Managing Two Capability Pools - Human and AI
Managing Two Capability Pools - Human and AI
The quiet arrival of a second workforce
Every organization now runs on two capability pools. One of them shows up to the Monday meeting, has opinions about the parking situation, and remembers exactly what you promised at the last performance review. The other one arrived by procurement, has no opinions it will admit to, and remembers nothing at all, which is either its greatest weakness or its most attractive feature depending on how the quarter went.
Most management thinking has not caught up. We still have headcount plans, hiring plans, learning and development budgets, succession charts. What we do not have is a language for the second pool. It sits in the technology budget, gets discussed in terms of licences and tokens, and is managed by people whose job title suggests infrastructure rather than talent. Meanwhile it is quietly doing work that used to be someone's career.
The result is a peculiar kind of organisational split brain. Half the leadership team is optimising a human portfolio with the tools of twentieth century personnel management. The other half is buying capability the way you buy electricity. Nobody is managing the whole.
Two pools, two physics
The temptation is to treat the second pool as a cheaper version of the first. This is the single most expensive mistake available to you right now, and it is available in bulk.
The human pool behaves like a slow compounding asset. Capability accrues through exposure, failure, mentorship and boredom. It cannot be provisioned. It arrives late, unevenly, and mostly through experiences that look inefficient on a timesheet. It also depreciates if unused, holds context across years, and comes bundled with accountability, which is not a feature you can buy separately.
The AI pool behaves like nothing in the classic management canon. Capability arrives in discrete jumps announced by a vendor blog post, unrelated to anything you did. It is infinitely parallel and instantly forgetful. It is superb at breadth and unreliable at edges. It has no stake in the outcome, no fear of being wrong, and no ability to tell you which of those two things is currently causing your problem.
Same word, capability, describing two entirely different substances. Managing them with one mental model is like running a fleet of sailboats and jet engines under a single maintenance policy because both technically move.
Nobody has an inventory
Here is the awkward part. Before you can allocate across two pools, you need to know what each pool can actually do. Almost no organisation knows this about either one.
Managers know job titles, not capabilities. You know that Priya is a senior analyst. You may not know she taught herself to build data pipelines during lockdown, negotiates beautifully under pressure, and is the only person who can talk to the finance director without an incident report. That is capability. Her title tells you approximately none of it.
The same fog covers the second pool. Ask an executive what the AI systems in their organisation can reliably do and you will receive a list of vendor claims, a pilot that ran in March, and a rumour. Nobody has done the unglamorous work of testing where the tools break, because that work is boring, slightly humiliating, and impossible to put in a board deck as a win.
So we get the modern spectacle - a leadership team allocating work between two capability pools it cannot describe, based on demos it did not verify, with a confidence level normally reserved for religious conviction.
Substitution happens at the task level, hierarchy happens at the job level
Almost every unproductive argument about AI and work comes from a category error. Roles are bundles. A job is thirty tasks stapled together by an org chart and a salary band. The technology does not replace jobs, it dissolves individual tasks unevenly across the bundle, and then everyone argues about whether the job survived.
This means the unit of capability management is no longer the role. It is the task, and slightly more usefully, the workflow. Which is genuinely irritating, because your entire management apparatus - budgets, spans of control, career ladders, recruitment pipelines - is built around roles.
The practical work is unfashionably granular. Take a workflow. Break it into steps. For each step ask three questions.
- What is the cost of being wrong here, and who absorbs it
- Does this step need context that lives in someone's head or in a relationship
- Is the output verifiable cheaply, or does checking it cost as much as doing it
That third question is the one everyone skips, and it is the one that determines whether you gained anything. If a system produces work in ten seconds that takes a competent human forty minutes to verify, you have not automated the task. You have converted a producer into an inspector and called it transformation. Inspectors, incidentally, tend to be less happy than producers, which is a morale cost you will pay later without ever attributing it correctly.
The apprenticeship problem, or how to eat your own seed corn
The tasks the second pool handles most convincingly are, by a cruel coincidence, precisely the tasks that used to train the first pool. First draft memos. Basic analysis. The tedious research. Routine code. The work of the first three years.
Nobody became a good senior lawyer without producing several hundred mediocre documents. Nobody developed engineering judgement without shipping things that broke. Expertise is not downloaded, it is sedimented, and the sediment comes from doing low value work at high volume while someone more experienced sighs at you.
Remove that layer and the arithmetic works beautifully for four years. Then you discover that senior capability has a fifteen year lead time and no substitute good, and that you have optimised the bottom of a pipeline whose top you still depend on entirely. This is not a technology risk. It is a capital allocation decision, made accidentally, by people who thought they were approving a software purchase.
The organisations that get this right will do something that looks wasteful in the short term - deliberately routing work to humans not because they are better at it, but because doing it is how they become better at everything else. Call it a training subsidy. Put it in the budget with a straight face. Defend it in the meeting where someone points out that the machine does it for a fraction of the cost. They will be correct and you will still be right.
Volume stops being a signal
When the marginal cost of producing a plausible artifact falls to nearly nothing, output volume dies as a measure of contribution. This has happened, and most performance systems have not been told.
Expect a rising tide of confidently formatted, structurally impeccable, faintly hollow work. Documents that answer a question nobody asked. Fifty slide decks where four slides would do. Analysis with the shape of insight and none of the load bearing weight. Each one takes ten seconds to generate and twenty five minutes for a colleague to read, absorb and quietly discard. The productivity gain of the sender is funded by a productivity tax on everyone downstream, and it does not appear in any dashboard.
Managing two capability pools therefore includes an unglamorous new duty - being the person who asks what a piece of work is for. Not sarcastically, although the temptation will be strong. The scarce resource is no longer production. It is judgement, attention, and the willingness to say that something plausible is not the same as something useful.
The political economy of the second pool
A capability pool with a large budget and no representation in the room creates predictable dysfunction.
Its successes are claimed by whichever function is nearest at the moment of measurement. Its failures are structurally orphaned, since blaming a tool feels unsophisticated and blaming the person who trusted the tool feels unkind. Governance therefore accumulates in the vacuum, usually in the form of a committee whose main output is a policy document that nobody in delivery has read.
Meanwhile the human pool is watching. Closely. Not the strategy deck, which they correctly ignore, but the small things - whether the announcement of a new capability comes with reassurance or with silence, whether the word efficiency is used as a synonym for something less pleasant, whether the people who raise concerns get thanked or managed. Your capability strategy is also a loyalty strategy, and the second document is written entirely in behaviour.
The organisations that lose their best humans to this transition will not lose them to redundancy. They will lose them to the slow realisation that leadership found the machines more interesting.
A working stance
Not a framework. Frameworks are how we avoid thinking in a well organised manner. Call it a stance.
Treat capability as a portfolio, not a headcount. Hold an honest inventory of both pools, including where each one reliably fails, and update it more often than annually because one of them changes quarterly. Allocate at the task level while managing careers at the role level, and accept that this is genuinely difficult rather than pretending the tension does not exist. Budget for verification as a real cost. Subsidise human learning even where it is locally inefficient. Measure outcomes rather than artifacts, because artifacts are now free. Concentrate human attention where consequences are irreversible, relationships are load bearing, or someone must be genuinely accountable, which remains a thing only humans can be.
And retain a small amount of scepticism about anyone, vendor or internal, who describes either pool as unlimited. One of them needs sleep. The other one needs supervision. Neither of them needs your enthusiasm as a substitute for your attention.
The management task itself has not really changed. It was never about managing people. It was always about allocating scarce capability against unbounded need, under uncertainty, with insufficient information and a deadline. There are simply two currencies now, and no published exchange rate.
Which is, admittedly, the most interesting the job has been in about a century.