Why Your AI Agents Need an HR Department
For the last year and a half I have been running a small experiment on a Mac mini in Dhaka: an AI workforce that researches, writes, builds, and reports — while I sleep. It works. But the more agents I added, the more one question kept me up at night: which of these agents actually earn their keep?
It turns out almost nobody can answer that question. Gartner forecasts that over 40% of agentic AI projects will be cancelled by 2027 — not because the models are weak, but because the costs are opaque, the quality degrades silently, and nobody is governing the workforce. Companies know how to manage human teams. Nobody has written down how to manage a team of AI agents.
So I wrote it down.
The idea: give your AI agents an HR department
My new working paper, A Governance-Layer Architecture for Human-Governed AI Workforces, specifies something I call the Governance Layer. The core idea is embarrassingly simple: treat AI agents the way a well-run company treats employees.
- Every agent gets a personnel file — a ledger recording what it did, what it cost, and what value it produced, week by week.
- Budgets are earned, not given. High performers get more resources; declining agents get reviewed.
- Agents can ask, but never take. A structured request queue lets an agent argue for a bigger budget or a better model — with evidence. A human decides.
- An HR evaluator reviews the fleet monthly — promote, coach, replicate, or retire. The trends are computed deterministically; no AI grading its own homework.
- Reward hacking gets caught. If an agent's quality scores go up while its business value goes down, a detection rule flags it for review. That rule ships as open code with tests.
None of this requires a specific vendor, model, or framework. It is an architecture — schemas, rules, and boundaries — that any team can implement.
What the paper is (and is not)
I want to be precise, because AI research is drowning in overclaims. This is a working paper: a design-science specification with a documented validation roadmap. It does not claim the architecture beats alternatives in a randomized comparison — that study is pre-registered and scheduled, and I will publish the results whichever way they land. What it does claim is narrower and, I think, more useful: to my knowledge nobody has published an open, schema-level architecture that treats agent governance as organizational performance management — and now one exists, with every table, rule, and failure mode documented.
Read it, run it, break it
- The paper (~44 pages): tashfin.com/research/v7-paper
- The artifacts (schemas, the reward-hacking detector with 5/5 passing tests, synthetic ledger data — all MIT): github.com/tashfindelwar/ai-workforce-governance-layer
- Thirty-second demo:
python3 scripts/rule_h10.py data/synthetic_agent_ledger.csv— it flags exactly one misbehaving agent.
The paper was born from running a real system in production — every design decision in it has a scar behind it. If you are building agent systems, deploying them for clients, or just tired of AI bills you cannot explain, I would genuinely value your criticism. Replicate it, poke holes in it, tell me where it is wrong.
That is how this gets better. And it is also, quietly, the whole thesis of the paper: AI works best when a human stays in charge of it.