Companies make a mistake when they discuss AI strategy as though agentic AI were a single capability to deploy uniformly across an organization. It isn’t.
It’s an understandable mistake. Watch a swarm of AI agents write code, then test and fix their own errors, and it’s easy to conclude that the same process suits the entire enterprise. The challenge is that AI excels at some types of work and is unreliable at others — sometimes dangerously so. The agentic AI that suffices for one process might be a poor fit in another.
Coding suits AI because it has clear rules and tests that confirm correct output, and that feedback is how models got good at it. Ask the same agent whether to cut the price of a struggling product, and it will answer just as confidently even if it lacks the data it needs to make an informed decision. Nothing will confirm it’s right before you need to act.
Fortunately, the agentic unevenness is predictable. Two questions explain where an agent is likely to excel or struggle in an enterprise:
- Has the LLM been trained on enough relevant data to be able to work its magic?
- Can you verify the answer to ensure it’s correct?
The answers aren’t binary and fall along a spectrum. Where a task lands on those two dimensions helps determine how companies should divide work between humans and AI.
Sometimes, AI agents can perform so reliably that a person’s main job is to delegate the tasks to AI and audit the results. Elsewhere, employees must capture missing context, curate AI’s output, or make and own the decision themselves.
Answering those two questions reveals four broad categories of tasks, each calling for a different role for humans. Below, we look at each in turn.
High verification and abundant context: Let the machine do it — but check its work
The first category starts with work where AI is familiar with the problem to solve and has an objective way to judge the result. Coding is the obvious example, but the same logic applies to many scientific and engineering problems. Ask an agent to identify a new material that stays pliable when frozen but doesn’t melt below a specified temperature, and the result can be tested against those requirements.
This is where AI’s gains compound. The same loop can extend beyond code, as expert rubrics and data on what people prefer and do give models a way to check work that technically has no single right answer. As models improve this way, agents will be able to take on more of it, with enormous potential for speed and productivity across an enterprise. The human role changes accordingly; instead of executing every step, people decide what to delegate and put checks in place to ensure the work is done correctly.
However, the risk runs in both directions. Delegate too slowly, and competitors may move faster. Delegate too early, and you can automate mistakes just as efficiently as successes.
Yet, a fair amount of work isn’t like that. Often, one of those two conditions is a problem, and sometimes both are.
High verification and missing context: AI could do this if only it knew what we know
A different problem arises, as this category suggests, when the work remains highly verifiable but the AI has not been trained or given feedback on tasks like this.
Consider getting a critical order to a customer on time. Whether it arrives by Friday is easy to verify. But an experienced employee may know that one supplier routinely misses its stated turnaround time, another will expedite an order if you call the right person, and a third can make a late shipment if given enough warning. None of that information may appear in the formal process.
Much of that knowledge lives in Slack threads, hallway conversations, and inside the heads of experienced employees. Why did the team skip a step in one market? Why did a salesperson make an exception for one customer but not another? Why did a manager ignore the standard process this time?
The human job is to stitch together what is known tacitly or from experience and to make that context available to the AI. Head-mounted cameras on service workers now record how the job actually gets done. Mercor, among others, has built a business interviewing experts and selling what they know to the AI labs.
Organizations that excel at turning tacit knowledge into usable context can improve their agents in ways competitors cannot match, even if they purchase the same AI model. Preserving that advantage, however, means preventing proprietary data and decision patterns from training external models.
Low verification and abundant context: AI can give a pretty good answer — and that’s the problem
A different risk appears when AI has plenty of relevant context but there is no objective way to verify that its answer is right, at least not in time to inform the decision. Much of marketing, positioning, and strategy falls into this category.
Ask an agent for a new marketing campaign and it may produce something polished enough to impress the team. But because it is drawing on past examples, it may produce work that feels new without being especially original. Two rival firms asking similar questions of a model may each get a confident, thoughtful answer that feels distinctive — only to find six months later that their campaigns, messaging, and perhaps strategies are starting to converge.
The human role here is to curate — to push beyond the generic, preserve what is distinctive, and decide when the AI’s answer is simply too familiar. Here is where human taste matters: the ability to recognize what feels distinctive, surprising, or true to the organization rather than merely competent.
The question is not simply whether the output is good, but whether it feels right for us. Does it reflect what customers value about the company? Does it sound like something our closest competitor could just as easily say? Does it offer something genuinely new or merely repackage what has worked before?
And when the eventual outcome can be measured, companies can use that evidence to sharpen what they look for the next time, potentially moving similar work toward a more verifiable category.
The risk here is not so much bad output as it is competent-sounding, unoriginal output that gradually makes companies and their ideas look and feel too similar.
Low verification and missing context: Nobody knows, somebody still has to decide
In this category, AI is least reliable when operating autonomously and there is no objective way to test its answer or little relevant precedent to draw on.
Imagine a company considering an acquisition in a market it has never entered. AI can analyze the financials, find comparable deals, and model different outcomes. None of that, however, can tell management whether the acquisition will strengthen the company, distract it from what it does best, or prove a costly mistake five years from now. If the situation is genuinely novel, there is no reason to assume the model will be able to generalize well from the sparse data like this it might have encountered during training. By the time there is enough evidence to know, the decision will have been made.
That doesn’t make AI useless; it simply changes its role. AI can surface assumptions, identify precedents, generate alternatives, and make the strongest case against the favored option. It can help executives think through a decision and its tradeoffs without pretending to settle it.
At some point, analysis gives way to judgment about what risks to take, which tradeoffs to accept, and what best reflects the company’s values. A human still has to make the call and own the consequences.
Match the human role to the work
For leaders, the takeaway is not to ask simply, “Where should we deploy AI?” but “What kind of work is this, and what should humans do here?”
The four categories outlined above are necessarily a simplification. Verification and context are matters of degree, and work can move from one category to another as more context and more durable rubrics become available. Companies learn to look for next time based on past outcomes and better ways emerge to check AI’s answers.
The point is not to force every task neatly into one of four boxes. It is to keep asking the two questions and to adjust the human role accordingly. AI is jagged. Enterprise strategy should be built around that reality, not pretend it away.
Learn more:
- Read the latest issue of the Futures magazine
- See if your business is ready for the advent of personal AI agents
