AI and Staff Productivity: What Four Studies Found, and What It Means for a Nigerian Service Firm

Dell’Acqua and colleagues ran a preregistered experiment with 758 consultants at Boston Consulting Group. Noy and Zhang tested writing tasks across 453 college-educated professionals. Brynjolfsson, Li and Raymond followed more than 5,000 customer support agents through a real deployment rather than a lab.

The gains are real, they are large, and they are far more specific than the sales pitch suggests. The same studies that found the gains also found the damage, which is the half nobody quotes.

Those are big numbers, and I want to be careful with them, because none of them came from a Nigerian service firm and none of them came from Paxlan. What they tell you is the size of the prize on the right task. They do not tell you that your firm will see the same thing on whatever task you happen to pick first.

This is the finding that gets ignored, and it is the one that should change how you think about the whole thing.

The writing study found the same pattern. Quality gains concentrated in the bottom half of the skill distribution, closing the gap between the strongest and the weakest writers rather than widening it.

So AI does not make your best people better. It pulls your weakest people closer to your best ones.

For a Nigerian service firm that matters more than the speed number, because of where your bottleneck actually sits. If you are the founder, you are probably the person everything escalates to the moment it needs judgment. The reason your junior brings you the draft is not that they are slow, it is that they get to a certain point and stop. When the floor rises, fewer things reach you, and the capacity you get back is your own.

Yes, and this is the part nobody selling it will mention.

The BCG researchers included one task deliberately designed to sit outside what the AI could handle. On that task, the consultants using AI were 19 percentage points less likely to reach the correct answer than the consultants who had no AI at all.

Not slower. Not merely unhelpful. Worse than doing it by hand, by a wide margin, and these were skilled consultants at a top firm.

The researchers called the boundary the jagged frontier, and the word carrying the weight is jagged. The line between what AI does brilliantly and what it does badly is not smooth and it is not obvious. Two tasks that feel equally routine to you can sit on opposite sides of it. Your team cannot see the line either, which is why they trusted the confident wrong answer.

This is the entire argument for finding out before you build, and it comes from a Harvard-run study rather than from a firm with automation to sell you.

Two things landed this year.

The first is cost. The Nigeria Tax Act took effect on 1 January, rewriting the PAYE bands, removing the Consolidated Relief Allowance, and folding four separate levies into a single 4% Development Levy. On top of salary, an employer is already carrying 10% pension on basic, housing and transport combined, plus 1% NSITF and 1% ITF once you pass five staff or ₦50 million in turnover. That lands near 12% before you count recruitment, equipment, or the time your managers spend supervising.

CareerBuddy put the fully loaded figure at 19.5% to 43% above face-value wages. That one is their own balance-sheet math rather than published research, so treat it as a direction rather than a number.

The response across corporate Nigeria has been to convert staff onto contractor retainers and strip the liabilities off the books. That solves an accounting problem. It does not solve the underlying one, which is that the same repetitive work still has to be done by a person.

The second is distance.

Inside the United States, lower-usage states are catching up. Between countries, the leaders are pulling further ahead. So the gap is opening, not closing, and waiting to see how it settles is itself a decision.

You do not need a consultant to start this, so here is the test we use.

Probably inside the frontier

• It runs on a trigger or a schedule rather than a hunch.
• Its inputs are already written down somewhere instead of living in somebody’s head.
• A competent new hire could be taught it in a week from a checklist.
• There is a right answer, and someone could verify it in under two minutes.
• It gets done the same way whichever client it is for.

Probably outside it

• It depends on something nobody wrote down, like which client’s MD hates being called on a Friday.
• Being wrong is expensive, and hard to notice until much later.
• It needs somebody to weigh competing priorities against each other.
• It is the thing your best person is actually paid for.

Then count. Take one week, have each person log what they did in half-hour blocks, and sort the blocks into those two piles. Multiply the first pile by what that person costs you loaded, not by their salary.

Most firms doing this honestly for the first time are surprised by the size of the first pile, and more surprised by how much of it belongs to their most expensive people. If you would rather see where you stand before running the exercise, the AI-readiness quiz is ten questions and gives you a scored answer.

Not buy a tool. The firms getting nothing out of AI mostly bought a tool and went looking for somewhere to point it.

Start with the count above, on one team, for one week. Take the largest thing in the first pile, the one that happens most often and has a checkable right answer, and automate only that. Measure it against what it cost before. Then do the next one.

A single working automation your team actually uses is worth more than a six-month programme, because it proves the model inside your own firm on your own numbers, and everything after it gets easier to fund and easier to trust.

Where the repetition sits differs by trade. We have written up what it looks like in recruitment and staffing, in marketing agencies, and in accounting practices.