01 the business
Software company
120 staff, twice-yearly reviews
At review time, managers try to remember six months of work and write mostly about recent weeks. Peer feedback sits in forms nobody reads closely, and review quality depends on which manager you have.
02 before
The problem
When review season comes, managers try to remember six months of work and end up writing about the last few weeks. Feedback from peers sits in forms nobody reads closely, and the quality of reviews depends on which manager you have.
- Reviews reflect recent memory, not the whole period.
- Peer feedback is skimmed.
- Quality varies by manager.
03 the build
What the AI does
- Step 1: Collects goals, peer feedback, self-reviews and one-to-one notes from Lattice, BambooHR, Google Docs or Notion for each employee.
- Step 2: Groups the evidence by goal and by the competencies in your review framework.
- Step 3: Drafts a balanced summary of strengths and areas to work on, linking each point to the note or feedback it came from.
- Step 4: Flags gaps, such as a goal with no evidence or feedback that contradicts the self-review, for the manager to look into.
- Step 5: Checks the wording for vague or loaded language and suggests plainer, more specific phrasing.
paused: waiting for a person to approve
How it was put together
- Collected goals, peer feedback, self-reviews and one-to-one notes for each person.
- Grouped evidence by goal and by the review framework's competencies.
- Drafted a balanced summary linking each point to its source.
- Flagged gaps and contradictions, and suggested plainer wording for vague or loaded phrases.
04 human in the loop
Where a person stays in charge
The manager edits and owns every review, and HR checks ratings across the team before anything is shared with the employee. Pay and promotion decisions are made by people, not by the summary.
- Managers edit and own every review.
- HR checks ratings across teams before anything is shared.
- Pay and promotion decisions are made by people, never by the summary.
05 rollout
From kickoff to live in 3 weeks plus one review cycle
- Weeks 1–2Connect sources and map the review framework; agree data access with HR.
- Week 3Pilot with three managers in the next cycle.
- Next cycleOffer to all managers.
06 after
What changes
Reviews cover the whole period, not just the last month, and every point can be traced to something real. Managers spend their time on the conversation instead of the first draft.
Assume 15 managers writing 8 reviews each, at 2 hours per review, falling to 1 hour with a draft.
15 × 8 = 120 reviews. 120 hours saved per cycle, or 240 hours a year with two cycles.
What to measure
- Manager time per review.
- Share of reviews citing evidence from the whole period.
- Employee views on fairness.
- Consistency of ratings across teams.
07 evidence
What the research says
AI used to recruit or select people, or to make decisions about promotion, termination, task allocation or performance, is classed as high-risk.
EU AI Act, Annex III: high-risk AI systems (opens in a new tab) · 2024
NIST's framework organises AI risk work into four functions, Govern, Map, Measure and Manage, with a profile for generative AI added in July 2024.
NIST, AI Risk Management Framework 1.0 and Generative AI Profile (opens in a new tab) · 2023
In a study of 453 professionals doing writing tasks from their own jobs, access to ChatGPT cut the time taken by 40% and raised quality, as rated by evaluators, by 18%.
08 lessons
Mistakes to avoid
- Using the draft as the verdict.
- Pulling in data people did not expect. Be clear about what is used.
- Skipping the cross-team check.
09 questions
Common questions
- Is this high-risk under the EU AI Act?
- AI used to make decisions about promotion, termination or performance is classed as high-risk. Keeping the summary as input to a human decision, with HR oversight, is central to the design.
- Will employees see the draft?
- Only the final review the manager writes and HR approves.