Document Underperformance Without Letting AI Characterize a Person

Workflow area: Hiring & Performance · Tier: Core · Hands-on time: ~30 minutes

This is the highest-stakes writing a manager does. Done well, it gives a struggling person their clearest possible chance; done sloppily, it becomes evidence — of your bias, in a dispute. The failure mode AI amplifies is characterization: models reach for fluent explanatory prose ("shows a lack of ownership," "doesn't seem invested"), and every one of those phrases is a claim about someone's inner state that no document can prove and any employment lawyer can attack. The discipline is the same one HR has always taught — facts, expectations, support, timeline — with AI doing the structuring and a hard rule doing the protecting.

When to use — and when not to

• Use it when informal feedback hasn't landedyou've given clear verbal feedback (pb-03) at least twice, documented in your notes, and the gap persists; documentation is the next rung, not the first.
• Use it to prepare the HR conversationarriving with a factual log and a drafted plan makes the HR partnership fast; arriving with frustration makes it slow.
• Use it for the plan's check-in cadencethe weekly written check-ins during a plan are the same facts-expectations-support structure in miniature.
• Skip it if the log doesn't exist yetdocumentation drafted from memory is recency-biased and thin; start the factual log today (step 1) and revisit in three weeks. An improvement plan built on vibes fails the person and the process.
• Skip it entirely if the real issue is something elsea performance frame on what is actually a role-fit, team-conflict, or personal-crisis situation compounds the problem; that's a conversation with HR first, not a document.

The recipe

1. 1.

   Keep the factual log before you need it. A running, dated record per concern: what was expected (with the artifact — the ticket, the agreed deadline, the design-review action), what was observed (what shipped, when, what broke), and the impact. One entry per event, written the week it happens. No adjectives, no theories: "missed the Aug 4 date agreed in sprint planning; feature shipped Aug 18 after two reviews found unhandled error paths" — not "keeps missing deadlines, work is careless."

2. 2.

   Draft the plan from the log, with the intent rule in the prompt. Enterprise tenant only, one person per session:

   Copy

   Draft a performance improvement plan from my factual log. HR
   will review before anything is shared.
   
   HARD RULES:
   - Describe observable behavior and work outcomes ONLY. Never
     characterize intent, attitude, motivation, or personality.
     Banned: "careless", "lacks ownership", "unmotivated",
     "doesn't care", "bad attitude", and any synonym.
   - Every stated gap must cite a dated entry from my log. If
     the log doesn't support a gap, leave it out and tell me.
   - Expectations must be specific, observable, and achievable
     in the plan period — behaviors a reasonable person could
     demonstrate and a reasonable manager could verify.
   
   Structure: (1) purpose and support framing, (2) specific
   gaps with dated examples, (3) expected standard per gap,
   (4) support we will provide, (5) check-in cadence and how
   progress will be assessed, (6) timeline.
   
   Log: [PASTE FACTUAL LOG]
   Role expectations: [PASTE LEVEL/ROLE STANDARD]
   

3. 3.

   Audit the draft for smuggled characterization. Models leak intent-language even when told not to. Run a second pass: "List every phrase in this draft that asserts or implies anything about the person's mindset, effort, or attitude rather than observable behavior." Good output catches the subtle ones:

   Copy

   - "failed to prioritize the migration" — implies a choice/
     mindset. Observable version: "the migration tasks agreed
     on Jun 2 were not started as of Jun 30 (standup notes)."
   - "did not take review feedback seriously" — intent claim.
     Observable version: "the same error-handling issue was
     flagged in three consecutive reviews (May 8, May 22,
     Jun 5) and recurred in each subsequent PR."
   

   Rewrite every flagged phrase. This audit is the playbook's core move — it's also worth running on documentation you wrote entirely by hand.

4. 4.

   Check the expectations are a fair path, not a trap. For each expected standard, ask: could a person meeting the role's level genuinely do this in the time given, with the support listed? Plans with vague standards ("significantly improve quality") or impossible ones ("zero review comments") read as pretext — because they usually are. The support section is not decoration: named mentor, adjusted scope, specific training; if you can't fill it, you're documenting an exit, not an improvement plan, and HR needs to know which this is.

5. 5.

   HR reviews before the person sees anything. No draft reaches the employee without HR sign-off — they check legal exposure, precedent consistency (is this standard the same one applied to others at this level?), and process requirements in your jurisdiction. Deliver the plan in person (pb-04 to rehearse the conversation), give the person space to respond, and log the check-ins with the same factual discipline; the plan period's record matters as much as the plan.

Make it yours

• Small team or startup: no HR department? Get outside counsel or a fractional HR advisor to review before delivery — early-stage companies lose employment disputes on process, and the factual-log habit is cheap insurance either way.
• Large org or GCC: use HR's template and process exactly — this playbook improves the content that goes into it, not the workflow around it. India-specific: involve HR early on notice-period, documentation, and severance implications; employment-law requirements differ materially from the parent org's home jurisdiction, and a US-template PIP can be procedurally defective locally.
• Remote teams: observable behavior must be genuinely observable remotely — commit history, review threads, meeting participation, written communication. Never let presence proxies (response latency, green-dot hours) into the log; they measure time zones and caregiving, not performance.

Quality check

• Every gap cites a dated log entrya claim without a date and artifact is an opinion, and opinions lose disputes.
• The characterization audit returned zero unfixed phrasesgrep the final draft for "attitude," "ownership," "motivated," "careless" and their cousins; any survivor is a liability sentence.
• The standards are achievable and the support is reala plan you privately believe cannot be passed is a termination process wearing a support costume; name it honestly with HR.
• HR saw it before the employee didno exceptions, including the check-in summaries during the plan.
• The same rigor applies to everyoneif this person's plan is documented to this standard, past and future plans at this level must be too; inconsistent rigor is itself discoverable evidence.