Case study · Working prototype · July 2026

HR Mission Control keeps a compliance update attached to the work it creates.

I built a private agentic AI prototype that checks a defined set of public sources, prepares a review packet, and keeps the source, reviewer, owner, action, and closeout in one case. The person still makes the call.

Status
Working prototype
Built by
Mike Winkler
Environment
OpenClaw
Test data
Public and made-up
Decision authority
Qualified person

What is HR Mission Control?

HR Mission Control is the prototype I use to test agentic AI in HR compliance work. Its job is narrow: notice a possible update, prepare the evidence for review, route the work to a named person, and keep the approved action from disappearing after the first meeting. It does not interpret the law or decide company policy.

The problem

Finding the update was not the whole job.

A compliance update can land in an inbox and still go nowhere. Someone has to check the source, work out which facts are missing, decide who should review it, assign the next step, and confirm the change was actually made.

I built HR Mission Control to test whether an agent could carry that administrative thread without pretending to own the HR or legal judgment. The claim is modest: the source and the work stay connected long enough for a person to finish the job.

HR Mission ControlExample review path
  1. 01 · Input

    Approved source

    A policy or compliance update

  2. 02 · Agent work

    Review packet

    Source, summary, missing facts, suggested owner

  3. 03 · Human gate

    HR reviews

    Interpret, question, change, approve, or stop

    Agent stops here
  4. 04 · Follow-through

    Named owner

    Action, due date, status, and closeout evidence

64/64
repeatable checks passed
10/10
known items found in a small source test
240
local requests in five minutes without a failure
A simplified workflow Mike Winkler uses to test agentic AI in HR. The system prepares and routes the work; a qualified person owns the decision. The dated local results are not a claim of production readiness or complete legal coverage.

The artifact

The review packet shows its work.

A polished summary is easy to over-trust. The review packet separates what the source says from what still needs interpretation, then makes the handoff visible.

Example review packetHuman review required
Source
URL, publisher, retrieved date, and the version checked
What changed
A short factual summary tied back to the source
Still missing
Jurisdiction, workforce impact, conflicting guidance, or another open fact
Reviewer
HR, Legal, Security, Privacy, or another named owner can change, reject, or stop it
Approved next step
Named owner, due date, and only the action the reviewer approved
Closeout
Status, evidence, final decision, and the person who made it

The useful failure

A source changed its title. The issue did not become new.

One test source changed its page title. A naive workflow could have opened a second case and left the earlier review, owner, and notes behind. That is tidy software and bad operations.

I changed the case identity so it could survive a title change. The lesson was small but important: the workflow has to recognize the continuing piece of work, not merely compare strings on a page.

Dated local checks

The test record, without the victory lap.

These checks tell me the prototype behaved as expected in a limited local test. They do not prove production reliability, complete legal coverage, or client outcomes.

July 15, 2026

64/64

Repeatable checks passed

The same expected behaviors held across the local test suite.

July 13, 2026

10/10

Known items found

The workflow found every item in a small source set I had already checked by hand.

July 10, 2026

240

Local requests completed

The prototype completed 240 requests without a failure during a five-minute test.

The public trail

The evidence does not stop on this page.

I publish the non-sensitive workflow notes so the design can be inspected. An AI for HR Community session also described the OpenClaw work independently. Neither link turns a prototype into production software. They do make the work easier to verify.

The boundary

What the prototype is not allowed to decide.

It does not provide legal advice, promise that every relevant source will be found, interpret a policy for the company, or make a decision about an employee. It can prepare and route the work. A qualified person owns the meaning and the action.

  • No employee discipline, termination, pay, promotion, leave, or accommodation decisions
  • No production access to company or employee data in this prototype
  • No action past the review gate without explicit approval
  • No claim that a local test equals a secure company deployment

What this demonstrates

This is the kind of AI-in-HR work I can help with.

HR Mission Control is one working example of how I approach agentic AI in HR: start with a dull but real operational problem, limit the authority, build the smallest useful version, test an ugly edge case, and leave the evidence where another person can inspect it.

I bring 16 years in HR, including 10 as an HRBP, to the build. That matters because the hard part is rarely making the agent produce text. The hard part is deciding where the work should stop, who has to review it, and what a useful handoff contains.

Start with a conversation

Tell me what is getting in the way.

Bring the half-formed question, the process that keeps failing, or the idea you can’t quite explain yet. We’ll figure out what is worth solving. No deck required.

Book a conversation