer, someone comfortable with ambiguity. Seth Roe described the litmus test he uses when staffing a squad. It’s not “does this person have a six-month roadmap in their head,” but whether they can sit with not knowing. That, Seth says, is the mindset the model needs — someone who can identify what’s in their control right now and go build it, rather than waiting for a fully-formed plan.

The business side matters just as much as the technical side. Seth frames it through a “pioneer” model. The enterprise needs pioneers willing to chart uncharted territory, and it needs them willing to be player-coaches rather than pure managers. “The days of just being purely a manager are quickly disappearing,” he says. As for who ends up strongest in the seat: “The best teams we’ve seen so far are the engineers who have good business sense, and the business people who have good product sense.”

The RIS in action: SMB Signals 

Seth’s own track record makes the case for the model better than any framework could. At Salesforce, in his first month on the job, Seth’s RIS shipped Small Business Signals (SMB Signals) in three weeks.

The problem SMB Signals solves is simple to describe and, until now, hard to build quickly. From Slack to email to Tableau, small business sellers had their leads, incoming customer data, and more evergreen customer information living in multiple places at once. It made them spend endless hours hunting for customer signals that should have notified them automatically like, for instance, license upgrade opportunities or abandoned cart patterns. 

Through the RIS, the team built SMB Signals. Every action a seller needs is ranked and routed directly to them in Slack, with the context already there. From there, sellers take action with one click: logging the call, creating the opportunity, everything pre-filled.

To develop Small Business Signals, it moved through the RIS four-phase validation model we outlined above: 

1. Desirability
2. Feasibility
3. Value
4. Go-to-market 

Phase by phase, the prototype was tested with a slightly larger, more representative group of sellers, and each pass added fidelity (better UI, more mature data access, tighter security) before the team invested further. Crucially, someone from Slack’s core product team was embedded in the project from day one, not brought in at the handoff. That was one of the biggest lessons from month one: know where a solution is going to live long-term, and make sure the team that will own it there has a hand in the decisions from the start, so the RIS can eventually hand it back to be maintained and scaled.

The results came almost as fast as the build. In its first cohort, SMB Signals hit 100% adoption, sent more than 1,600 signals, and surfaced over $461,000 in open pipeline that sellers would otherwise have missed entirely.

Try this activity: Build your own RIS 

We’re not gatekeeping the RIS. AI empowers us all to be builders. Use this exercise to help create your own. First, figure out if your team is ready and, if they are, find a problem that would make an excellent first use case. 

Scale agents without starting over

When it comes to scaling agents, no one is sitting on their hands –– even from the start. We recently discovered in our State of Agentic AI in the Enterprise report, organizations are piloting an average of 10.4 agents at once. 

It seems obvious that your second deployed agent is where scaling actually starts. However, adding more agents to your roster doesn’t automatically equal scaling. How so? Think about the efforts that would go into building dozens of agents if every one gets built from scratch. You’d have to design its own data connections, its own guardrails, its own way of handling errors. You’d be repeating the same setup cost over and over. Scaling is about sequencing your agents so each makes the next agent easier to build and deploy. The organizations getting scaling right are building systems where agents can find each other, talk to each other, and hand off work without a human stitching it together manually. 

This chapter is about what it actually takes to build that system, so you’re running a group of specialized agents that work together instead of a pile of agents that are each on their own path, doing it alone.

From deployed to orchestrated

As you may know from reading Becoming an Agentic Enterprise: A Step-By-Step Guide, Salesforce’s Agentic Maturity Model has five stages — starting with simple chatbots (Level 0) up through multi-agent orchestration (Level 4). 

Many organizations reading this are already sitting at Level 3: Complex Orchestration, Multiple Domains (but if you’re not, visit our first playbook to help you get started).

At this level, agents work across departments like sales, service, and finance, pulling from a harmonized data source instead of pulling data from siloed buckets. 

The jump from Level 3 to Level 4 is what this chapter is about: agents that autonomously orchestrate multiple workflows with harmonized data across many domains. A good example of this: an agent handling complex cross-department workflows, such as managing a sales pipeline while pulling data from multiple sources to create a holistic customer view.

The three most important factors in advancing from Level 3 to Level 4:

Maturing to Level 4 requires defining how agents talk, how they find each other, and how you keep “any-to-any” (meaning agents that should be collaborating together) from becoming “any-to-everything” (meaning agents pulling any unrelated agent in to collaborate with and potentially risk sharing sensitive information with): 

There are three technical pillars that will help you achieve these objectives:

1. Shared communication layer

Agents that can’t talk to each other can’t hand off work. They need a shared way to communicate – a common protocol to exchange messages, state what they are individually capable of, and coordinate multi-step work with another agent using the same rules. That protocol needs to work across vendors and it needs a lightweight way for agents to describe themselves to each other so that discovery and communication aren’t two separate problems.

Now, a shared protocol lets agents exchange information, but it doesn’t automatically mean they can do everything. A shared protocol alone doesn’t allow agents to negotiate terms, verify each other’s claims, or know when to stop and escalate to a human. Organizations building toward Level 4 need to solve for both the protocol layer that lets agents talk at all, and the trust layer that lets them collaborate on anything that actually matters.

2. Dynamic agent discovery 

Once an organization has agents built by different teams on different platforms, someone has to actually know what each agent does, what data it touches, and whether five teams just built the same agent under five different names. Without a system that does this automatically, agent “discovery” becomes a spreadsheet someone updates when they remember to (spoiler: this is low on a priority list). 

You need a system that continuously detects new or updated agents the moment they go live, with data extraction detailed enough to know what an agent can do and what it’s authorized to touch. Plus, deduplication logic that catches redundant agents before they multiply; and a live, queryable catalog that both humans and other agents can check at runtime.

3. Control and governance 

Once agents start handing work to other agents across systems, access control stops being a single checkpoint and becomes something that has to hold at every step of a chain. A gap anywhere in that chain is enough for an agent to take an action it shouldn’t.

This requires:

• Identity that travels: If an agent acts on a user’s behalf, that user’s identity and permissions govern every downstream step. If an agent acts autonomously, its own identity is just as traceable. 
• Least-privilege enforcement: Agents only get the access a specific task needs, not standing access to everything they might someday touch. 
• Centralized policy enforcement: Ensure the same rules apply whether a request comes through an API, a UI, or another agent along with a complete audit trail, so every action can be reconstructed after the fact. 

Salesforce’s own platform is one live example of this operating together rather than as separate bolted-on layers. Identity and authorization propagate through every step of a multi-system agent workflow. Data access is enforced centrally, no matter how many systems or agents are involved. And every interaction generates trace data for audit. The result: identity, data governance, API enforcement, and AI trust operate as one system instead of four.

Twenty agents, one foundation

SaaStr runs the world’s largest community for SaaS founders with only three full-time employees but more than 20 AI agents. These agents span an assortment of roles including: sales, marketing, and event operations — but they all have the same mind. 

To do so, SaaStr made Sales Cloud its system of record, bringing nine years’ worth of customer interactions, event registrations, engagement signals, and account records into one single source of truth. Then, they used Salesforce’s AIforce architecture to give every agent, including the ones built outside Salesforce, direct access to the same customer data. The result is that adding agent #20 didn’t require rebuilding what the first 10 already had. Each new agent inherits years of customer history, event data, and engagement signals on day one with no custom integration, no new connector, no starting over necessary. 

The clearest proof it’s working is in Hexi, SaaStr’s SDR agent. Hexi runs outreach to more than 3,000 previously unresponsive leads with a 72% open rate. It’s the highest open rate across SaaStr’s entire AI stack. Why? Because Hexi draws on the same continuously updated customer profile every other agent contributes to, instead of working from fragmented data. Read more on SaaStr’s agentic success here.

Turning wins into a system 

The real difference between having a pile of agents and a system of agents is about making sure they can find each other, talk to each other, and hand off work without you standing in the middle. However, to build that system, you need a human team to design it. That’s not a job for whoever on your team happens to have bandwidth. It takes a team built for this work. That’s where we go next.

Try this activity: Level 4 readiness check 

Get a gut check on whether your agentic enterprise is ready to move up on the maturity model. Ask yourself these questions based on the three technical pillars Level 4 requires below.

The next horizon? Human-AI partnership

Two shifts are unfolding at the same time, and they only make sense together. 

• Humans are becoming builders, coaches, and taste-makers inside their own companies. 
• Agents are becoming ambient, adaptive, and able to continually learn and grow.

What’s exciting about this moment is that neither shift can exist without the other. Agents without human judgment produce confident nonsense at scale. Humans without agentic leverage stay trapped at the ceiling of what a pre-agentic team could carry.

At Salesforce, we’re working every day to scale our own enterprise to match the pace of a world of business that reinvents itself every quarter. We’ve hired some of the best AI research scientists in the world. Here’s our honest read on where this is going — and where humans and agents go together from here.

The skills that matter most are the ones agents can’t fake

Ask any leader running agents in production what worries them most, and you’ll hear a version of the same answer: It’s not that agents can’t do the work. It’s that they can do it confidently wrong. An agent that hallucinates a policy, misroutes a customer, or writes a beautifully-worded email based on a misread of the situation doesn’t sound broken. It actually sounds fluent. “AI slop” is, at best, lengthy prose that’s light on meaning or value, and at worst, downright inaccurate. It’s a challenge we never had to consider before the advent of generative AI. Which is why one of the human skills that will matter most in an agentic company isn’t prompting. It’s taste. Discernment. Expertise grounded in their own intellect, curiosity, and experience. The best workers in an agentic enterprise have the ability to look at a plausible-sounding agent output and say, “That’s wrong, and here’s why.” 

It’s what a great editor does with a draft, what a senior engineer does at a code review, what a seasoned rep does when a lead sounds promising but feels off. Joe Inzerillo says it more simply: “AI can only take you so far. What you need on your team right now are people with deep knowledge and really good taste.”

AI fluency is the leverage point

If discernment is the individual skill, AI fluency is what compounds it across a company. And the data is clear on where the leverage lives.

Salesforce research on the future of work has found that AI-confident managers are 20% more likely to lead AI-fluent teams. AI confidence, in turn, is one of the strongest single predictors of a team’s ability to innovate, ship, and retain talent through this transition. Which means the highest-return investment most companies aren’t making right now is around building the confidence of the managers who set the tone for everyone reporting to them.

Building that confidence means giving every employee access to agentic tools by default (not by request), giving them a shared workspace where they can see peers building alongside them, and giving them role-specific guidance for how their work is actually changing — not in the abstract, but this quarter, in their function. If that sounds familiar, it should. It’s the foundation the 4Rs framework was built on.

We’ve noticed a pattern, as have some of our customers; a “shadow version” of AI fluency creeping into workplaces that is worth keeping an eye on as we scale: Employees who forward an agent’s output along to a colleague without reading it themselves, treating a first draft as though it were finished thinking. Most of the time this isn’t laziness so much as a natural instinct to trust a tool that seems to know what it’s doing. Still, it’s worth naming and discussing, because it can hinder agentic progress – not to mention employee satisfaction. Open dialogue about these should be encouraged. It helps remind teams what real AI fluency actually looks like: judgment on top of speed, not speed instead of judgment.

A quick note on the 4Rs

In our first playbook, Becoming an Agentic Enterprise, we introduced the 4Rs: Redesign how work gets done, Reskill your people, Redeploy talent into growth areas, and Rebalance what humans and agents each do best. That framework is still the foundation of how Salesforce is approaching workforce transformation, and it’s still really relevant. If you haven’t worked through it with your team yet, check out check out Chapter 2: Design your AI-human workforce.

What’s changed since we published it is the texture: the everyday, in-the-flow-of-work examples of what the 4Rs actually look like when they’re running. As agents scale across the organization, tasks for humans will inevitably look different. The 4Rs will remain a guidepost for leaders in every department to align and aspire to.

Agentic enablement: an ongoing climb

Nobody becomes an agentic pro in a single training session. The organizations getting real leverage out of agents treat enablement the way a climbing team treats a mountain: as a steady ascent with distinct pushes along the way, each one building on the last. You start at the base by getting people set up on the tools. Then you help them see what AI actually looks like for their specific role. Then you go deeper on how the role itself is evolving. And at the peak, you shift the conversation from individuals to teams — how small, cross-functional, AI-enabled groups operate together.

The real work is figuring out which surge your people need next and meeting them there in the flow of work. Use the readiness check below to figure out where your organization is on the climb, and where to invest the next push or “surge.”

Try this activity: Enablement “surge” readiness check 

Get a gut check on whether your people are actually climbing the enablement mountain, or stuck at base camp with tools they don’t quite know how to use. Ask yourself these questions across the four surges below.

Welcome to the dawn of the Self-Evolving Enterprise

A little over 100 years ago, a French artist named Jean-Marc Côté made a series of lithographs he titled “En L’An 2000” (In the Year 2000), predicting what daily life would look like at the turn of the next century. He was working with the only visual vocabulary he had: brass telephony gear and hand-cranked machinery. And still, he got eerily close.

We’re going to do something similar in this chapter. No one has a crystal ball, Salesforce included. But we do have a decade of research in the space behind us. Salesforce AI Research has been publishing and shipping enterprise AI research for over ten years, with more than 250 million downloads of our models and datasets on Hugging Face. That vantage point, combined with the patterns we’re already seeing in production with our own customers, gives us a real view into what’s coming next.

The agentic foundation you’re building today is exactly what makes the next horizon reachable. 

Here’s how we see it unfolding, and what you can start doing now to be ready.

Three shifts to watch

1. How we’ll interact with technology: the agent becomes the interface

The phone in your pocket hasn’t fundamentally changed in almost 20 years. It’s still a grid of app icons, each one a separate destination you have to navigate to. Imagine a device with only two apps on it designed to communicate with friends and co-workers, for example. iMessage and Slack. Everything else: your calendar, your bank, your shopping, gets handled by a personal agent that has access to all of it in the background, mostly through voice. 

We’re already starting to introduce this model with a personalized AI teammate for every Slack user in Slackbot. It’s the direction Apple and Google are quietly building toward with new device concepts. The user interface as we’ve known it for the last three decades is starting to fade into the background, and personal agents are becoming the primary way work gets done. 

How to prepare: Think about where your employees and customers still hit friction with a traditional UI. As feedback comes in, consider how agents might reduce friction in the user flow. These are the places a personal agent is likely to land first. 

2. How we’ll make decisions: from guesswork to foresight

For most of the last two decades, business intelligence has been about looking backward. Dashboards, reports, quarterly reviews. Agents are turning our attention toward the future. Anyone in the company, not just the analytics team, is starting to have access to the power of foresight: the ability to model scenarios, compare outcomes, and make informed choices between options that a human alone couldn’t have surfaced in time.

And a growing share of those decisions are being made autonomously. According to our recent State of Agentic AI in the Enterprise 2026 report, 36%of agentic workflows are autonomous. This is not to say there isn’t a human at the helm; it means that humans are in the workflow alongside them, monitoring decisions made. This is the shift we’ve been calling operational intelligence: intelligence woven through the entire operation, informing decisions at every level, and increasingly making some of them on its own.

This doesn’t mean humans are excluded from decision making. Every agent will have at least one person accountable for it. If you own the agent, you own the decision. What’s changing is that the human at the helm is increasingly comfortable letting the agent make the call without checking in first.

How to prepare: Identify the decisions in your business that are currently slow, repetitive, or bottlenecked by human bandwidth. These are your first candidates for autonomous decision-making, with the right human accountability wrapped around them.

3. How we’ll drive innovation: agents that teach themselves

The third shift is one that has a major influence on how companies will innovate in the future. Agents are starting to learn continuously and evolve on their own through a method called recursive self-improvement. Frontier models are already accelerating past human speed in helping to design and develop their own successors. What that looks like inside an enterprise is agents that don’t just execute a fixed capability, but get better every time they run, grounded in your company’s proprietary knowledge, your workflows, your integrations, and the traces of every interaction they’ve already had.

This is the piece that turns a collection of agents into a Self-Evolving Enterprise. The company itself, not just the model layer, becomes the source of intelligence. And because every enterprise’s data is unique, that intelligence compounds into a real competitive advantage that no one else can copy.

How to prepare: Take stock of your proprietary data, especially the “dark data” you’re sitting on without utilizing: transcripts, traces, interactions, business logic. That’s the fuel your self-evolving agents will run on.

We’ll build this next chapter with you

The pace of change is the story now. Waterfall, the sequential, plan-everything-upfront approach born in 1970s manufacturing and formalized for software soon after, took roughly thirty years to become the default operating model of the enterprise. Agile, which broke that plan into two-week sprints and put working software in customers’ hands sooner, took another twenty to move from the 2001 manifesto through a generation of certifications, coaching, and org redesigns before it truly stuck.

Agentic didn’t get thirty years, or twenty. It arrived in months. The frameworks and change-management playbooks that carried companies through the last two transitions were built for a slower world, one where you could plan a quarter, run the pilot, form the committee, and still be early. The one we’re in now runs on a different clock entirely, where the cycle from idea to shipped, learning, evolving software can be measured in days.

We hear the anxiety in the industry. We choose to lead differently. We think this is a time of empowerment. It’s a chance to give every employee access to leverage that used to be locked away in a few departments, and to build companies where humans do the work only humans can do, while agents handle the rest.

This playbook is here as a symbol of Salesforce’s partnership with you. It’s designed to grow. We’ll keep updating it as we learn, as the workforce transformation gets more concrete, as the self-evolving enterprise moves from research to production, as our customers show us patterns we haven’t seen yet. We are on this journey too. And we will be here with you every step of the way.

Still piloting AI agents or want a refresher on the basics? Read this step-by-step guide first