Why Agentic AI Is the Enterprise's Next Wave

Every real wave in enterprise software has been a price collapse wearing the costume of a new capability. What has collapsed this time is the cost of handling an exception — which is why the wave is genuine, and also why so much of it is going to fail anyway.

An invoice arrives at a shared mailbox on a Tuesday and does not match. The purchase order says one quantity, the goods receipt says another, the vendor has attached a credit note referencing a contract amendment that lives in a different system entirely, and the three-way match in the ERP fails cleanly and quietly, exactly as it was designed to. From that moment the invoice stops being a transaction and becomes a small research project. Someone in accounts payable opens a ticket, someone in procurement is asked whether the amendment was ever loaded, someone in the receiving organisation is asked whether the goods actually turned up, and eleven days later the invoice gets paid, having consumed a couple of hours of attention spread across four people who each held one fragment of the answer. None of this is dysfunction. It is the system working precisely as intended: the rules covered the ordinary case, the ordinary case went through untouched at effectively zero marginal cost, and the case the rules did not cover went to human beings, because for the entire history of enterprise software that was the only place it could possibly go.

That routing rule — covered by a rule, handled by the machine; not covered by a rule, handled by a person — is the load-bearing assumption underneath nearly every system an enterprise has bought in the last forty years. It is so deeply assumed that most organisations have never costed it directly. They have costed licences, integrations, headcount and cycle times, and treated the flow of uncovered cases into human queues as weather rather than as a line item. It is, in fact, one of the largest line items they have, and it is the thing that has just changed price.

A wave is a price collapse, not a product launch

New capabilities appear in enterprise technology constantly and mostly change nothing. What produces an actual wave — the kind that reorganises budgets, org charts and the vendor landscape — is narrower and duller: the unit price of something the organisation already does an enormous amount of falls far enough that the organisation stops rationing it. Enterprise resource planning was a wave because it collapsed the cost of holding one consistent record across functions that had previously each kept their own. Software-as-a-service was a wave because it collapsed the cost of distributing and operating an application. Cloud was a wave because it collapsed the cost of provisioning capacity, and in doing so quietly killed the annual argument about how many servers to buy. In every case the thing people pointed at was a product, and the thing that actually moved the industry was a price. The reliable test for whether you are looking at a wave or a demo is not how impressive the artefact is; it is whether organisations changed shape around the new price, and how quickly they stopped noticing they had.

By that test, most of the last decade of enterprise machine learning was a capability rather than a wave. Classification and prediction became dramatically cheaper and better, and they were deployed almost entirely inside the existing routing rule. A model scored the transaction, or ranked the queue, or flagged the anomaly — and then handed the uncovered case to a person exactly as before, sometimes with a helpful summary attached. The routing survived intact, which is why the productivity numbers so often disappointed people who had seen genuinely remarkable demonstrations. The demonstration was real. The price of an exception had not moved.

What is different about the current generation is specific and worth stating without embellishment: a system can now take a case nobody anticipated, work out what it is asking, assemble the context from wherever that context happens to live, reason about it against the organisation's own policy, and carry it toward a resolution rather than toward a queue. That is not a better classifier bolted onto the same pipeline. It relocates the boundary that every enterprise architecture has been drawn around. It is, on the evidence, a genuine price collapse, and it is why the people calling this a wave are not simply repeating themselves.

The exception was the last expensive thing left

The reason this particular price matters more than it sounds is that the covered path was eaten decades ago. Straight-through processing, workflow engines, robotic process automation and integration middleware between them took most of the volume out of human hands, and what they left behind was the residue: the cases that fell out. Anyone who has ever costed a process end to end knows the shape that results. The exception rate looks small when expressed as a percentage of transactions and enormous when expressed as a share of the labour, because a covered case costs almost nothing and an uncovered one costs a chain of people, several context switches, an ageing ticket and a decision made by whoever happened to be available rather than whoever knew most.

It also helps to be precise about why exceptions were expensive, because the answer is not that they were intellectually hard. Very few of them are. They were expensive because the information needed to resolve them was dispersed — a fragment in the contract repository, a fragment in the ledger, a fragment in an email thread from March, a fragment in the memory of someone who has since changed roles — and the human being in the middle was functioning as the index. The cost of an exception was overwhelmingly the cost of reassembling context, plus a small amount of genuine judgement at the end. The judgement was the part everyone talked about. The reassembly was the part everyone paid for. Reassembling dispersed context against a body of enterprise knowledge is exactly the thing that has just become cheap, which is why the collapse is real and not merely announced.

Cheap to handle is not the same as answered for

None of which means the deployments will work. Gartner has predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, and warning separately about "agent washing" — existing chatbots, assistants and rule engines relabelled as agents without the substance underneath changing at all. It is tempting for anyone selling in this category to treat that forecast as a comment about immature buyers or about other people's products. It is more useful, and more accurate, to read the three named reasons as one reason wearing three hats.

Consider what the old routing rule was actually doing, beyond moving work. When an uncovered case went to a named person, that routing did not just allocate labour; it allocated accountability, silently and for free. The accounts payable clerk who released the mismatched invoice owned that release. If it was wrong, there was a person whose judgement it had been, a manager who had a view about it, and an audit trail that ended somewhere specific. Nobody designed this. It was a by-product of the fact that only humans could handle uncovered cases, and humans come with accountability attached whether or not you ask for it. Remove the human from the exception path and the labour transfers cleanly while the accountability does not transfer at all — it simply evaporates unless somebody deliberately rebuilds it. That is what the three failure modes are describing. Costs escalate because no one owns the scope of what the agents are allowed to attempt. Business value stays unclear because no one owns a number that was supposed to move. Risk controls stay inadequate because no one has decided, in advance and in writing, who is answerable when an autonomous resolution turns out to be the wrong one. Agent washing thrives in the same vacuum, because the only test that reliably separates a real agent from a relabelled rule engine is what happens on the case nobody anticipated — and an organisation that has not decided who is on the hook for that case has no particular reason to run the test.

This reframes what human oversight is for. Treated as a safety blanket, a human-in-the-loop step is a tax you try to minimise, and the natural instinct is to remove gates as confidence grows until one day the interesting decisions are being made by something with no name attached. Treated properly, the loop is an accountability register: a small, explicit set of points where a decision is handed to a person because that person is the one who will answer for it, chosen for that reason rather than because the model was uncertain. That is the design question buried inside platforms like StudioX, where autonomous workers run missions across systems while observations and human gates mark the places where judgement is deliberately retained — and the hard part of the deployment is never the agents. It is the argument inside the customer about which decisions belong on the register and whose name goes next to them. Much of the emerging literature on what an autonomous enterprise actually requires converges on the same unglamorous conclusion, which is that the governance work is the work, and that organisations which skip it do not get autonomy cheaply — they get an expensive pilot and a quiet cancellation.

So the mental model worth carrying is not that agents automate exceptions. It is that an enterprise has always had two distinct budgets: the cost of handling the cases nobody wrote a rule for, and the capacity to answer for how they were handled. For forty years those two budgets were fused, because the same person did both, and no one had to think about them separately. The wave is real because the first budget has genuinely collapsed. The failures will come from organisations that believed the second one collapsed with it, and discovered — usually during an audit, occasionally during a lawsuit — that answerability was never a cost of processing. It was the price of being allowed to make the decision at all, and it has not gone down by a single point.