Your AI pilot is making a make-or-buy decision before anyone admits it

The word pilot gets embarrassing once people have started depending on it.

The word stays on the slide because it keeps everybody’s stakes low. Out in the work, the thing has learned your product names, your old exceptions, your customer language.

Then somebody notices the usual person is on leave and says, almost casually, “Use the pilot.

It knows.“

You weren’t careless for trying something small. Small things are how a busy organisation learns anything at all.

You’ve got queues to clear, customers to answer, people waiting on a decision, and a budget that didn’t arrive with a research lab attached.

A team that tests a useful tool isn’t secretly voting to give its work away.

It’s trying to make Tuesday possible.

Put the science fiction down for a moment.

Something plainer happened.

Your experiment became a boundary decision before anyone found the meeting room to call it one.

2018 — Your team buys a tool for one presentation. The files are somewhere safe, it assumes.

2022 — The trial account gets renewed, because the files are there and the presentation has become a routine.

2024 — The AI pilot gets real customer data. It drafts better, because the clean example was never the real work.

2026 — Nobody can say when it stopped being optional.

Your pilot already chose

The older name for this is make or buy. It doesn’t ask whether a supplier is clever enough, or whether your team is loyal enough.

It asks how your recurring activity gets governed. By your employer’s own authority, by an outside supplier through price and contract, or by some hybrid with shared safeguards and retained decision rights.

And the activity isn’t “customer service” or your whole job. It’s a bundle small enough to have edges. Drafting replies for password-reset and billing questions from approved material, say.

Make, buy and ally are three forms of governance, not a grand switch between insourcing and outsourcing. Your employer can buy a hosted model and application layer while making the retrieval, identity, evaluation and workflow integration that turn it into an actual service.

That mundane unit is the whole point.

Which is why the strategy conversation doesn’t start when somebody proposes a large contract. It starts when your pilot takes a recurring queue, a body of context, and an operating role.

You may still send every reply. The provider may still be paid out of an innovation budget.

But the work has begun to live across a boundary somebody will eventually have to govern.

From demo to service

A real pilot starts from a measurable problem, not “improve customer service with AI.”

It names your queue, who gets the draft, which sources the system may use, what it must never answer, what stays yours alone, and the event that stops it.

Then it records a baseline. Weekly volume, median handling minutes, recontact rate, quality score, escalation rate, cost per completed case.

That isn’t corporate throat-clearing. Without it, “time saved” is a compliment paid to a demo.

The willingness to stop matters more than the willingness to start.

In Gartner’s 2024 enterprise survey, 41% of generative-AI prototypes and 42% of non-generative prototypes reached production on average. The missing ones aren’t all cowardice or bad engineering.

A pilot that finds inadequate source material, an impossible rule, or too little benefit before it reaches your wider queue has done exactly what a pilot is for.

Then the demonstration asks for the one thing demonstrations never need.

Live records.

The neat answer got built on a selected file. Your queue brings superseded guidance, other languages, unusual account types, regulated disclosures, and the quietly decisive question of who may see what.

A production data product has a named owner, defined meaning, access rules, and a way for you to correct a defect. It records source, version, audience, permission. It can remove an old document and revoke an entitlement.

That is not a pile of PDFs with better search.

None of this is hypothetical. Gartner found 48% of high-maturity leaders naming security threats among their top three implementation barriers.

So your pilot has now acquired controlled access, continuous source maintenance, and a dependence on the people who keep those sources current.

Whether or not the renewal still says trial.

The real asset is starting to look less like the model and more like the knowledge, the permissions, and the repair work around it.

The cheap query is not the service

Your external layer costs inference, seats, storage, cloud compute, security tooling, supplier support.

Your internal layer costs process-owner time, data engineering, integration, evaluation, security review, legal and procurement work, training, help desk, monitoring, incident response.

The third layer is change. Redesigning the work.

Running in parallel. Auditing output.

Absorbing errors.

That one costs even when no invoice arrives.

Your pilot doesn’t remove any of these. It introduces the obligation to decide who carries them.

There are three older tests for that obligation, and they arrive in an unhelpful order.

How specific is the asset. How uncertain is the future work. How often does it repeat.

Specificity asks whether your knowledge, tooling or investment loses value outside this relationship. Uncertainty asks whether the future work can be specified and verified. Frequency asks whether repetition justifies a dedicated arrangement.

They aren’t a scorecard that reveals destiny. They’re a way of noticing that a low subscription price hides the cost of searching, qualifying, specifying, monitoring, arguing, and adapting when the work changes.

The other side has costs too. Supervision. Bureaucracy. Internal bargaining. Weak incentives. Honest managerial error. Yours included.

So this was never a moral test of whether to keep work inside.

A common, transferable activity can sensibly be bought. A distinctive, repeated one deserves a hybrid. Buy the general capability, keep your context and your controls, and hold enough understanding to change the arrangement when the supplier or the data or the work moves.

Dependence arrives as ordinary use

The feeling sharpens once the service is released narrowly and you start doing real work with it.

Your measure was never that employees opened the feature.

It’s eligible cases, drafts shown, acceptance without material edit, material edits, overrides, time to send, sampled accuracy, and what happened to the customer afterwards.

Acceptance can mean trust. It can also mean you, at four o’clock, agreeing.

A low rate with useful tags can point at one stale source, or one bad interface.

High-maturity organisations report keeping AI initiatives in production for at least three years far more often than low-maturity ones do. At that duration, source refresh, support, model changes and outage handling aren’t future concerns.

They’re the work.

Calling something a pilot after you’ve built your afternoons around it doesn’t preserve anybody’s freedom. It leaves the continuing obligation unnamed.

And your minutes saved are capacity evidence. They become cash only when spending actually changes. Hiring avoided. Overtime or contractor spend cut. A contracted service removed. Infrastructure closed.

Otherwise those minutes mean better throughput, better service, more resilience, or time for work that deserves you.

Those can all be good reasons to continue. They aren’t the same reason, and your ledger shouldn’t pretend they’re.

The theory can’t flatter itself

Make-or-buy has its own temptation. It likes to make the eventual choice sound inevitable.

Its most important costs — your lost learning, bargaining delay, management attention, unplanned adaptation, informal trust, internal lobbying — are inferred, never directly observed.

Coase’s 1937 insight generated more than six hundred articles without turning any of those into a clean line on an invoice.

The evidence underneath is uneven too. Asset specificity has been studied heavily. Volume uncertainty, technological uncertainty and behavioural uncertainty have far thinner support, and frequency barely enters the analysis at all.

That’s not an argument for ignoring your boundary. It’s an argument against turning your pilot’s continuation into a story where the economics made everyone do it.

Firms select arrangements for reasons that also affect performance.

Power matters. So do politics, executive incentives, legal conditions, and who keeps discretion when the model or the interface or the source changes underneath you.

None of those are transaction costs in better tailoring.

A decision allowed to end

At review, your process owner and finance reconcile the original baseline against what actually happened. Volume, case mix, training time, rework, supplier consumption, source maintenance, and the capacity nobody could redeploy.

That’s where your pilot earns a name more useful than success or failure.

Extend it. Hold it at its present boundary. Or retire it with the lesson intact.

Retiring isn’t an admission that you were foolish to try. It may mean your data wasn’t ready, the quality wasn’t safe enough, the value was too small, or the right boundary sat somewhere the demo never suggested.

A business that can’t say no to a pilot isn’t experimenting. It’s outsourcing its ability to choose.

You’re not late to the strategy discussion. You didn’t miss it while somebody else bought a model.

It began the moment the pilot touched live work. When the files became a data product. When the draft entered a queue. When a person relied on the result. When somebody quietly started maintaining the whole arrangement.

That faint embarrassment is useful. It’s what it feels like to notice the temporary thing has become part of the building.