MIT’s NANDA initiative puts the problem most clearly in its 2025 report, The GenAI Divide: State of AI in Business 2025 — a title that should give any executive approving an AI budget pause. Its main finding is hard to ignore: among the companies studied, about 95 percent had seen no measurable return from generative AI, meaning little of that investment had shown up in the profit-and-loss statement.
The report deserves to be read with its limits in mind. It labels itself preliminary and rests on a modest evidence base: 52 structured interviews, 153 survey responses, and a review of roughly three hundred publicly disclosed deployments. It has also been fairly criticized because “success” can mean very different things from one company to another. Still, the headline is difficult to dismiss, especially because other major studies point in the same direction. McKinsey’s 2025 survey found that while most companies now use AI somewhere in the business, only about 39 percent could show any enterprise-level profit impact, and that impact was usually below five percent. BCG put it more bluntly: about 60 percent of companies had seen no material value from their AI spending, while only around 5 percent were capturing value at scale. The takeaway is clear: the money is being spent, the ambition is real, and AI adoption is spreading fast but true business transformation remains rare.
The report’s deeper point is that the divide is not mainly about model quality or regulation, even though those are often the easiest explanations to reach for. The bigger question is whether the technology has been woven into the way people actually work. The report calls this a learning gap. A generic chatbot can certainly help an employee in the moment, but it does not learn the company’s routines: the approval path, the exception rules, or what changed last week in the systems around it. More usage alone does not make it better at doing a specific job. It can answer questions, and that has value. But it still leaves the real work sitting with the people around it.
This is the divide the report is pointing to. On one side are the many companies where AI is widely adopted but has not really changed how the business works. People may have a chatbot open in another tab and use it throughout the day, but the underlying processes, decisions, and handoffs remain much the same. On the other side is the smaller group that has managed to close the gap. What is revealing is how they got there: not usually by building everything from scratch, but by working with partners who could embed the technology into a specific, high-value workflow and help it get better over time.
The lesson is not that AI is all hype. It is that using AI and transforming a business with it are not the same thing, and most companies have only taken the first step. That leads to the next, more practical question: what does the technology look like when it truly helps a company change how work gets done?
From chatbots to agentsThe tools that stalled all had the same basic limitation: they talked, but they did not move anything forward. You asked a question, they gave an answer, and the work still landed back on the person at the keyboard. Each new chat started from scratch. Someone still had to check the response against policy, enter the details into the right system, follow up with the next person, and make sure nothing fell through the cracks. That can be useful, sometimes very useful. But it is assistance, not delegation.
Agentic AI marks the shift from software that simply responds to software that can actually act. Instead of answering one question and stopping, an agent is given a goal and works through the steps needed to complete it: pulling data from one system, applying the company’s rules, using another tool, checking its own work, and returning a finished task rather than a suggestion. That matters because it goes straight to the failure the MIT report identifies. A well-built agent keeps the context a generic chatbot loses at the end of every exchange: the workflow it belongs to, the state of the task, and the rules it must follow. It can improve over time, too, but only when that learning is deliberately built in through feedback, evaluation, correction, and governed updates. Left alone, an agent does not magically teach itself. Designed properly, it stops asking people to rebuild the same context every time they need help.
Consider a bank's customer service. The common chatbot answers “how do I dispute a charge?” with a paragraph of instructions and leaves the customer to do the rest. An agentic version, given the goal of resolving the dispute, can look up the transaction, check it against the fraud rules, open the case in the bank's own system, notify the customer of the reference number, and escalate to a human only when something genuinely requires judgment. The first tool describes the process. The second one runs it.
Or take procurement. A general AI tool can summarize a supplier contract if you paste it in. An agent pointed at the procurement inbox can read each incoming quotation as it arrives, compare it against the approved vendor list and budget ceiling, flag the two clauses that deviate from the company's standard terms, and draft the response for a manager to approve. The afternoon's work becomes a quick review of the exceptions. Nobody had to prompt it question by question; it was given the job and did the job.
Finance offers a third example. Month-end reconciliation sounds routine until you watch a team lose hours matching thousands of transactions across systems and chasing the few that refuse to line up. It is repetitive, rules-heavy work, but the exceptions still require judgment. An agent can do the matching continuously, separate the real problems from the noise, and present each exception to an accountant with the likely cause already identified. The team stops spending its time hunting for mismatches and starts focusing on the decisions that actually need human judgment.
All three answer the diagnosis the same way. None is a smarter chatbot; each is embedded in a specific workflow, operates on the company's own data and rules, and produces a finished outcome rather than advice. That is adoption turning into transformation, made concrete.
That said, not every step in these examples deserves an agent and the excitement around agents often runs ahead of the actual need. If a fixed rule or deterministic script can do the job reliably, that is usually the better answer: cheaper, faster, simpler, and easier to trust. In the reconciliation example, for instance, most transaction matching may belong to ordinary automation, not an agent. An agent earns its place only when the work genuinely requires judgment across several steps: interpreting an unusual case, choosing which system to touch next, or adapting when the inputs do not fit the template. Using an agent where a rules engine would do the job is not innovation. It is just expensive automation with better marketing.
What “doing it right” looks like in practiceTo see what this looks like in the real world, it helps to look at a company built around the very thing the studies say separates the winners from the rest: going deep into a client’s operations instead of simply handing over a tool and hoping the organization can figure out the hard part later.
Palantir is the clearest case. Its public image is often wrapped up in defense and intelligence work, but the more useful lesson is in how it delivers. Palantir does not simply sell software, hand over the keys, and leave the customer to struggle through implementation. It embeds its own engineers — forward-deployed engineers, in the company’s language — inside the client’s organization for weeks or months at a time, usually in small teams. Their job is not to write a slide deck full of recommendations. It is to build working software against the company’s real data, learn the awkward details of how the business actually runs, and stay close enough to the work until the system is useful in production.
The real lesson is not the software itself, but the way the work is divided — and that division mirrors the chatbot-versus-agent distinction almost perfectly. People still do the hard, messy, judgment-heavy work: untangling the current business process, deciding what should change, and translating the company’s data, rules, and exceptions into something a system can use. Once that work is done, the narrower and more repeatable tasks can be handed to AI agents, connected to the workflow through Palantir’s AI platform with governance and audit controls built in. In other words, humans redesign the process; agents carry out the specific jobs inside it. Each does the work it is actually suited to do.
The point of looking at Palantir is specific: its delivery model makes the implementation problem impossible to miss. It shows, in operating-model terms, what the studies are really describing — transformation that comes from working inside the workflow, not from buying a generic tool and hoping value appears later. One company, of course, does not prove the entire thesis. But it does offer a particularly clear case, and the pattern is hard to ignore. The model has become influential enough that other major AI firms are now building their own embedded-engineer teams around the same idea: if the value lives in the workflow, the people building the system need to be close to the workflow too.
This is not a model to romanticize. Two caveats are hard to ignore. First, it is expensive and slow by design: embedding engineers inside a company for months does not scale like selling software licenses, and many organizations simply cannot fund that kind of bespoke engagement. Second, the same depth that makes the model effective also creates dependency. Once software, data models, workflows, and operational knowledge are tightly connected, replacing the system becomes difficult and costly. That may be a side effect of deep integration rather than the goal, but the result is the same: the customer becomes more reliant on the vendor. Both truths matter. The model works because it goes deep — and going deep is exactly what makes it hard to unwind.
So the transferable lesson is not “hire Palantir.” It is simpler, and more important: the companies crossing MIT's divide are not treating AI transformation as something a vendor can deliver fully formed. Whether they build, buy, or partner, they are embedding it inside the business, close to the workflows, data, rules, and people that determine whether value appears at all. That is what separates adoption from transformation. The winners are not merely adding AI to existing work; they are redesigning the work so AI can change the outcome.
Agents don't remove the riskNone of this means agentic AI is guaranteed to work — and any serious argument for it has to say that plainly. In fact, the risk case is strong enough to take seriously. In June 2025, Gartner predicted that more than 40 percent of agentic AI projects would be canceled by the end of 2027, undone by rising costs, uncertain business value, weak risk controls, and models that are still too immature for some of the autonomous work being assigned to them. That forecast does not undercut the case for agents; it clarifies the terms. Agentic AI can create value only when the work is well scoped, the economics make sense, and the organization is mature enough to govern what it has set in motion.
The two figures are making different kinds of claims, and that distinction matters. NANDA’s 95 percent is an observation drawn from the organizations in its dataset, not an audit of every company’s books. Gartner’s 40 percent is a forecast about projects that have not yet reached their outcome. But the two warnings point in the same direction. Both suggest that the real fault line is not the model itself, but everything around it: the economics, the workflow fit, the governance, and the discipline of implementation. Gartner is more cautious about the technology, warning that today’s models are still not mature enough for some autonomous work. Yet neither finding lets companies put most of the blame on the models. The harder truth is that AI more often fails or succeeds because of the organization built around it.
Gartner also points to a trap that brings the whole argument back to the chatbot problem. It calls the practice “agent washing”: vendors relabeling ordinary chatbots and older automation tools as “agents” without delivering the autonomy, workflow integration, or learning loop that would make an agent worth paying for. By Gartner’s estimate, only about 130 of the thousands of firms claiming to sell agentic AI are selling the real thing. That matters because an agent-washed chatbot does not move a company across the divide; it simply repackages the same weak implementation model at a higher price. The buyer may think it has purchased transformation, but in practice it has bought another interface. The lesson is not to avoid agentic AI. It is to demand proof that the system can do more than talk — that it can act, learn, and create value inside the workflow where the business actually runs.
The response is not glamorous, but it is where the real work begins. Agents that take meaningful action need serious guardrails: clear limits on what they can do without human approval, an audit trail for every decision, and especially careful handling wherever money, personal data, or regulated outcomes are involved. A system that can resolve a customer dispute can also resolve thousands of them incorrectly if the rules, data, or escalation paths are weak. That is why the lesson from successful implementations is not to unleash autonomy and hope for the best. It is to design autonomy carefully — with tight scope, strong integration, visible controls, and human oversight where judgment still matters.
That is what successful AI implementation looks like in practice: not a flashy tool everyone chats with, but a focused system that quietly takes responsibility for a real piece of work. It is scoped tightly, governed visibly, measured honestly, and improved deliberately over time. That is how AI stops being a novelty in the workflow and starts becoming part of how the business actually performs. In the end, the companies that win will not be the ones with the most AI, but the ones disciplined enough to turn it into better work.
The harder questionSo far, the argument has been written for companies trying to adopt AI. But it raises an even sharper question for a different kind of reader: the company that wants to sell this transformation — the B2B provider promising to help other firms cross the divide. If AI value is created inside workflows, then the real challenge is not selling a smarter tool. It is proving that you can enter a client’s messy operations, understand how work actually happens, and help turn AI from a promise into measurable business value.
The answer is demanding because the real product is not the software alone; it is the ability to deliver change inside a client’s operating reality. If the business depends on deep, custom transformation across complex clients, the capability quickly starts to look like the forward-deployed model itself: engineers who can embed inside a client’s operation, understand an unfamiliar business fast, translate messy data and informal rules into working systems, and keep shipping until the solution survives production. And there have to be enough of them to do it again and again, not just once for a flagship customer. The bar rises with the complexity a provider promises to solve. For the kind of transformation that actually moves the numbers, it is high. This is not a challenge a polished platform solves on its own; it is an organizational muscle, expensive to build, slow to scale, and very hard to fake.
And even that is not enough. Gartner’s forecast — that more than forty percent of agentic AI projects could be canceled by 2027 — is the clearest reminder that capability alone does not guarantee results. Even strong teams can lose control of the economics, chase value that was never clearly defined, or move faster than their governance can support. The point is not that capability does not matter; it matters enormously. But it is only the price of entry. Success still depends on disciplined scope, measurable business value, and the controls to keep the system trustworthy once it starts acting in the real world.
That is why the obvious shortcut is so tempting — and so flawed. A company can imagine hiring a Palantir-style partner to build the solution, operate it for a while, and then hand over the capability at the end through a neat build-operate-transfer model. But the very thing that makes the model valuable is the hardest thing to transfer. Code can be documented, systems can be handed over, and playbooks can be written. What cannot be moved so cleanly is the judgment a team develops by wrestling with a hundred messy implementations: knowing which exception matters, which workaround will break later, which stakeholder constraint is real, and which “simple” process is actually full of hidden dependencies. That knowledge can move, but only slowly — through paired delivery, apprenticeship, repeated joint implementation, and time inside real client work. The deliverable was never where most of the value lived. The value lived in the learning accumulated along the way, and learning is not something a company can simply check off on a transfer schedule.
The conclusion brings the argument full circle. AI transformation is not a finished product a company can simply buy; it is operational work that compounds inside the business over time. The same is true one level up: the capability to deliver that transformation cannot be purchased fully formed either. It can be built, strengthened, and even transferred in part, but only through delivery — through real projects, difficult trade-offs, mistakes, corrections, and repeated contact with how businesses actually run. That is why the long-term winners may not be the companies that adopt the most advanced AI first, or the providers with the most impressive demos. They will be the ones that learn fastest, turning each implementation into deeper capability and each deployment into a better way of working. The lasting advantage will not belong to the companies that automate the most work, but to the ones that learn the most from the work they automate.
Komentar (0)