Embrace the Mess

Every large organisation I work with has a pilot graveyard.

Picture the steering committee. Twelve proofs of concept, all of them demoed, all of them applauded. A contract summariser, a service-desk assistant, an agent that drafts tender responses. Eighteen months later, none of them is in production. Nobody killed them. They were simply asked to wait: for the data clean-up, for the platform decision, for the governance framework, for the review board. By the time the wait was over, the business sponsor had moved on, and the model they were built on was two generations old.

The models are good enough, and have been for a while. What holds these pilots back is a reflex: the large organisation’s instinct to establish order before it allows value. That reflex served us well in finance and in safety. With AI, it is the single biggest reason pilots never scale.

The pilot graveyard

The numbers are now familiar. MIT’s NANDA initiative reported in 2025 that around 95% of organisations saw no measurable return from their generative AI investments. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The usual explanations are cost, risk and immature technology.

In my experience, most of the pilots that die worked in the lab. They die when they are asked to meet the organisation’s conditions for scale.

The cleanup fallacy

The first condition is always the data. “Before we do AI, we need to get our content in order.”

I have heard this sentence for thirty years. In the Documentum era, it justified multi-year taxonomy programmes. With SharePoint, it justified migrations that would finally give us one source of truth. Each wave promised that this time the information would be clean, classified and findable. None of them delivered it, because large organisations produce working information: drafts, versions, exceptions, emails that contradict the policy, and the policy that everyone knows is out of date.

A clean-up programme in a large organisation has no end date. Making AI wait for one means it waits forever.

The bane of platform thinking

The second condition is the platform. Nobody wants a tactical solution. Everyone wants the unified, enterprise-grade AI platform that every future use case will run on.

Two arguments carry this. The first is “let’s not reinvent the wheel”: why build ten point solutions when one platform could serve them all? The second is IP protectiveness: our data is too valuable to touch someone else’s tools, so we must build our own, inside our own walls.

Both sound responsible. In practice, they guarantee that nothing useful reaches the business for a year or more. The platform needs requirements, so it needs use cases. The use cases need a platform, so they wait. The internal chatbot built to protect the IP arrives late, a generation behind what employees use at home, and they notice.

Platforms are worth having. But starting with the platform is a near-certain way to lose the short- and medium-term interest of the business stakeholders who fund the whole thing. The better order is the reverse: ship tactical wins, watch which patterns repeat, and let the platform grow out of what proved useful.

The human-in-the-loop trap

The third condition sounds the most reasonable: a human must review everything the AI does.

For a single pilot, that works. At enterprise scale, it collapses. Every agent produces proposals, drafts, approvals and escalations, and every one of them lands in someone’s queue. The small group of experts trusted to review them becomes the bottleneck. The oversight meant to make AI safe ends up moving the work to another desk.

It is also bad for the reviewers. In “The Last Interface Is Language” I wrote about what close supervision of AI does to the people doing it. A 2026 Boston Consulting Group study found that intensive AI oversight was the most draining way to use it, and that exhausted overseers made more major errors. The psychologist Lisanne Bainbridge predicted this in 1983: the more you automate, the harder the human’s remaining job becomes. Tired reviewers skim and then rubber-stamp, so a blanket human-in-the-loop rule mostly produces the appearance of control.

The impulse is sound, but it needs limits. Oversight should be proportionate to risk: human judgement where the stakes call for it, automated checks everywhere else. The technology is moving that way. Models such as Jev, launched by TypeSafe AI in September 2026, don’t generate text at all. They return structured decisions, such as a choice, a score or a yes or no, with a confidence value attached. That is what routine triage needs: let the machine decide the obvious cases, and send people only the ones it is unsure about. It is early, and the claims are the vendor’s own. But it points to a future where human attention is spent where it counts.

Built for mess

For seventy years, computers demanded order. Structured fields, controlled vocabularies, exact syntax. If information was messy, the machine couldn’t use it, so we spent fortunes cleaning it up.

Language models are the first technology built the other way around. I am a linguist by training, and what strikes me most is that these models handle ambiguity the way people do. They read the outdated policy and the email that contradicts it, and they can tell you that the two disagree. They don’t need a taxonomy to do it.

Insisting on clean data before using AI throws away the one advantage that makes it different from everything that came before.

What embracing the mess looks like

In practice, it comes down to a handful of habits:

Start where the work is. Deploy into the real environment, with the real documents.

Go tactical first. Solve one painful problem for one team, fast. Fund the platform from what repeats.

Measure outcomes. Ask whether the work got better.

Make oversight proportionate. Decide per use case where people must decide, and automate the rest of the checking.

Let structure emerge. Usage shows you which data matters, which processes need redesign and which controls are needed. That is where clean-up should start.

But what about…

“We can’t risk it with our data.” Some data, no. Much of it, yes, and with the right contracts, the leading providers now offer stronger controls than many internal environments. Protecting IP by blocking every tool mostly protects it from being used.

“Without a platform, we’ll end up with a hundred point solutions.” Probably fewer than you fear, and they will teach you what the platform actually needs. A platform designed before anyone has used AI in earnest is designed on assumptions.

“Regulators expect a human in the loop.” For some decisions, rightly so. But “where the law requires it” is a much smaller set than “everything”. The EU AI Act asks for human oversight in proportion to risk.

Where the mess stops

Embracing the mess still has limits. Security, privacy and accountability are non-negotiable. Regulated decisions, such as those about credit, health, employment or legal rights, need real human judgement and an audit trail. Somebody must own every AI system in production. Keep the controls, and stop making every one of them a precondition for starting.

Order is the outcome

Large organisations will not become tidy before they adopt AI. They never have, and they never will. The ones that get value are the ones that start in the mess, learn from it, and let order emerge from use.

Order is the return on the investment.