AI, Outsourcing and the Boundary of the Firm

In 1937 the economist Ronald Coase asked a simple question: if markets work so well, why do firms exist? His answer was that using the market has costs of its own—finding suppliers, negotiating, enforcing contracts—so firms do in-house whatever is cheaper to organise than to buy. That logic sets the boundary of the firm, and for twenty years it sent banks' and consultancies' analytical work offshore. AI is about to change the calculation.

AI won't just automate knowledge-process outsourcing (KPO). It changes what that work leaves behind.

Once a workflow runs on AI, it can produce reusable know-how: prompts, evaluations, exceptions and decision logic. That weakens labour arbitrage and changes the make-or-buy decision. Commodity execution can stay outsourced; proprietary workflows and institutional memory may move into firm-owned capability centres.

Outsourcing is getting squeezed from both sides

For twenty years, KPO moved repeatable financial work—research, credit analysis, drafting—from expensive labour markets to cheaper ones. Vendors sold analyst-hours; the wage gap was their moat.

AI cuts the human time needed for routine production. Stanford's 2026 AI Index reports that AI agents' success on OSWorld, a benchmark of everyday computer tasks, rose from about 12% to 66.3% in roughly a year. That is not a test of analyst work: real tasks still need reliable sources, judgement and sign-off. But even partial automation reduces the hours a client must buy.

The harder question is who owns the reusable assets the work produces.

1. AI replaces analyst-hours with software and expert review

The comparison is no longer a London analyst against a Bangalore analyst. It is an offshore analyst against an AI workflow supervised by experts.

Take a sector-research pack. Retrieval, extraction, comparisons and first drafts can be automated. The analyst's job shifts to framing the question, resolving ambiguity, checking sources and signing off.

What clients pay for is shifting from hours to verified output

Professional-services firms are built like pyramids, with a wide base doing routine production. AI can flatten that base without eliminating the profession.

That changes what gets measured. A provider selling an analyst team is rewarded for utilisation. One selling a verified credit memo must manage throughput, error rates, source traceability and expert review. Bespoke judgement may still be billed by the hour. In repeatable work, automation breaks the link between hours supplied and revenue earned.

McKinsey describes a retail bank where AI agents draft credit-risk memos from current data, prior memos and relationship-manager feedback, citing sources for their assumptions. McKinsey estimates a 20–60% productivity gain and 30% faster credit decisions. That is one workflow, not proof that a whole analyst role has vanished. But if buyers can do the routine preparation in their own systems, vendors have fewer hours to bill. The human reviewer remains; the production pyramid changes shape.

2. Execution becomes IP creation

A firm used to separate process execution from creating intellectual property (IP): outsourced work simply returned a deliverable. With AI in the loop, each run can also produce training examples, evaluations, feedback, exceptions and process logic.

The most valuable rules often live in exceptions, not policy manuals: why a sound-looking credit was rejected, how EBITDA is adjusted in an odd case, when a customer gets escalated. Encoding those judgements makes them portable.

That asset can compound. A reviewer’s correction becomes an evaluation; repeated edge cases become rules; traced decisions make the next run easier to check. The value is not a generic prompt but a record of judgement in real cases. It accrues only to whoever can lawfully keep and reuse those artefacts.

Whoever builds and keeps the workflow may learn how the business makes decisions. Make-or-buy is therefore partly a decision about who owns IP, not just a procurement choice. That holds in consulting, law, accounting, engineering, healthcare administration and banking.

The client need not own the vendor’s foundation model or general-purpose software. The real line runs between the provider’s reusable tools and customer-specific learning: labelled decisions, exception histories and evaluations shaped by the client’s work. A good contract can leave the platform with the vendor while keeping that learning portable and under the client’s control.

3. Longer data chains make leaks easier to cause and harder to spot

A vendor may add a model provider, orchestration layer, vector store and subcontractors, and any of them may log prompts, outputs and feedback. Agents also need access to client systems. Each link is another place to lose control.

The old path was Bank → KPO → Bank. The new one may pass through orchestration, model and infrastructure providers, each with different retention, access and contractual terms.

Contracts that cover only “customer data” may miss derived assets: prompts, embeddings, evaluations and workflow logic. The question is no longer only “Who owns my data?” but “Who may retain, reuse or train on what was learned from it?”

That turns procurement into a design problem. Buyers need to know who may reuse inputs, outputs and derived artefacts; whether reviewer feedback can improve tools sold to other clients; and what can be exported or deleted on exit. Segregation, auditability and the terms passed down to model providers decide whether an outsourced service quietly becomes someone else’s asset.

An investment-bank example

Consider an unannounced acquisition. Under deadline pressure, a KPO analyst pastes part of a confidential information memorandum into a personal AI account to draft comparable-company commentary.

No malice is required. The bank may have no agreement with that provider, no visibility into retention and no record of the disclosure. Weak separation between client workspaces could open another route out, to a rival client.

What can leak is not only the file but the reasoning around it: prompts, chosen comparables and valuation assumptions.

Not every model trains on every prompt; that depends on the service and its settings. The immediate failure is lost visibility and an unapproved path for confidential material. Even if no breach follows, the bank may be unable to show a client, regulator or deal team where the information went or whether it was deleted. That uncertainty has its own cost.

4. Outsource the commodity work; keep the intelligence

None of this kills outsourcing. It splits it.

Routine, standardised work can stay outside the firm, and may become more attractive to outsource as providers automate it. Much of it will look less like labour outsourcing and more like services-as-software: the client pays for an outcome, and the provider decides the mix of people, software and AI infrastructure needed to deliver it.

Work that generates proprietary data, encodes how you make decisions or captures what your firm knows is different. There, the case for keeping the intelligence in-house is much stronger. The team can still sit in India, Poland, the Philippines or anywhere cheaper. I expect more firms to run it through a captive Global Capability Centre (GCC) than an outside KPO.

A GCC keeps the cost advantage of location but holds the assets and governance inside the firm. Where the work happens matters less; who owns what it produces matters more.

This is a decision rule, not a call to bring everything home. Ask three questions. Does the judgement set us apart? Can derived assets be segregated and transferred? Can the provider’s output be audited? A vendor may still win on scale or specialist tools, and a captive centre brings its own fixed costs, hiring needs and operating risk. The case depends on what must be owned, not on who employs the most analysts.

5. Software competes for the workflow

Tools such as Rogo are built to help bankers and investors produce models, memos and presentations from a firm's own systems and data. They target the junior-analyst tasks on which KPO pricing has relied.

The plumbing matters too. Connectors such as the Model Context Protocol let AI tools reach market data and internal systems. That makes fetching and reformatting easier to automate, though checking results and controlling access still take work.

The advantage is not just a better model. It is a product that knows who may see what, cites its sources, writes back to approved systems and leaves a usable audit trail. A standalone tool without those links may still need a services team around it. But once a firm has trusted connections and reusable workflows, it can replace routine external work far more easily. The bottleneck shifts from cheap labour to access, integration and control.

Legal offers a preview. Harvey targets document review, due diligence, contract extraction and drafting—the same tasks legal-process outsourcers sell. Adoption alone does not prove that outsourcing is shrinking, but it gives law firms a way to do that work in-house.

Two responses follow. The client licenses the tool and brings production back, or the outsourcer licenses it and sells integration, quality assurance and accountable delivery. The second still pays where data is messy, exceptions are frequent or a human reviewer carries liability. For outsourcers, the defensible move is to build these tools into verifiable outcomes, not compete with them on the price of an hour. Software does not erase demand for expertise; it weakens the claim that more junior hours equal expertise.

6. Who captures the productivity dividend?

KPO revenue could grow even as employment and margins move differently: clients may buy more analysis but need fewer hours per unit. So the investor question is not whether revenue grows, but who captures the productivity dividend.

Take an illustrative contract worth 100 in vendor revenue: 55 in production labour, 20 in other costs and 25 in EBIT. This is a model, not a forecast.

Now AI halves production labour to 27.5 and adds 10 in platform costs; other costs stay at 20. Vendor costs fall from 75 to 57.5. With volume and quality held constant, three outcomes are possible.

Illustrative contract economics (units per 100 baseline revenue)

In A, the client negotiates the fee down to 70: vendor EBIT falls to 12.5 despite automation. In B, the fee stays at 100 and EBIT reaches 42.5—until competition or renewal forces it to share the savings.

In C, the client buys software directly and the outsourcing contract disappears. The zeros show that contract's revenue and EBIT, not the client's total spending or the software firm's profit. Switching costs, supervision, demand growth and new AI expenses would change all three outcomes.

The arithmetic shows where the negotiation turns. Net savings are 17.5, not the full 27.5 labour cut, because the platform costs 10. A fee of 82.5 leaves vendor EBIT at its original 25; a fee of 70 gives away more than the gain and eats into the old margin. If volume grows, total profit could still rise while the margin per contract falls.

The split is stark: hour sellers face pricing pressure; fixed-fee outcome providers may keep the savings for a while; platforms collect recurring fees; clients and GCCs can keep both the savings and the knowledge the work creates. Lasting margins follow control of the workflow, not just the volume of work.

Investors should watch pricing and ownership, not AI announcements: how much work is fixed-fee rather than time-and-materials; whether cost per verified output is falling; how much of the savings renewals pass to clients; and who keeps the evaluations and exception histories. Rising revenue with flat headcount can signal operating leverage, but only if gross profit and contract retention back it up. Otherwise someone else is collecting the productivity dividend.

7. Agent permissions are the control test

Once agents do real work, access control must cover what they can do, not just what they can read: which systems they can reach, what they may change, who approves significant actions and how each step is logged.

For a provider, that means client segregation, minimum necessary permissions, audit trails, human sign-off and a way to shut an agent down. Prompt injection (hidden instructions planted in content an agent reads) and cross-client leakage make these controls part of the product, not a compliance afterthought. A provider that can prove them has a real basis for trust.

These controls also drive unit economics. Broad, standing agent access is easier to run but makes failures costlier; task-specific credentials, separate client memory and tested rollback add expense. Human review cannot simply be cut from the P&L: high-stakes decisions may need sign-off even when drafting is cheap. Providers that measure error rates and set oversight to match have a better claim to both safety and lasting margins.

The new make-or-buy question

The first outsourcing wave asked: where can this work be done most cheaply? The AI wave asks: where should the knowledge created by doing the work live?

Coase's test still holds: firms do in-house what is cheaper to organise than to buy. AI changes the price on both sides. Buying analyst-hours gets cheaper, but buying now also means handing over the judgement, exceptions and decision logic the work leaves behind. For commodity work, the market still wins. For workflows that encode judgement, institutional memory and competitive advantage, the firm increasingly does.

Outsourcing isn't disappearing. Coase's boundary of the firm is moving.