Your AI works. Nobody owns it.

I collect the AI rollouts that quietly stopped working, figure out why, and I am convening a closed room for the people who lived through them. Then I help teams take ownership of the systems they were handed.

Talk to me about a workflow Bring a case to the cohort

The quiet failure

Your rollout went great. Then it didn't.

The system launched. Cycle time dropped, the business case held, the dashboard went green and stayed green. Everybody moved on to the next thing.

A year later the workflow no longer matches how the work actually gets done, and nobody left in the building can tell you why it was built that way.

It is the gap between the system runs and somebody still understands it.

Almost every AI programme measures the first. Almost none measure the second, which is exactly why this failure is so quiet. So one of two things happens. The workflow freezes exactly as delivered and drifts out of true, or people route around it while usage stays healthy because nobody is looking at who.

I could not find anyone publishing these cases honestly, so I started collecting them.

What's here

Ownership over adoption.

I collect the cases nobody publishes, test them against a working framework, and bring the people who lived through them into the same room.

Research notes

The question, what the research says, my take, the implications, and where it breaks. Sources come with the discount on each one, because what a citation supports is usually narrower than the use it gets put to.

Read what I have published

A cohort that tells the truth

Ten to twelve practitioners, one hour a month, closed room, Chatham House rule. The format: somebody brings a case and the rest of the room takes it apart. No pitch decks, no vendor demos, no highlight reels. The price of entry is one de-identified case. The first cohort is forming now.

Bring a case

Consulting

Five gates that take a workflow from intent to something your team can change without calling anyone. Scoped so the end is defined before the start, and measured on whether ownership actually moved.

See how the work runs

Writing

Published so far.

Six pieces, all on LinkedIn. The research notes format described above is where this is going next. These are the argument as it stands.

Working together

Five gates, and the last one decides when I leave.

Most engagements end on a calendar. This one ends on evidence. The gates run in order, and each one has to clear before the next opens.

Gate 1

Destination

One page, vision only, before anybody looks at the current workflow. Inspect the workflow first and you anchor to it, then spend the budget automating what already exists instead of asking whether it should.

Gate 2

Reality

The workflow as run, not as documented. Recorded rather than interviewed, because people describe their idealised process when asked and reveal the actual one when observed. Out of it: the real decision points, who actually holds authority at each, and genuine judgment separated from inherited habit.

Gate 3

Mesh

Lay the as-run map against the destination and sort every step into serves the vision, fights it, or is irrelevant. A step whose rationale nobody can reconstruct is a deletion candidate before it is an automation candidate. This is the gate that saves the money, and the one most frameworks skip by jumping straight from current state to tooling.

Gate 4

Leverage

Only now do tools enter, and they enter governed: the system prepares, the human decides. What you get is a working proof of concept on your own data in about four hours, not a roadmap.

Gate 5

Ownership

Adoption is a usage metric owned by whoever sold you the system. Ownership is a capability held by your people. Gates 1 to 4 govern whether the project proceeds. Gate 5 is the only one whose criteria govern whether I leave, which is what stops an engagement quietly becoming a dependency.

It is measured on usage shape, not throughput: how often your team changes the thing without asking, whether they apply it to a problem I did not build for, whether they can say what it does badly, whether it survives their busiest week, and whether they consciously route around it where it should not be trusted. That last one is the measure nobody else takes. A team at full utilisation has not found the boundary, it has stopped looking for it.

The withdrawal date is announced at the start, because fading is scheduled or it does not happen. If that sounds like the kind of engagement you want, the first conversation costs nothing and is usually about half an hour.

Start a conversation

Who I am

Adam Helbig

I started in learning systems, building intelligent tutoring software whose entire design goal was to disappear. You model what the learner knows, you support them exactly where the model says they need it, and you withdraw the support as they improve. If the system is still helping at the end, it failed. The withdrawal is the product.

Then I spent several years putting AI into enterprise workflows across legal, compliance, finance and HR, and it took me longer than it should have to notice the obvious thing. This scaffold never fades. The AI does the preparation permanently and the human holds the judgment permanently, so the job changed shape and almost nobody wrote that part down.

That gap is what the five gates are built to close, and it is why the M.Ed is load-bearing here rather than decorative. Everything on this site is the working out of that one observation.

The Festivus

The session where people tell you how your system let them down.

Named after the airing of grievances, and only about half as a joke. "Do you have any feedback?" produces nothing, because it asks people to volunteer a complaint against social pressure. "Tell me how this disappointed you" inverts the burden. The complaint becomes the expected answer and silence becomes the deviation. It runs at handover, then thirty days, then ninety.

  • It goes on the calendar.Feedback that is not scheduled does not happen. "Always open to feedback" is what everyone says and nobody delivers.
  • Nothing defensive gets said.Notes and clarifying questions only. One explanation of why a complaint is unfair ends the practice permanently.
  • What changed gets published.A grievance with no visible consequence teaches people to stop bringing them.
  • I go first.Name something you got wrong before anyone else has to. It sets the floor for the room.
"It gives me an answer. I have no idea whether it is the right one."
"We stopped using it in April. Nobody told the dashboard."
"I spend longer checking it than I used to spend just doing it."
"Nobody can explain why it works this way. The person who knew left."
Illustrative. Written to show the shape of a grievance, not taken from a session. Real ones replace these once the cohort produces them.

Open to connecting

Write to me. I answer everything.

I am genuinely open to a conversation, and not only the kind with a budget attached. The argument on this page is built from method and from what I have run inside an enterprise, not from a case library. The cases are what I am collecting, and the best of them will come from people who wrote with no agenda.

If you are somewhere in the middle of one of these and it has not gone the way the business case said it would, that is exactly the conversation I want to have.

Worth writing about:

  • A workflow that stopped holding, and you are not sure why
  • Scoping an engagement, or just pressure-testing whether you need one
  • A case for the cohort, de-identified
  • Disagreeing with something I published
  • Speaking, podcasts, or a working session with your team

Also on LinkedIn.