How to Use AI to Write a Self-Evaluation From Your Work (2026)

To use AI to write a self-evaluation, collect the evidence first and let the AI draft second. Pull a list of what you actually did this year (projects shipped, problems solved, people you helped, feedback you got) from your sent mail, calendar and issue tracker, paste it into the assistant with your review form's questions, and tell it to use only those facts. Then edit the draft until every sentence is something you'd say to your manager out loud, and check every number against its source.

The order matters. Most bad AI self-reviews come from asking a chatbot to "write an impressive self-evaluation for a project manager". With nothing real to work from, it fills the gaps with confident, generic praise and the odd invented percentage, and managers can tell. The AI is good at structure and wording. The facts have to come from you.

The steps in short

1. Copy your review form's questions (or your company's competencies) into a note.
2. Collect evidence from the whole year, not just the last month: sent mail, meetings you ran, tickets you closed, docs you wrote, praise you received.
3. Give the AI the questions, the evidence and the prompt below, which forbids invented numbers and asks it to flag gaps.
4. Fill the gaps it flags with real figures, or cut the claim.
5. Rewrite in your own voice and check your company's AI policy before you paste anything confidential.

Where to find a year of evidence

Recency bias is the main reason self-evaluations undersell people: you remember October and forget March. Spend 20 minutes on these sources before you write a word.

• Sent mail. Search for words like "shipped", "launched", "fixed", "thanks for", "great job" and the names of your big projects. Your sent folder is a record of what you delivered and who you unblocked.
• Calendar. Recurring meetings you organized, workshops you ran, interviews you did, onboarding sessions for new hires.
• Issue tracker. Tickets you closed in Linear, Jira or similar, grouped by project. Count them only if your team cares about counts.
• Docs and wiki pages. Specs, postmortems, runbooks and guides you wrote in Notion, Confluence or OneDrive.
• Feedback. Kudos messages, customer replies, notes from your one-on-ones, last year's review and the goals you set in it.
• Numbers you can prove. Revenue, time saved, error rates, response times, but only where a dashboard or report backs them up.

Write each item as one plain line: what happened, your part in it, the result, and the month. "March: led the billing migration, 4 people, finished two weeks early, no customer tickets" is perfect input.

An AI prompt for your self-evaluation (copy and paste)

This works in any chat assistant. Paste your review questions where it says so, then your evidence list.

You are helping me write my annual self-evaluation. Use ONLY the facts
in my evidence list below. Do not invent projects, numbers, percentages,
names or outcomes. If a question can't be answered from my evidence,
write [NEED EVIDENCE: what's missing] instead of guessing.

My role: [job title, team, how long in the role]
My goals for this year were: [paste last year's goals, or "none set"]

Review questions:
[paste the exact questions from your review form]

Evidence list:
[paste your one-line items, each with the month]

For each question:
1. Answer in 120-200 words, first person, plain and specific.
2. Lead with my two or three strongest examples for that question.
   For each, say what I did, why it mattered, and the result.
3. Separate my own work from my team's. Use "I" only for what I did.
4. Include one honest area to improve, with what I'll do about it
   next year. Make it real, not a disguised strength.

Then:
- List any claim you made that a manager might ask me to back up.
- List three goals for next year that follow from the evidence.
- Avoid these words: passionate, synergy, leverage, spearheaded,
  dynamic, go-getter, rockstar.

Three details in that prompt do most of the work. "Use only the facts" plus the [NEED EVIDENCE] marker stops the invented metrics. "Separate my own work from my team's" stops it claiming a team launch as yours. The banned-words line removes the phrasing that makes reviews sound machine-written.

Can AI write my whole self-evaluation?

It can write a good first draft, but it shouldn't write the final one. Your self-evaluation is the one document where your manager wants your view of your year, and the reflection (what you'd do differently, what you're proud of, where you need support) only you can supply. Use the AI to turn rough notes into clear paragraphs, find the strongest examples, and match each one to the question or competency it answers.

Check two things before you start. First, your company's policy: some employers allow AI help as long as the final text is yours, and others ask you to use only an approved tool. Second, what you paste: leave out client names, salaries and anything under NDA unless the tool is one your employer has approved for that data.

How to pull your evidence automatically in Zemith

Collecting evidence is the slow part, and it's where an assistant connected to your work apps saves real time. In Zemith, a chat can search your Outlook mail, read your Outlook calendar, list your Linear issues and search your Notion pages, then draft the review from what it finds and hand you a Word file.

You need the Plus plan (US$15.99 a month, as of October 2026). Plugins aren't on Free.

1. Connect your apps. Open Plugins in the sidebar, pick Outlook Email, Outlook Calendar, Linear or Notion, and choose Connect account. Sign-in is with your own Microsoft, Linear or Notion account. Some work Microsoft accounts need your IT team to approve new apps first.
2. Start a New chat and pick Claude 5.5 Sonnet under Select model. It's a strong writer and it can use tools; you can switch to GPT or Gemini mid-chat to compare drafts.
3. Ask it to gather first, not write. For example: "Search my sent Outlook mail and my calendar from January to October 2026, and my closed Linear issues, and list what I delivered by month. One line per item, with the source. Don't write the review yet."
4. Correct the list. Delete anything that wasn't really yours, and add what it can't see (hallway wins, side projects, feedback from calls).
5. Paste the prompt above with your review questions, and ask for the result as a Word document. The .docx appears in the chat's Files panel, ready to paste into your HR tool.

Reading your mail, calendar and issues runs without interruptions. Anything that writes to your apps, like sending an email or saving a file to OneDrive, asks for your approval first, so the assistant can't send your draft to your manager by accident.

What does it cost? Credits depend on how much the assistant reads. As a rough guide, a run on Claude 5.5 Sonnet that reads about 40,000 tokens of email and tickets and writes a 2,000-token draft uses around 23,000 credits, a small part of Plus's 1.5 million a month. Searching a big mailbox several times costs more, so ask for one bounded date range at a time. You can see which apps Zemith connects to on the AI agent and plugins page, and how product teams use the same Linear and Notion setup on the page for product managers.

How do I make an AI self-evaluation not sound like AI?

Read it aloud. Any sentence you wouldn't say in your review meeting goes, and so does any adjective without a fact behind it. Then do three passes:

• Swap claims for examples. "Strong communicator" becomes "wrote the incident updates for the June outage, which support reused for customers".
• Cut the hedging and the hype. No "I believe I may have", and no "exceptional", "outstanding" or "world-class".
• Keep your quirks. If you'd say "fixed" rather than "remediated", say "fixed". Your manager knows how you talk.

If you used Claude for the draft, ask it for one more pass: "Rewrite this in a plainer voice, shorter sentences, and keep every fact." It's a quick way to strip the polish without losing the substance. Notion users who want this inside their workspace may also find our comparison of Notion AI and Zemith for Notion Plus users useful.

How should I rate myself on a self-evaluation?

Rate yourself where your evidence puts you, and make the rating and the examples match. If your form uses a 1-5 scale, a 4 needs at least two examples where you clearly beat what your role expects, and a 5 needs results that changed something beyond your own team. A 3 is not an insult: it usually means you did the job well. AI can't pick the number for you, but it can help with the check: "Here's my rating for each competency and my evidence. Where does the evidence look weaker than the rating?"

Honest caveats and alternatives

• If your employer gives you Microsoft 365 Copilot, try it first. It works inside your company's Microsoft 365 data, and guides like Duke's April 2026 walkthrough use it to gather a year of email, Teams chats and calendar for a review. Zemith makes more sense when your evidence is spread across Microsoft and Linear or Notion, or when you want Claude to write the draft.
• ChatGPT's apps (it renamed connectors to apps) can also reach some of these services on paid plans, depending on your plan and region. If you already pay for ChatGPT and your tools are covered, that works too.
• The assistant only sees what you connect. Slack threads, Jira, Google Workspace and anything that happened in person won't be in its list. Zemith has no Slack or Jira plugin, so add those items by hand.
• Your HR tool is still where the review lives. Zemith gives you a draft and a Word file. You paste the final text into Workday, Lattice or whatever your company uses.

The bottom line

The AI isn't the hard part of a self-evaluation. Remembering a whole year of work is. Gather the evidence first, give the AI a prompt that won't invent anything, and edit the draft until it sounds like you. If your work lives in Outlook, Linear and Notion, Zemith Plus can do the gathering for you and write the first draft with Claude in the same chat.