The Six Weeks Before Access Are Not Evidence About the Hire

What do six weeks without the tool show about the hire?

Less than the calendar suggests, and much less than a manager forming an early view believes. A hire selected for AI capability, working six weeks without the tool, produces output under conditions the job never has. The work is real and the sample is not.

Whatever goes in the week-three check-in describes the constraint at least as much as it describes the person.

What the wait withholds has been measured, and it is worth most to the person with the least behind them. Across a staggered rollout of a generative AI assistant to 5,179 customer support agents at one software firm, access raised issues resolved per hour by 14% on average ¹. Inside that average, novice and low-skilled agents gained 34%, and the experienced ones gained almost nothing ¹.

Everything in that number comes from one software firm and one occupation, from an assistant trained on the company's own transcripts and rolled out in 2020 and 2021. What travels is the direction rather than the size.

Different work, a larger pool, the same shape. Three company-run randomized trials at Microsoft, Accenture and an anonymous Fortune 100 firm pooled 4,867 developers. Access to an AI coding assistant raised completed tasks by 26.08%, and the less experienced developers both adopted it more and gained more ². The standard error on that number is 10.3%, so the honest range runs from roughly 6% to 46% ².

So in both settings the tool helped the least experienced most, and the six weeks withhold it and then hand you the output as evidence about them.

But the ramp was already two numbers.

Time to usable output and time to unsupervised judgment came apart in AI-heavy work, and the delay blocks only the first. What this team refuses to delegate, and who to ask when a figure looks wrong: none of that waits on a licence.

Start the access request when the offer is signed

File the request the day the offer is accepted rather than the day the person arrives. A security or vendor review runs on its own queue, and that queue starts when the ticket exists. A request filed on day one puts the whole wait inside the ramp.

Name the person, the tool, the data it will touch, and the first deliverable that depends on it.

The requisition is where the dependency belongs, because whoever wrote it usually knows best which work needs the tool. The manager writes the 30-60-90 before the start date, and the access line goes in the same document with a date beside it.

A finalist who asks about tooling during the offer conversation has already told you something worth knowing.

Some requests never clear at all, which is a different problem wearing the same clothes. Requiring AI skills while security blocks the tools publishes a contradiction, and hiring against it produces someone selected for work nobody will let them do.

Since 2 February 2025, an EU employer has carried an obligation on the other side of the access date. Article 4 of the AI Act asks providers and deployers to take measures supporting the AI literacy of their staff, and an employer running an off-the-shelf tool is a deployer ³.

The Commission's own questions and answers on Article 4 say there is no need for a certificate and no obligation to measure an employee's knowledge of AI ³.

On the Commission's reading, then, a training certificate is not a condition of access.

What belongs in the six weeks that does not need a model?

Everything the tool would have skipped past. Last quarter's work and the accounts behind it, reviews attended without a deliverable, the claims this team checks before they leave the building, and the short list of decisions nobody here delegates. Hand over one real task with a genuine ambiguity in it, and watch how it gets handled.

Newcomer research points at a similar list. The standard meta-analysis of newcomer socialization, fitted on 70 unique samples, models early adjustment as three things: knowing the role, believing you can do the job, and being accepted. Those three carry through to satisfaction, performance, intent to remain and turnover ⁴.

The meta-analysis is a path model fitted over correlational studies from before remote onboarding, so read it as a direction and not a dose.

A licence buys none of the three.

And week one is where an unrehearsed look at how someone works is cheapest to get, because the habits still on display are the ones they arrived with.

Say the tool rule out loud before the first Monday, because the alternative is already ordinary.

Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 knowledge workers across 31 markets fielded in February and March 2024, found 78% of the people who use AI at work bringing their own tools to it ⁵.

In the same survey, 52% of those users said they are reluctant to admit using AI for their most important tasks ⁵.

So a new hire with a blocked account and a personal one has a reason to reach for the personal one, and the work read in week four may have been done somewhere nobody approved.

Decide now what may never be pasted into an unapproved tool, name the fallback for everything else, and put both in the onboarding note.

Put the first AI-assisted deliverable after the access date

Move the first AI-assisted deliverable to the week after the licence lands, and write the date in the plan. A check that drifts tends to drift past probation, and the question everybody had at the offer stage then outlives the period meant to settle it.

One piece of genuine work, done with the tool, read by somebody who knows the domain.

What to read for is the set of questions the work was supposed to settle. What got framed before anything got generated, which claim was checked and against what, what the person kept out of the model's hands, and what came back and was thrown away.

Read the decisions inside the deliverable rather than the deliverable.

A first AI-assisted piece of work is also where a second reader earns the seat, and when that work stops needing one is a call better made by type of decision than by a date.

If the tool does not move this work, the six weeks withhold nothing.

Sixteen experienced open-source developers took on 246 real tasks in a randomized trial, on repositories they had known for about five years. Early-2025 AI tools made them 19% slower ⁶. They had forecast a 24% speedup, and afterwards still believed they had gained 20% ⁶.

One setting, sixteen people, code they knew intimately. The part worth carrying is not the 19%. It is how far their sense of the work sat from what was measured, and that distance runs in both directions, including yours about a new hire.

Write the six weeks into the requisition, not into the review

Put the delay where it is a cost, which is the requisition, and keep it out of the performance record, where it becomes somebody else's constraint charged to the new hire. Six weeks off the front of every AI-selected hire belongs next to time to fill, in the conversation about what this role costs to staff.

The figure usually reached for will not carry that argument. The claim that a new hire takes about eight months to become fully productive comes from a March 2012 online survey of 500 US HR professionals, sponsored by a moving company ⁷.

The same survey reports 58% of respondents saying their company does not measure new-hire productivity at all ⁷.

So the only ramp number with your name on it is the one recorded from now: the date the request was filed, the date access arrived, and the work that could not start in between. Three fields per hire. Keep them, and the wait stops being a complaint about IT and becomes a figure the next requisition can carry.

A ticket filed on day one tells you one thing, which is how long the queue was. Filed at offer acceptance, dated, and named in the onboarding plan, it becomes the line in the file that says which six weeks of this person's work were never evidence.

See the benchmarks >