I know the project belongs on the page. What I can’t remember is what it was worth - the number is in an inbox somewhere, a SharePoint folder, and four versions of the same model, and I have an evening to find it.
Most of us have had that evening. The work was real, the outcome was real, and the detail that would make it legible to anyone else has quietly dissolved into storage nobody organised. So we write the thin version and move on.
That evening happened to me recently, while I was preparing for an internal recruitment process. What I took out of it had almost nothing to do with recruitment and a great deal to do with how we prepare for a review of any kind: the self - assessment in the appraisal system, the budget round, the business case, the monthly review - choose your fighter.
A recruitment process is just a review where somebody bothered to write the requirement down.
A review is a proposal, not a presentation
Think about it as a deal. There’s a request for proposal, and there’s the proposal answering it. A CV is a proposal. So is a business case, a budget submission, a summary of a monthly financial performance, and the self-assessment sitting in your appraisal system very soon. In each one, somebody in the room is holding a decision, and the job is to meet the requirement behind it rather than present a complete account of our work and hope the relevance is obvious.
The difference sounds academic until you watch both versions land. The account of work is thorough, reconciled, and strangely inert. The proposal moves something. Under time pressure we tend to produce the first one - accurate, exhaustive, and answering a question nobody asked.
In a recruitment the requirement is handed over. A job description is an unusually honest document: someone has written down what they need, in their own order, with their own emphasis. Before I opened any of my own material I put it in front of me and treated it as the specification it is.
Most reviews don’t come with a clear requirement, so it has to be reconstructed - who is sitting there, what they’re being asked to approve, what they’re personally accountable for, and what would make them say no. That reconstruction is the preparation; everything after it is assembly. It also tells you what isn’t required, which is the harder half. We over-prepare the material we enjoy: five slides on the market that’s performing, one on the thing that needed addressing. We rarely actively and consciously decide to do that. It’s what happens when the comfortable section is the one that expands at eleven at night.
Understand the request, then understand what you have, then work out which of it meets the requirement, then rehearse the gaps rather than the answers. The order carries more weight than the steps - starting at the second part (what we have) is how you end up with a beautiful deck answering the wrong question.
What AI did with four years of my own material
Here’s where I expected the tool to do something clever, and the useful part turned out to be elsewhere.
I had material: four years of self-assessments, my manager’s assessments, objectives, development items, notes written as projects closed. None of it was written for this particular purpose. It existed because I keep a record, and it sat in an archive I almost never reopened.
I put the relevant years in front of Copilot together with the job description, and told it to invent nothing, exaggerate nothing, and help me surface what genuinely met the requirement.
Lens work: filtering what I already had
It didn’t find anything that wasn’t there. Everything it surfaced, I had written myself. What it did was read four years of my own material against a specification and filter it - work I could have done by hand over a weekend, and would have done worse, because by hour three I’d have been skimming.
That’s worth being precise about, because it’s the version most people recognise and the version most posts overstate. The tool is a lens. The material is yours, the judgment is yours, and what comes back is your own record sorted by relevance to a question you supplied.
Retrieval: finding what I’d lost
Then it did something I couldn’t have done at all.
There was a transformation project that belonged on the page, and I couldn’t recall the detail. The archive had never quantified it - I’d written down that it happened, not what it was worth. But the evidence existed: in the thread where the scope was agreed, in the files, in whichever version of the model turned out to be final. A tool sitting natively inside that environment reads across all of it.
What came back was the scale, the unit and a figure of an annualised saving. I verified it against the source before it went anywhere near the page. Verification isn’t optional, and it never becomes optional. But honestly, I I wouldn't have found it in reasonable time without working through my context with an AI tool.
Two different jobs, worth separating. Against material you deliberately kept, the tool does lens work - filtering what you already hold against a requirement. Against the exhaust of everyday work that nobody organised, the threads and attachments and version four of the model, it does retrieval. The second only works because the tool lives where the work happens, and it’s the one that saves you from the thin version.
The constraint that doesn’t make it into the headline
The retrieval half comes with a condition attached, and it’s worth saying plainly because it cuts against the enthusiasm.
It only works inside the environment where the work actually sits, with your permissions, under whatever your organisation has already agreed to. That makes it a real argument for the platform you’re on - which is also an argument for lock-in, and anyone selling you the capability will be slower to mention the second half.
Three things follow. The filtering is portable: any tool you can hand your own documents to will do it, so you can use the one you prefer and the one your policy allows. The retrieval isn’t, and no amount of prompt craft substitutes for access. And both inherit whatever your organisation permits - if retention cleared the 2023 mailbox, there is nothing to retrieve, and if the model that mattered lives on someone’s personal drive, it isn’t in scope either.
Worth establishing which of those applies to you before you plan an evening around it.
How to use AI as your reviewer, not your supporter
Assembled context earns its keep when you stop using it to support the case and start using it to attack it.
I asked for the reviewer’s position instead of my own: what would this be assessed on, where is the evidence thin, what gets pushed on first. I asked for disagreement explicitly, because a tool that agrees with us is producing a transcript of our own assumptions in a more confident way.
The gaps were the output. Not all of them could be closed, but knowing which ones were real changed what I prepared and what I was willing to say I didn’t yet have.
The same move works through the lens of the role rather than the person. Annual reports, strategy documents, external announcements - all of it reads differently against a specific question than it does in general.
Then rehearse against the actual room: how long, who’s in it, where the first question usually comes from. Less about having answers ready, more about finding out which parts of your own case you can’t yet say out loud.
Why AI can’t prepare you without an archive
None of this works without the record, and the record is the part nobody writes posts about.
The tool didn’t build my context. I did, over the years, in a habit that felt purely administrative at the time. A paragraph at the end of a project. The self-assessment I’d have written anyway. Notes that existed because writing them took ten minutes and I’d already decided that was cheaper than reconstructing them later.
It’s worth being clear about what that habit actually buys, because it does two separate things. It gives the filtering something to filter: without it you get your own job description handed back to you in different words. And it gives the retrieval something to anchor to. I knew the project existed and roughly when, which is what made the search tractable in the first place. Retrieval across four years of everything, with no idea what you’re looking for, returns noise.
So, the honest version of this isn’t about tooling at all. Keep the record: what a project was for, what changed, what it was worth, written down while you still remember. Everything above is downstream of that, and the preparation only felt quick because the slow part had already happened, in pieces, years earlier.
The tool isn’t the differentiator. The context is.
Questions people frequently ask about using AI for review preparation
Does this only work with Copilot? The filtering works with any tool you can hand your own documents to. The retrieval doesn’t. Finding a number inside four years of emails and file versions needs a tool that already lives in that environment, with your permissions and inside your organisation’s policy.
Can AI write my performance review self-assessment? It shouldn’t, and the output gives it away. What it does well is read material you wrote yourself against the requirement you’re answering, and tell you which parts are relevant. The writing stays yours.
What should I give it before asking anything? Two things, in this order: the requirement - a job description, a meeting agenda, the decision being taken - and your own material. Without the first you get a summary of yourself. With it you get a filter.
What if I haven’t kept a record? Then this is a slower month, and the useful move is to start now. The next review is always coming.