How do you preserve capability across vendors?

Outsource execution where it makes sense, and retain enough internal knowledge to specify the work, evaluate its quality, manage the exceptions and replace the supplier. That sentence is thirty-five years old. It was written about manufacturing, and it transfers to buying an AI capability without needing much alteration.

Which is the useful thing to say about this question. It is not new, and treating it as new is why organisations are repeating a set of mistakes that were documented in detail before most of the people making them were working.

The original warning

Prahalad and Hamel defined core competence as the collective learning in the organisation, particularly the capacity to coordinate diverse skills and integrate streams of technology. The Core Competence of the Corporation (1990).

Their warning was about what outsourcing does underneath the numbers. Cost advantage can be bought while the competence that made the firm able to compete is hollowed out, and the hollowing is invisible in the accounts because it shows up as savings. Their examples were component supply and manufacturing partnerships in which Western firms progressively surrendered the ability to make things while retaining the ability to brand them.

The transfer to AI is close enough to be uncomfortable. A capability bought as a service arrives with the same accounting signature: cost down, output maintained, internal expertise quietly unexercised.

What has to stay

Four things, and the list is the same as for any outsourced function.

Enough to specify the work. An organisation that cannot state precisely what it needs will get what the supplier finds convenient to provide.

Enough to judge the output. This decays first and matters most. It is also the loss nobody notices, because approving work is indistinguishable from evaluating it right up until the moment it is not. Who supervises work they cannot do.

Enough to handle the exceptions. The cases outside the supplier's competence come back, and they come back as the hard ones.

Enough to replace the supplier. The capability to switch is what makes every other term in the relationship negotiable. It is also expensive to hold and worthless until needed, so it goes first.

The specifically new part

Three things do differ from the manufacturing case, and they are worth separating from the parts that do not.

The thing being bought changes without notice. A vendor updating the model underneath a workflow changes its behaviour without the buying organisation deciding anything. If the process was tuned around the previous model's failure modes, that tuning silently expires. Worse, the tuning is usually informal and undocumented, because it accumulated as habit rather than as design.

Perrow supplies the frame: this is tight coupling to a component you neither control nor inspect. Normal Accidents (1984). The tighter the coupling, the less warning a change gives before it propagates.

The failures are less visible than a supplier's. A component arriving out of specification is detectable. A plausible wrong answer arrives in the same register as a correct one, which is what makes it undetectable at the point of receipt.

The substitution happens task by task rather than by contract. Nobody signs anything. Work migrates, and the point at which the organisation stopped being able to do it itself has no date.

Agents running core processes

The question has a sharper form when the vendor's system is executing rather than advising, and the answer does not change much. What has to stay in-house is the ability to tell whether the output is right, to decide the cases outside the system's competence, to intervene with real authority rather than a nominal veto, and to run the process another way if the supplier becomes unavailable.

Weick and Sutcliffe's principle of deference to expertise rather than to rank matters here specifically. Managing the Unexpected. Retained capability sitting in someone without the standing to use it is not retained capability for this purpose.

What this page does not establish

Prahalad and Hamel argued from cases rather than measurement, and their article is a work of strategy advocacy that became influential partly because it was well written. The outsourcing literature that followed is mixed on outcomes, and firms that outsourced heavily did not uniformly hollow out.

No evidence here establishes that organisations buying AI capabilities lose internal capability, or at what rate. The mechanism is available by analogy and the measurement has not been done.

There is also a real argument on the other side. Retaining the ability to do everything internally is how organisations become slow and expensive, and specialisation exists because it works. The question is which capabilities are load-bearing rather than whether to outsource at all, and this page offers a way of asking rather than a list.

Key sources

• Prahalad, C. K. and Hamel, G. (1990). The Core Competence of the Corporation. Harvard Business Review, 68(3), 79-91. In the essential works.
• Perrow, C. (1984). Normal Accidents: Living with High-Risk Technologies. Basic Books. In the essential works.
• Teece, D. J., Pisano, G. and Shuen, A. (1997). Dynamic Capabilities and Strategic Management. Strategic Management Journal, 18(7), 509-533. In the essential works.
• Weick, K. E. and Sutcliffe, K. M. (2001). Managing the Unexpected. Jossey-Bass. In the essential works.

Related SuperSkills research

On what is being preserved, organisational capability. On testing whether it is still there, the capability audit. On withdrawing a deployment, deployment is not a ratchet. On judging work you can no longer do, who supervises work they cannot do and who owns verification. On agents acting independently, AI agents and human judgement.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Core competence is Prahalad and Hamel's and the coupling argument is Perrow's. This page is an application of both rather than a new claim.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.