Train, borrow or hire, and then step back

The story

The expert everyone waits for.

By the time Maya Okafor had been in the role for half a year, there was one number she looked at every Monday. It was the wait for Tomás Reyes. Tomás was Hartwell's senior ML engineer, and on paper a lucky find. In practice, the wait to get an hour of his judgment averaged about six weeks.

Every team had a reason to need him. Warranty Claims needed him to look at why the model missed an exclusion. Dispatch needed him for a new service-log format. Customer Support wanted a second opinion on a vendor's claims. Tomás said yes to all of them, did all of them in a hurry, and was visibly tired.

Helen Brandt, the COO, did what a reasonable executive does with a queue. She asked how many more people Maya wanted. Maya said she did not know yet. She wanted to understand what the queue was made of.

She sat with Tomás for a morning and sorted a month of requests into three piles. About a third were questions anyone competent could answer if they had the record of the last time the question came up. About half were real judgment calls: when is the familiar answer wrong? And a small remainder were real research, work no one at Hartwell had done before.

That sorting changed the problem. The first pile needed a library. The second needed people who could do what Tomás did, in the cases they would actually meet. Only the third needed Tomás himself. The queue was telling her something about the shape of the work. It was not a hiring plan.

The problem

Four ways to get a skill. None of them is free.

When a company is short of a scarce skill, there are four routes. You can train someone you have. You can borrow a person from another team for a while. You can hire, and wait. Or you can contract, and pay for time you can use now. Each has a cost that people leave out.

Training has two seats to pay for: the learner leaves a job to practice, and the mentor leaves another to teach. It can also fail, finish late or build a skill the company stops needing. Borrowing works only if the lender gives real time, and a lender who loses a good person has little reason to be generous. Hiring costs months before the first useful day. Contracting buys speed and leaves no knowledge behind.

The question is rarely “which is best?”. It is “which is best for this work, given how soon it lands, how long it lasts, and how much of it there is?”

What the research shows

Each answer wins somewhere, and the margins are thin.

Training versus hiring versus contracting. The lesson on train, hire or contract puts all of them in one small world and lets them compete. In the favorable case, where training is certain to work, the best training plan earns 31 illustrative units. Starting recruitment now, or bringing in contractors on the later dates, earns 30 each. Training wins by one. Make training a coin flip, with a one-half chance that the trainee qualifies, and training falls to 24 while the outside routes stay at 30. In another case, all routes tie at 30.

What I take from that is the thinness of the margin. Training does not win by a mile; it wins when the odds are good and the skill is still needed when the trainee arrives. It loses when it is a gamble. The decision depends on three things: the odds that the training works, what the learner and the mentor would have delivered otherwise, and whether the work still exists when the training finishes.

Generalists, specialists and the queue. The lesson on staffing finds that each answer wins somewhere. A broad person who understands both the models and the business can remove a handoff. A specialist is needed when the task requires depth no generalist has. A mix wins when both appear. The same total hours fit when deadlines are spread out and fail when everything is due on Friday. Two urgent jobs that need the same person on the same day are a counterexample to any argument that counts only hours.

The practical lesson: before asking for headcount, name the missing depth and the work that arrives at the same time. A queue is a signal. It does not say what to hire.

Support that lets go. The lesson on support separates three things that organizations run together. Learning is what happens when a person works a task with help. Passing is an assessment of fresh work by someone other than the teacher. Permission to act alone is a decision. Success with an expert at the elbow does not show that the person can make the same judgment alone.

The most useful result in that lesson is uncomfortable. Suppose the support function gains a little from staying involved. In the model, it keeps advising a team that is already ready. The team sees this coming, and skips the last step of learning, because why bother? The company ends up slightly worse off and the team never reaches full readiness. A promise to “hand over when ready” did not repair this. A rule that forced the handover did, and the adviser gave something up to follow it. So the exit has to be written into the mandate.

Honest limits: all of this is small models with invented numbers and assumed-accurate assessment. Real assessment is noisy. The research shows how each choice works and the shape of the trade-offs, not a prediction for your company.

What Maya decided

Train two, bridge with a contractor, do not hire yet.

Maya's reasoning was in three parts.

The middle pile, the judgment calls, was recurring. It would be there next year. That pointed to developing people. Training only wins when the odds are good, so she lowered the stakes in two ways. She picked two learners, not one: Ilse Novak, a Warranty analyst who already read exclusions for a living, and Kofi Mensah, a Platform engineer who knew how the base ran. She put the chance that each would qualify at about one in two. If the two are independent, the chance that at least one qualifies is three in four.

She also booked a contractor for the first ten weeks. The research says outside routes win in the short run, and she took it at its word. The contractor handled the first pile and some of the second while the learners trained. The contract had one clause that mattered: write down everything you do. The company was buying time, and it was determined to keep the notes.

And she did not hire. Recruiting an ML engineer would take about five months. By the time one arrived, the learners would have qualified or not, and the queue might be gone. Five months in, the wait for Tomás had fallen from about six weeks to about two. A hire would have walked into a gap.

The learners' time was borrowed, which made the lending the hard part. Priya Natarajan owned Ilse's time. She did not want to lend her best analyst on the strength of a promise that Ilse would “help out a bit”. So Maya wrote the loan down. It named the hours (two days a week), the end date (twelve weeks) and the receiving owner for what Ilse learned. It also named the lending manager's gain: Ilse would return knowing how to catch exclusion failures before they reached a claim. Priya signed. “A network without reserved time just becomes another queue,” Maya said. I think she is right.

Kofi qualified five months after training began. Ilse did not, because a recall wave pulled her back to claims work in the fourth month. Her training was not wasted, though. She became the best evaluator in the program, the person who could look at a model's output and say whether it was plausible. One of two learners qualified, and one did not. That was inside the odds Maya had priced, and it is why I would never have staked the plan on a single trainee.

The framework

The Handover Ladder.

Everything above is a way of getting a skill to a team. The harder half is letting go of it. Teams that learn a task should need less help with it, and the support function should be able to step back from that task without a fight. I use a four-rung ladder.

Learn. The learner works a real case with an expert beside them. This is where most training programs stop.

Pass. Someone who did not teach the learner hands over a fresh case. The learner explains the risk and chooses the test. The teacher cannot be the examiner, because the teacher has a stake in the result.

Act alone. Permission is its own decision, made by a named owner for a named task family. Not “the person is good now”, but “this person may decide this kind of case in production”.

Support recedes. Help steps back from this task and goes somewhere harder. The record of what was learned stays in the library, and a light spot-check stays on the calendar. If the task changes, the ladder starts again, usually near the bottom.

The ladder is my proposal. The three-way separation between learning, assessment and permission comes from the support lesson. The four rung names are mine.

The ladder in use

How Dispatch graduated.

For the first three months after the shared base went live, Tomás ran a weekly clinic. Anyone from Dispatch could bring a case where the model's routing suggestion looked wrong. He would work it with them, explain the pattern and add a line to the library. This was the learn rung, and it was good.

In the third month, Maya changed the format. Dan Whitfield's dispatchers picked a fresh batch of cases, ones Tomás had not seen, and wrote the check for the model-suffix problem themselves. A reviewer from Warranty, not Tomás, graded it. That was the pass rung. The test they wrote caught two failure patterns Tomás had not noticed.

In the fourth month, Dan's senior dispatcher signed the permission: Dispatch could change the routing rules for this task family without asking. The clinic ended that week. Tomás did not argue. He had a full queue and was glad to lose a row.

Support did come back, once. In the eighth month, a new service-log format arrived from a plant in Brazil, and the team's tests started failing. Dispatch asked for one week of Tomás's time. That was the loop on the ladder working as intended. Help left, the record stayed and when the task changed the team knew how to ask for it back.

The harder part was Tomás. Handover threatened the thing he was valued for. Maya put the exit in his mandate in one line: part of Tomás's review is how many tasks other people now do without him. “The expert's status must not depend on staying the only person everyone needs,” she said. It was awkward. It was the only way to stop the research's failure case from becoming Hartwell's.

Borrowing

A network without reserved time is another queue.

Maya's loan rule is the part of her charter I would copy first, so I will set it out. A borrowed person comes with five things written down: the hours, the end date, the receiving owner, what the lender gets back, and who decides if the loan is extended. If any of those is missing, it is not a loan. It is a favor with a half-life.

The lender's manager decides whether to lend, because they hold the person's time and the consequences of losing it. The AI Capability Manager cannot order a loan. This is the same logic as the three keys in chapter 2: the person who carries the cost holds a key. A network that does not respect this will find its lenders starting to hide their best people.

Breadth helps here. Kofi, who knew the Platform stack and the models well enough to carry a case across both, could carry a case across both without a handoff. That is the research's point about breadth: it removes a handoff. It cannot add a working day. When two recall waves landed in the same week, Kofi was exactly as busy as Tomás. I raise this to head off a mistake. The answer to a queue is not always a generalist.

Where this fails

What I would not claim.

Most of this chapter is judgment, and I say so once here. The research gives me the shapes: training can win by a hair, a mandate can fix an adviser's drift, borrowing needs reserved time. It does not give me the numbers. My view that two learners are better than one, that the contractor should cover the first quarter, and that the exit belongs in the mandate is mine.

The least comfortable case is the learner who passes and then leaves. Training makes people more valuable on the market. A company that trains must be willing to lose some of them and be better for having trained them. If you cannot accept that, buy the skill instead.

And remember that Kofi was one of two. Plan for the one who does not qualify. Ilse became an evaluator, but she could just as well have left the program unhappy. A person who fails a training program needs a next step, and giving them one is part of the plan.