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.

Share this line as a card

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.

Whether capability travels has been tested on surgeons and on aircraft#

The four-item list above assumes something that can be checked: that capability is a stock an organisation either holds or has let go. Two measurements say it is stranger than that, and both make the vendor question harder rather than easier.

Robert Huckman and Gary Pisano looked at cardiac surgeons who operate at more than one hospital within narrow periods, and used patient mortality as the outcome. A surgeon's results at a given hospital improved with recent procedure volume at that hospital, and did not improve with volume at other hospitals. Their conclusion is that performance "is not fully portable across hospitals", and their tentative explanation is familiarity with the assets of the particular organisation. Graded entry.

The buyer's version of that finding is unwelcome. Part of what a supplier delivers belongs to the pairing of that supplier with your organisation, and accrues through repetition inside it. Switching suppliers therefore costs more than the contract says, and so does bringing the work back in-house, because the returning team has the general skill and not the specific accumulation.

Benkard's aircraft data gives the same shape from the manufacturing side: experience depreciates over time, and only some of a firm's production experience transfers from one generation of an aircraft to the next. Graded entry. A capability held through one version of a model is not automatically a capability held through the next one.

Which makes the fourth item on the list, enough to replace the supplier, a rate rather than a reserve. It is not a document that can be filed and left. On the only evidence anyone has measured, it survives in proportion to how often it is exercised inside this organisation on this work, and the rates at which unexercised capability disappears are not slow.

Strategy advocacy, and a literature that disagrees with itself#

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.

The two measurements in the section above were made on cardiac surgery and on aircraft assembly, neither of which is a vendor relationship for an AI capability. Huckman and Pisano describe their own asset-familiarity explanation as preliminary, and this page takes their result from the published abstract rather than from the tables, so it carries the direction of the finding and none of its magnitudes.

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.
• Huckman, R. S. and Pisano, G. P. (2006). The Firm Specificity of Individual Performance: Evidence from Cardiac Surgery. Management Science, 52(4), 473-488. Graded entry.
• Benkard, C. L. (2000). Learning and Forgetting: The Dynamics of Aircraft Production. American Economic Review, 90(4), 1034-1054. Graded entry.

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.