The agent works. The workflow does not.

The demo went well. It usually does. Somebody pasted in a messy customer email, the agent read it, pulled the order, drafted a refund and wrote a reply that was better than the one the team writes at 5pm on a Friday. Everyone in the room agreed this was going to save a lot of time.

Six months on, nobody can find the saving. The team still processes the same number of cases a day. The agent is running. The invoice for it is running too.

Two numbers from the same month

McKinsey's State of AI survey, published in August 2026 across 1,719 executives in 97 countries, defines an "AI high performer" as an organisation attributing at least 5% of EBIT to its AI use. That group is about 6% of respondents.

Anthropic's 2026 State of AI Agents Report, drawn from more than 500 technical leaders, found 80% reporting measurable ROI from agents.

Both can be true at once, and that is the interesting part. Local savings are real and easy to see: a team of four handles what five used to. Getting that to the bottom line requires something to change downstream — a vacancy not filled, a queue that stops needing a supervisor, a customer who renews because the response came in an hour instead of a day. Most agent programmes never arrange for that second thing to happen.

The pilot flatters the agent

A pilot is designed by the person who wants it to work, and it inherits four advantages that production will take away.

One happy path. The demo case is the case somebody chose. Real intake is 40% of cases that do not match the shape anyone described, and the interesting question is not how well the agent handles the 60% — it is what it does when it cannot.

A human watching. During a pilot, someone reads every output. That person is a safety system nobody accounted for, and they are also the reason the error rate looked acceptable. Remove them and the error rate does not change; the consequences change.

Clean data in one system. The pilot ran against an export. Production runs against a CRM where the same customer exists three times, a ticketing tool with a different ID scheme, and an accounts package that will not let anything write to it after month-end close.

No accountability. Nobody asked who is responsible when an agent issues a refund it should not have. That question does not have a technical answer, and it stops deployments dead when it surfaces at the end rather than the start.

Bolting it on changes nothing

Here is the finding worth pinning to a wall. In the same McKinsey survey, nearly three-quarters of AI high performers said they had fundamentally redesigned workflows because of AI — against one-quarter of everyone else. Of all the factors tested, intentional workflow redesign made one of the strongest contributions to business impact.

Redesign is a precise word here and it is not a synonym for adoption.

Take a six-step process: request, triage, agent drafts, human checks, approve, close. The agent made step three faster. Steps four and five did not move, and they are where the days were. Work does not sit still because somebody is typing slowly; it sits still in queues, waiting for a person who has other things on.

Redesigning means asking what the process would look like if the agent were reliable at that step — then removing the steps that only existed because it was not. That is an uncomfortable conversation about approval limits, about who is trusted with what, and about a supervisory role that may no longer be needed in its current shape. It is a management decision wearing a technology costume, which is exactly why it gets deferred.

The boring part is the part that fails

Ask technical leaders what is actually blocking them and the answer is not model quality. In the Anthropic report, the top three obstacles are integration with existing systems (46%), data access and quality (42%) and change management (39%).

"Integration" sounds like a connector. In practice it is five unglamorous pieces of engineering:

An identity of its own. The agent needs its own service account with its own permissions, not a borrowed admin login. Nothing is auditable if every action in the log belongs to whoever set it up, and nothing can be revoked in a hurry either.

Permissions narrower than a human's. Read the order, draft the reply, refund up to a limit, escalate above it. Scope is the control that lets you deploy at all, and it belongs in the system being called — not in the prompt. A prompt is not a permission.

A trace for every action. What it did, which record it touched, what it was working from, and which version of the instructions was live. Six weeks later somebody will ask why a particular customer got a particular answer, and "the model decided" ends the conversation badly.

Writes that survive a retry. Agents retry. Networks fail mid-call. If the same refund can be issued twice because a retry produced a second write, that is not an AI problem — it is the same idempotency discipline any integration needs, and skipping it is how a pilot becomes an incident.

A defined failure path. What happens when the agent does not know. "Escalate to a human" means a queue, an owner and a response time, or it means the request disappears.

None of this is exotic. It is ordinary systems work, and it is the reason an agent that demos in a fortnight takes a quarter to put into production. Teams underestimate it because the model did the impressive part first.

What Gartner actually said

The most-quoted figure in this whole debate is Gartner's: over 40% of agentic AI projects will be canceled by the end of 2027. It gets repeated as proof that legacy systems kill agents. Read the release and the reasons given are escalating costs, unclear business value and inadequate risk controls — three management failures, not an infrastructure one.

The same release names the other trap: agent washing, where existing chatbots, assistants and RPA scripts are relabelled as agents. Gartner reckoned only around 130 vendors of the thousands claiming the category were doing anything genuinely agentic.

The buyer's test is simple enough. Ask what the thing does when the answer is not in its instructions, what it is permitted to change without asking, and what happens to the audit trail. Rebranded software answers those badly, because it was never built to do anything on its own.

The quieter shift underneath all this

One number in the McKinsey survey did not get the attention it deserved: 32% of respondents said their organisation decided against buying one or more software products or features because they could be built internally with agentic coding tools.

For a mid-sized business that is a genuine change in the economics. Tools that were only ever worth licensing — a small internal portal, a reconciliation job, a customer-facing form with three integrations behind it — are now worth building, owning and changing when the process changes. The constraint is no longer typing speed. It is whether anyone on the team can specify the work precisely, review what comes out, and run it safely afterwards.

That last part is where this connects back to the top of the post. The organisations getting EBIT impact are not the ones with better models. They are the ones who redesigned the work, wired the agent properly into the systems that hold it, and kept someone accountable for the result.

Where to start

Not with a tool. With one process you can name, that runs at least weekly, where you already know roughly how long it takes and how often it goes wrong.

Write down every step, who does it, and how long each waits before someone picks it up. The waiting is the finding — that is almost always where the time is, and no agent improves a queue by drafting faster into it. Then ask the harder question: if the drafting step were reliable, which approvals would you be willing to remove? If the answer is none, the project will not pay for itself, and it is better to know that in week one.

We do the integration and build side of this — service identities, scoped permissions, audit trails, the systems plumbing an agent needs to touch production safely — as part of our IT and software services. The process side, the redesign conversation that nobody wants to chair, sits in our business consultancy work. And if the gap is that your own team needs to build and review this work rather than outsource it, that is what our AI development training is for.

If you already have an agent running and cannot find it in the numbers, the useful first conversation is about the process, not the model — tell us which one it sits in.