AI Employee Performance Tracking From Work Records, Not Recall

Ask a manager to rate someone’s last six months and you get a clear memory of the last three weeks. The useful version of AI employee performance tracking attacks recall rather than judgment: the system assembles the evidence out of work records the company already holds, and a person still decides what it means. We run one weekly written scorecard across staff and agents, and the rubric underneath it is the same for both.

Your agents are measured, your people are remembered

The asymmetry shows up in the tooling. Google Cloud’s agent observability documentation, last updated 3 September 2026, lists what an AI agent is measured on by default: total invocations, latency at the 50th, 95th and 99th percentile, error rates per tool, token usage, and quality metrics covering response quality and hallucination rate. That is instrumented telemetry, collected whether anyone asks for it or not.

Now look at the person sitting next to it. Deloitte’s 2025 Global Human Capital Trends, a survey of roughly 10,000 business and HR leaders across 93 countries, found 75% of organisations rate their own ability to accurately evaluate individual value creation as not very or not at all effective. Only 6% said they were doing great things with data and evidence. Those are self-reported figures about a process that itself runs on self-reported opinion, the same gap a commission plan hits when nobody can evidence the number.

The trust numbers follow. In the same 2025 survey, 61% of managers and 72% of workers could not say they trust their organisation’s performance management process.

What the weekly scorecard actually scores

Ours reads task records, not people. Every task carries an assignee, a department, a priority, dates for start, due, any revision and completion, plus the closing evidence. The score comes from four inputs: base points set by priority, a business impact multiplier, a timeliness multiplier, and a quality modifier. Sales roles add an attainment score against the target committed to that week.

Every number decomposes. Open any score and you get the per-task breakdown: which multiplier applied, the reason recorded for it, and the task record it came from. The subjective parts, impact and quality, get judged once per task against a fixed rubric and then cached, so re-running last month does not quietly produce a different answer.

The weights and thresholds live in configuration rather than code, which means retuning the rubric after seeing real results is a settings change instead of a deploy.

Verification runs before scoring

A task marked done is a claim, not a fact. Before anything is scored, a verification layer checks completed tasks against typed, machine-checkable conditions written when the job was defined: a linked page returns a live response, a file exists where it was meant to land, a record exists in the system it was meant to reach.

Tasks with no checkable condition pass quietly. A failed check bounces back to the assignee. An inconclusive result goes to a human supervisor to confirm by hand, because a verifier that guesses is worse than no verifier at all. The chasing side of that layer is written up in ai task follow up automation.

The two loopholes we had to close

Version one could be gamed two ways, and both showed up in real data.

The first was leaving work open. An unfinished task is not a failure yet, so an overdue-but-open item scored better than a late-but-finished one. Open work now still counts toward workload but is excluded from the quality score until it closes. You cannot look productive by never finishing.

The second was the self-granted extension. If moving your own deadline counts as a legitimate revision, nobody is ever late. A date change now only counts once the person who assigned the work confirms it. Otherwise the original date stands and the timeliness penalty keeps compounding.

One rubric for people and agents

Scoring agents on the same engine came after the human version was already running. It needed almost no change, because the inputs are identical: a task, an owner, a due date, an acceptance condition, and evidence of what was produced. An agent that ships late and gets reopened twice looks exactly like a person who does the same.

That matters more this year than last. Stanford HAI’s 2026 AI Index, published 13 April 2026, reports agent task success on real-world benchmarks rising from 20% in 2025 to 77.3% in 2026, which is measured benchmark data rather than a survey. The same report puts actual agent deployment in the single digits across nearly every business function, which is self-reported. Capability is not the constraint any more. Knowing whether the thing did its job is.

As of 2026-06-30 we ran 41 registered agents, 9 live in production and 5 in beta. Each one is a tomte, our word for one production agent with one defined job, and each one gets a line in the same weekly report the people get.

Three ways to answer “how did that person do”

	Annual review	Activity dashboard	Evidence-scored weekly report
Source of the rating	Manager recall	Counts of activity	Task records plus verification
Covers AI agents	No	Sometimes, separately	Same rubric as people
Can you audit one score	No	Partly	Full per-task breakdown
Work left open on purpose	Invisible	Reads as activity	Excluded from quality until closed
Task claimed done but not shipped	Invisible	Invisible	Verification runs first
Effort per cycle	Days of manager time	Dashboard upkeep	A query

What the score is not allowed to do

It does not fire anyone, reassign work, or message a person. The evaluation agent scores and reports, nothing else. Low scores and business-critical failures surface as flags to leadership and to the chief-of-staff agent running our operations layer, and a human makes the call.

We kept that boundary on purpose. A system that acts on a performance score automatically is a system nobody tells the truth to, and the whole method depends on the underlying task records staying honest.

What we got wrong

Version one scored activity and read beautifully. Everyone improved and nothing downstream changed, because a closed task and a useful closed task scored the same until the impact multiplier existed.

The second mistake was trusting closing notes. Work got marked done with a comment about an unrelated topic, sometimes from a copied template. We added a plausibility check comparing the completion criteria against the closing evidence, and the documentation says plainly what it is: a shallow overlap check, not comprehension. A well-worded answer that misses the point still gets through, which is why the supervisor path stays in place.

What running this costs

The build is scoped: task records, verification conditions, the scoring rubric, and the weekly report. That is an AI Workforce Sprint at $9,500, delivered in 4 to 6 weeks, and the code, the database and the accounts are yours at handover. If you do not yet know how many tools and automations your teams are running, the AI Operations Audit at $1,900 counts them first.

Keeping the rubric honest afterward is the recurring part. Weights go stale, new roles appear, agents get added, and nobody owns any of it by default. Our fractional AI officer engagement covers that at $2,900 a month, cancel monthly, with one build shipped each month rather than a retainer that buys meetings.

The fleet this report scores, with the run and failure counts behind it, is on our agent fleet proof page. The cockpit the weekly numbers are queried from is described in ai operations dashboard.

If you want this running on your own work records, send us how your team tracks tasks today, what a completed task has to show before it counts as done, and who signs off on a deadline change. Start async at /start. You get a written reply within one business day, and there is no call to book.