The KPO promise was never just cheaper thinking, and neither is enterprise AI

The answer is immaculate. It’s also unusable, in a way you can’t immediately prove.

The deck has the headings.

The brief has the citations.

The model wrote the first page before lunch.

But the small private map in your head is gone. Which source was weak.

Which term changed meaning halfway through the search.

Which person knew the obvious comparison was false.

You’re not precious for caring about that map.

You’re tired. You’ve got a decision waiting, messages arriving, a family that still needs dinner, and no reason to spend another evening rebuilding a chain of reasoning that something has offered to compress.

Accepting the help isn’t a betrayal of your profession. It’s what people do when the work is more than one person can hold.

The machine’s intentions are the least interesting thing in the room.

Professional work only becomes movable after somebody decides what can be standardized, inspected and retained.

KPO learned that long before enterprise AI started promising to make every analyst more productive.

2005 — Research comes back from an offshore team in a tidy deck.

2014 — The source notes are in a shared folder.

2023 — The model produces a first brief before lunch.

2026 — The conclusion circulates without anyone remembering the question.

What had to move first

KPO was never cheaper thinking in another time zone.

A client asking whether to fund a distributor search in one country in one quarter can’t responsibly send you the instruction “find all competitors.”

The question has to become a search universe, an inclusion rule, known blind spots, permitted sources, excluded activity, a confidence requirement, and an answer the client can accept.

The result is a one-page decision charter and a question log.

That isn’t paperwork around the research. It’s the beginning of the research.

The work travels digitally. The network doesn’t erase your boundary. It makes the boundary easier to overlook.

A regulated claim stays a regulated claim. A professional decision stays with the client. A UK personal-data transfer still needs its role and path understood, because an overseas sub-processor can matter even when the controller sends the file directly.

That’s the part your rate card couldn’t buy.

Your offshore team needed enough context to know what the product meant, where the sales channel failed, and why a familiar claim was misleading.

It didn’t need every customer file anybody could find.

Public sources plus a sanitized brief might do. When controlled access was necessary, you named the workspace, the user group, the purpose and the retention rule.

Too little context makes the analysis generic. Too much makes it exposed, slow and expensive.

The professional skill was never access to everything. It was knowing what had to cross the boundary.

Then the answer had to keep its path.

An evidence table records source, date, claim, relevance, confidence, limitation. A web lead could discover a company without proving its ownership, revenue, regulatory status or technical claim.

At the planned midpoint, you and your provider have three possible instructions. Continue as scoped. Redirect. Or stop, because the evidence can’t carry the conclusion.

“Keep looking” was not a fourth instruction. It was usually a way to spend the remaining budget without learning anything.

Then came the work people call overhead because it doesn’t look like the finished answer.

A new service needs two case sets. Calibration cases, where you and the provider discuss why the answer was right. Blind cases, where the work gets judged before that discussion.

Calibration transferred judgment. Blind work tested the transfer.

Peer review checks the brief, the citation trail, the arithmetic, the completeness, the uncertainty. Your domain owner checks whether any of it answered the decision.

One is evidence discipline. The other is relevance. Neither stands in for the other.

Acceptance had a grammar too. Accepted. Accepted with revision. Returned for correction. Withdrawn because the brief changed.

Those states sound drab until the source notes vanish and you can’t tell whether a conclusion was wrong, incomplete, inconvenient, or simply aimed at a question nobody wrote down.

On closure, the final work, source register, working papers and uncertainty statement had to survive the presentation.

An exit rehearsal tests three actions. Retrieve the work. Remove access. Assign the open questions.

Your provider can leave. Your ability to ask and inspect the question can’t.

The invoice was never the bargain

The KPO ledger used two cost pools.

One was the provider’s billed cost. The other was your retained cost of framing, orientation, access, review, feedback and decision.

Compare total cost per accepted output. Not a wage in one country against a wage in another.

If you rewrite every deliverable, you haven’t bought capacity cheaply. You’ve bought an extra draft.

There were four common billable units. Hours. Named full-time equivalents. Projects. Accepted outputs.

Each pushes the uncertainty somewhere else.

Time and materials lets your question change, and rewards activity unless you watch. A fixed price suits a defined scope. A dedicated team keeps vocabulary and context, and creates idle-capacity risk. Output pricing makes the deliverable legible, and makes difficult research quietly leave your scope.

Your contract model never settled whether the insight was useful. It arranged who discovers the failure first.

Rework has three contributors. A provider correcting a defect. You changing a requirement. Genuinely new information.

Treat all three as provider fault and you get a tidy dashboard and a stupid service. Treat all three as your indecision and you get the same thing.

The useful metric tree has three levels. Production, capacity, decision.

On-time delivery and evidence completeness sit at the first. Qualified work your internal team can now handle sits at the second. A changed choice, reduced uncertainty in time, or an avoided loss sits at the third.

A quick deck can be excellent at one level and worthless at the next.

None of which made the market small. The World Bank’s cited analysis put global ICT-enabled services demand at US$755 billion, with emerging-market growth at 14% against 4.9% globally.

That doesn’t prove your research brief pays. It shows how a large market makes a boundary look inevitable when it’s still a decision.

And your retained capability belongs in that decision, because the exit has three workstreams. Recover files and access. Transfer open work and service knowledge. Re-establish the operating capacity.

That cost isn’t a sentimental premium for keeping experts nearby.

It’s what lets you inspect a conclusion, remember why you trusted a source, and move again without reconstructing your method from ash.

The same threshold, smoother interface

Now put the enterprise AI pitch beside that history.

The model drafts your reply, summarizes your case, produces a market brief. The interface makes knowledge work feel cheap because the first response is nearly immediate.

The query got cheaper. The service did not become free.

A real service starts with a bounded queue, a named production owner, an operating budget, a baseline, the sources it may use, the work it must never do, and an event that ends your pilot.

“Make analysts more productive” isn’t a service boundary. It’s an ambition looking for a bill.

Your knowledge base has to become an owned data product. Current material. A source of truth. Named owner, version, audience, permission, provenance, freshness. And a way to repair a defect.

A held-out evaluation set stays separate from the examples used to configure the service.

Buying a hosted model is fast. Building the retrieval layer, the identity, the evaluation set, the workflow integration and the controls is what makes it your service rather than somebody else’s demonstration.

The smooth answer is only the visible surface.

Under it sit the question, the evidence, the test, the exception, the approval and the maintained source.

And acceptance still tells a truth that output can’t.

Opening a feature doesn’t mean you used a draft in a real case. You need eligible cases, drafts shown, accepted-without-material-edit rate, material edits, overrides, time-to-send, sampled accuracy, escalation, and what happened to the customer.

Your financial claim needs the same restraint. Lower handling time becomes cash only when spending changes. Otherwise it’s capacity, speed, quality or resilience.

The right comparison was never a model call against an analyst salary.

It’s the whole service ledger. External consumption, data work, integration, evaluation, security, legal review, training, support, monitoring, incident response, redesigned work, parallel running, audits, errors.

The old KPO client paid for framing and acceptance. You pay for them too, whether or not your business case wrote them down.

Not a defence of drudgery

You’re not defending busywork.

You’re defending the chain of reasoning that makes knowledge portable without making it hollow.

That chain can be shared. Made faster. Handed to a provider, stored in a service, retrieved by a model.

But its point was never to preserve the ritual of an analyst making a deck slowly.

It’s to let you, later, see the question, inspect the evidence, understand the uncertainty, reject the bad route, and decide whether the conclusion still belongs in the world it entered.

No theory makes that boundary choice neutral.

Coordination costs get inferred, never observed. An organization confuses its choice of provider with the effect of using one. Power and politics don’t become harmless by being called transaction costs.

The tidy explanation is always available after the decision. Your retained record is what lets somebody ask whether it was true.

KPO’s promise was never that thinking ships like a container.

It was that a bounded question could travel with enough evidence, review, calibration, acceptance and retained judgment to come back useful.

Enterprise AI makes the same promise, delivered at the speed of a tab.

That’s why the polished answer can feel lonely. It arrived alone.

What you miss isn’t the clerical weight of producing it.

What you miss is the path by which you knew it was yours to trust.