Teams
Retire the Second Reader One Decision Class at a Time
A new hire's AI-assisted work stops needing a second reader one decision class at a time, never on a date. List what the role decides, then retire review for the reversible internal classes first, after three or four consecutive pieces where the new hire found the error before the reviewer did. Customer-facing work graduates later. Regulated filings, cited authorities and changes nobody can roll back keep a reader permanently, no matter how long the person has been there.
The takeThe calendar rule survives because it asks nothing of the manager. Ninety days is a date anyone can read off a spreadsheet; whether this person caught the wrong figure in the reconciliation is a judgment somebody has to make and write down. So teams pick the date and call it a standard. The date then holds a careful hire under review long after the review stopped teaching anyone anything, and releases an incautious one on schedule. Pick the harder rule.
Where Olive fits
Open a role and see what the work shows
The same acts sit under every rung of that ladder: framing before generating, demanding a source for the claim that matters, keeping the judgment that should not be delegated, and refusing an output on stated grounds. Olive reads those from a role-grounded assignment done with an AI assistant, with a human reviewer writing or editing each finding against a timestamped excerpt, rather than from a self-assessment.
Assess your shortlistWhy does the ninety-day rule keep getting written?
The ninety-day rule keeps getting written because it costs nothing to write and nothing to enforce. One line in the onboarding document, legible to HR, to the manager, to the new hire and to whoever inherits the team in March. An evidence rule needs somebody to watch a piece of work and record what they saw. Teams take the legible line, then find it measures the calendar and nothing else.
The number underneath the rule is thinner than the rule looks.
In March 2012 a moving company, Allied Van Lines, surveyed human resources professionals online and reported that a new hire takes about eight months on average to reach full productivity 5.
Fifty-eight percent of the survey's 500 respondents said their company does not measure new-hire productivity at all.
The standard ramp figure in hiring is a self-report about a quantity most of the people reporting it admit they never track, collected by a relocation company, before generative tools existed.
But a real ramp curve would not settle the question anyway. Time to productivity describes output. A second reader exists to catch one class of error before it leaves the building, and that class does not shrink on a schedule.
Nothing about month four makes an unrecoverable number recoverable.
The first ninety days are still worth planning. How to structure an AI-heavy new hire's first 90 days has real answers, and not one of them is a date on which the reading stops.
Sort the work by what a wrong answer costs
Sorting starts with a list of the decision types the role actually touches, ranked on two axes: how far a wrong answer travels, and whether it can be pulled back. An internal draft and a filed figure are different objects even when the same person produces both in the same week. Aim for six or eight lines, not thirty.
Three tiers do most of the work.
- Reversible and internal. A draft, a query, a first-pass analysis another person reads before anything happens. A wrong answer costs a rework.
- Customer-visible. A message, a proposal, a number in a deck that leaves the building. A wrong answer costs credibility and a correction somebody has to send.
- Unrecoverable. A filed return, a released report, a migration with no rollback, a citation offered to a court. A wrong answer lands in a system nobody reopens.
Medicine has already written a version of this ladder down, and the vocabulary is worth borrowing because it is precise.
The Accreditation Council for Graduate Medical Education's supervision guidance for radiology residency programs says a program must use three classifications of supervision: direct supervision, indirect supervision and oversight. The guidance defines oversight as the supervising physician being available to review procedures and encounters, with feedback provided after care is delivered 1.
Oversight is easy to mistake for a control. A reader who sees the work after it shipped is a coach. A reader who sees it before is a control.
The difference decides where each class can safely land when review comes off. Reversible internal work can drop to oversight and lose nothing. Customer-visible work usually cannot.
What ends review for a class of work?
One event ends it: the new hire finds the error before the reader does, on that class of work, several pieces running. Three or four consecutive catches is a defensible bar, and the exact count matters less than writing it down in advance. The event is observable, dated, and it belongs to the class rather than to the person's tenure.
A new hire's own account of how carefully they checked cannot end it.
Neil Perry, Megha Srivastava, Deepak Kumar and Dan Boneh, computer scientists at Stanford University, ran a controlled study of people writing security-relevant functions with and without an assistant built on OpenAI's codex-davinci-002 model. Participants who had the assistant wrote significantly less secure code than those without it, and were more likely to believe they had written secure code 3.
Confidence moved in the wrong direction. A person's sense of having checked is the one signal a graduation rule must not read.
The errors that survive a read-through are the near misses. In Stack Overflow's 2025 developer survey the top-reported frustration with AI tools was output that is almost right but not quite, named by 66% of respondents 4.
Almost-right is the expensive class, and it is why a second reader is not a formality for anyone. A wrong answer that reads wrong gets caught by the author. A wrong answer that reads right gets caught by the reader, or by the customer.
Medical training's own guidance names assessments at exactly this point.
The same radiology guidance says a program's supervision policy should briefly summarize the processes and assessments the program uses to advance residents incrementally, moving from direct to indirect supervision and to oversight. That material may include, but is not limited to, clinical competency committee discussions, resident portfolio reviews with case logs, and entrustable professional activities with entrustment scales 1.
Not one of the examples it names is a month count.
Which classes never come off review?
The classes that never come off are the ones where consequence sets the rule instead of trust: a filed return, a released audit opinion, a citation offered to a court, a database migration with no rollback. Those keep a second reader at every level of seniority. The reason for the reader was never doubt about the writer. Regulated fields already run this way and say so in their own standards.
The Public Company Accounting Oversight Board requires an engagement quality review on every audit engagement conducted under its standards. A reviewer from the firm issuing the report must be a partner or another individual in an equivalent position, and should not make decisions on behalf of the engagement team. The firm may grant the client permission to use the report only after that reviewer provides concurring approval of issuance 2.
Nobody serves a probation out of it. A partner with thirty years of clean files gets the same second reader as one in a first season. The rule is about what an unreviewed audit report can do, not about who wrote it.
Write that rule into your own ladder.
Naming the permanent classes early settles two arguments at once. The first is about which parts of the work to decide not to use AI on. The second is about who answers when a piece of AI-assisted work turns out wrong, and naming the human who checks it is only possible where a checker is assigned by rule rather than found afterwards.
Say the whole ladder in week one
Hand the new hire the ladder on day one, in writing, with the classes named and the graduation event stated. A reviewer who disappears without explanation reads as a verdict nobody delivered. A reviewer who comes off a named class after four recorded catches reads as a milestone, and the new hire can work toward the next one.
Role clarity is not a soft extra. Talya Bauer and four colleagues tested a model of newcomer adjustment across 70 unique samples. They reported that adjustment, meaning role clarity, self-efficacy and social acceptance, mediated the effects of socialization tactics and information seeking on job satisfaction, commitment, performance, intentions to remain and turnover 6.
The authors reported that the results generally supported the model. The design is a path model fitted over correlational studies, so it carries a direction and not a magnitude.
A ladder is role clarity with an event attached to each rung.
Counting catches has one real weakness: a missed error is not always a missing habit.
A new hire who does not yet know that your reconciliation excludes intercompany balances will not catch a wrong figure there, however carefully they read. That is the question of whether a missed error is a verification failure or a knowledge gap.
Counting misses punishes the wrong thing. Counting catches, and recording what was checked against what, does not.
The line in the onboarding document is the thing to rewrite, and it comes out longer: a short table of decision classes, the event that retires each one, and the ones with no retirement at all.
Audit firms working under PCAOB standards have a fixed date for their own rule change: the amended engagement quality review standard takes effect on 15 December 2026 2. Your table has no such deadline. Write it before the new hire's first piece of work, not after the first mistake.
Common questions
How long should a new hire's AI-assisted work be reviewed?
There is no single duration, because different work carries different consequences. Reversible internal work can come off review within weeks if the person is catching their own errors before the reader does. Customer-visible work takes longer, because the errors that matter there are the ones that read as correct. Some classes never come off: regulated filings, cited authorities, and changes with no rollback keep a second reader regardless of tenure. Set the ladder before the first piece of work arrives, and record the graduation event for each class.
What counts as evidence that review can come off a class?
A recorded event, not an impression. The event worth using is the new hire finding the error before the reviewer does, on that class of work, several pieces in a row. Three or four consecutive catches is a defensible bar, and the count matters less than fixing it in advance and writing each instance down. The reviewer records what was checked and against what, so the file shows a habit rather than a run of easy work. A new hire's own report of how carefully they read is not evidence, since in one controlled study, people using an AI assistant wrote less secure code and were more likely to believe it was secure.
Does a senior hire start higher on the ladder?
On classes where they arrive with a record you can inspect, yes. On classes that turn on how your systems, data and customers actually behave, no, because seniority elsewhere is not evidence about the reconciliation your team runs or the client who reads the deck. The practical version: start a senior hire on the same rung as anyone else for domain-specific classes, and expect them to clear it in a fraction of the pieces. Say that at the offer stage, so the ladder reads as a standard rather than as distrust.
Should the second reader check everything or sample?
The class decides, which is the same principle running underneath the whole ladder. On reversible internal work a sample is enough, because a missed error costs a rework and the next sample catches the pattern. On unrecoverable work sampling defeats the purpose, since the reason for the reader is that one wrong item cannot be pulled back. Between the two, sample heavily at first and thin out as catches accumulate, and record which pieces were read so the graduation decision rests on something.
What if the new hire never produces a catch?
Check the work before checking the person. A run with no catches can mean the assignments contained no errors worth finding, which is common when a new hire is handed low-stakes tasks for a month. Give them a piece of work with a known problem in it and see what happens. If catches still do not appear, review stays on, and the conversation is about the specific class and the specific misses rather than about readiness in general. Record it either way, because an absence of evidence and evidence of a gap look identical when nobody wrote anything down.
Does the second reader's own AI use change the rule?
Yes. A reader who runs the work through a model and accepts what comes back is not a second reader, because the same class of confident error survives both passes. Name what the reader checks against: the source document, the filing, the ticket, the query rerun on real data. A reviewer who cannot say what they checked against has not produced a check, and a graduation record built on those reviews overstates what the new hire can do.
References
- 1. Review Committee for Radiology Supervision Policy Guidance ✓ acgme.org Supports the requirement that a program use three classifications of supervision (direct supervision, indirect supervision, oversight), the definition of oversight as the supervising physician being available to provide review of procedures and encounters with feedback provided after care is delivered, and the guidance that a supervision policy should briefly summarize the processes and assessments used to advance residents incrementally, which may include but is not limited to Clinical Competency Committee discussions, resident portfolio reviews with Case Logs, and entrustable professional activities with entrustment scales.
- 2. AS 1220: Engagement Quality Review ✓ pcaobus.org Supports the claim that an engagement quality review is required on every audit engagement conducted pursuant to PCAOB standards, that a reviewer from the firm issuing the report must be a partner or another individual in an equivalent position and should not make decisions on behalf of the engagement team, that the firm may grant the client permission to use the report only after concurring approval of issuance, and that the amended standard is effective on 15 December 2026.
- 3. Do Users Write More Insecure Code with AI Assistants? ✓ arxiv.org Supports the claim that participants with access to an AI assistant based on OpenAI's codex-davinci-002 model wrote significantly less secure code than those without access, and were more likely to believe they wrote secure code.
- 4. 2025 Stack Overflow Developer Survey: AI ✓ survey.stackoverflow.co Supports the claim that the top-reported frustration with AI tools was solutions that are almost right but not quite, named by 66% of respondents.
- 5. 2012 Allied Workforce Mobility Survey: Onboarding and Retention ✓ allied.com Supports the provenance of the eight-month time-to-full-productivity figure: a March 2012 online survey of 500 US human resources professionals, in which 58 percent said their company does not measure new-hire productivity at all.
- 6. Newcomer Adjustment During Organizational Socialization: A Meta-Analytic Review of Antecedents, Outcomes, and Methods ✓ pdxscholar.library.pdx.edu Supports the claim that adjustment (role clarity, self-efficacy and social acceptance) mediated the effects of socialization tactics and information seeking on job satisfaction, commitment, performance, intentions to remain and turnover across 70 unique samples of newcomers, with the results generally supporting the model. No effect size is claimed.
6 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.