Give AI to your newest hire first

Summary

The instinct in most companies is to give new tools to their strongest people, who will "know what to do with it." Three of the best field studies of generative AI at work point the other way: the largest gains go to the least experienced, the top performers gain little, and the clearest benefit is speed to competence. For a small business, where every new hire is a large share of the team, we argue that the ramp is the main event. We also cover the risk that a fast ramp produces dependence instead of skill.

The evidence

Brynjolfsson, Li and Raymond studied customer support agents given an AI assistant [1]. Issues resolved per hour rose about 14% on average and about 35% for novice and lower-skilled workers, with little or no effect for the most experienced. Agents with two months of tenure who used the tool performed as well as agents with six months of tenure who didn't. Attrition also fell.

Dell'Acqua and colleagues ran an experiment with management consultants using GPT-4 [2]. On tasks within the model's abilities, below-average performers improved by 43%, compared with 17% for above-average performers. Noy and Zhang found the same compression in professional writing tasks: time fell, quality rose, and the gap between weaker and stronger writers narrowed [3].

The likely mechanism is that these tools spread the tacit know-how of top performers, the phrasing that works, the step people forget, to everyone. Top performers already have it. Novices get it on day one.

Why novices gain the most

Most work that looks routine runs on tacit knowledge: the phrasing that calms an upset customer, the step that prevents a callback, the question that saves an hour. Experienced people carry it; new people learn it slowly by watching and failing. AI tools built from records of past work make some of that knowledge available on day one, in the moment it's needed. For a veteran, the suggestion repeats what they already know. For a novice, it's the missing piece.

What the studies did and didn't measure

These are strong studies by the standards of the field, and their limits matter. The support study measured issues resolved per hour inside one company's software. The consulting study graded task output in an experiment lasting days, not careers. The writing study used short, realistic tasks with paid participants. None followed workers for years, and none measured whether people who ramped up with AI could later work well without it.

So the claim we make is narrow: in the first months on a job, the evidence consistently favors the least experienced. What happens after that is still an open question, and our measurements are designed to help answer it.

The ramp is the main event

For an owner-operator with 5 to 50 people, onboarding is expensive. Each new hire pulls time from the best people, and mistakes in the first months land on customers. Compressing a six-month ramp into two is worth more to a business that size than shaving minutes off the work of its veterans.

AI is a ramp, not a raise: it helps your newest people most, so give it to them first.

The uncomfortable parts

Speed to performance is not the same as speed to skill. In a school setting, students who used an unrestricted AI assistant did better on practice and worse once it was taken away [4]. A new hire who performs like a veteran with AI may not be one without it.

There's also a frontier. In the consultant study, on a task outside what the model could do well, consultants using AI were 19 percentage points less likely to get the right answer [2]. Novices are the least equipped to notice when they've crossed that line.

Finally, a technology that makes novices nearly as productive as veterans changes the value of experience, and with it, pay and promotion. That's a question for every business owner, not just economists.

The veteran's new job

If AI narrows the gap between new and experienced staff, the veterans' value shifts. They become the editors and teachers of the system: the people who catch what the AI gets wrong, whose outcome notes become onboarding material, and who handle the cases that fall outside what the model does well.

That shift only works if veterans are willing to share what they know. Research on knowledge hiding in organizations shows that people often withhold knowledge when they feel it protects their position [5]. A business that wants its best people to teach the system needs to reward them for it, not quietly treat their know-how as something to extract and replace.

Why small businesses feel this most

Large companies have training departments, documented procedures and enough staff to absorb a slow ramp. A business with twelve people has none of these. Onboarding falls on the owner or the one person who knows how things work, and every hour they spend teaching is an hour not spent on customers. That is why we think the ramp, more than any productivity gain for existing staff, is where AI pays off first for owner-operators.

How Crew uses this

New hires get a Crew member from day one, and Teach builds their onboarding from the business's own outcome notes: how work was actually done and decided, not a generic manual. The goal is that the AI works as a ramp the person climbs, not a crutch they lean on. Coach flags tasks that look like they sit outside what the model does reliably, and sends them to an experienced person.

An onboarding plan

1. 01Week 1: the new hire shadows their Crew member and reads Teach lessons built from the business's own outcome notes.
2. 02Weeks 2 to 4: real work with AI assistance, with Coach routing anything unusual to an experienced colleague.
3. 03Days 30, 60 and 90: a short block of standard tasks done with AI switched off, to check that skill is building, not just output.
4. 04After 90 days: the new hire's own outcome notes start feeding the lessons for the next person.

Running the AI-off check fairly

Switching the AI off at 30, 60 and 90 days only helps if new hires don't experience it as a trap. We frame it as practice, not evaluation: a short set of standard tasks, done alone, with the results discussed with the new hire and used to decide what Teach covers next. The point is to find where skill hasn't formed yet while the stakes are low, not to catch anyone out.

Starting this week

1. 01Pick the role you hire for most often and write down the five tasks a new person does in their first month.
2. 02Give the next new hire an AI assistant from day one for those tasks, and ask your most experienced person to review their work for the first two weeks.
3. 03Ask every person, new or experienced, to close tasks with a short outcome note. Those notes become next quarter's onboarding material.
4. 04Measure weeks to independence, and compare with the last person you hired without it.

Where we might be wrong

• The studies cover customer support, consulting and writing. Trades, healthcare and other hands-on work may behave differently.
• Short-run gains may fade. Most studies measure weeks or months, not years.
• Veterans may matter more than the data suggests, because they're the ones who catch the errors novices can't see.

What we're measuring

• Time to independent proficiency: weeks until a new hire handles a standard task without help.
• Performance with AI switched off at 30, 60 and 90 days, to separate skill from dependence.
• Veteran correction rate: how often experienced staff fix work produced by AI-assisted new hires.