Julia Broberg · governed AI and decision systems
AI can already run your planning.
When it gets one wrong, who has to answer for it?
Plenty of teams have wired an AI into something this year. Almost none of them can still tell you, six months on, what it did, who approved it, or how to undo it. Doing it so it holds is the hard part, and it is the one nobody is selling. I get the definitions to agree first, prove it on one real decision, and leave you a record of what happened and who said yes.
Here to check the engineering? See the work →A meeting you have already sat in
Somebody asks a question that sounds simple. How much do we actually spend with this supplier?
Three people answer. Procurement has one number. Finance has a different one. The site that receives the material has a third. Nobody is lying. Every number is correct inside the system it came out of.
It ends the way it always ends, with somebody agreeing to go and check. Next month it happens again. That is not a data problem. It is a definition problem, patched every month by one person who knows where the bodies are.
Now automate on top of it.
The disagreement does not go away. It gets faster, it stops being visible because nobody is patching it by hand any more, and it arrives sounding completely certain.
You do not get a faster company. You get the same wrong answer at machine speed, in more places, with the one person who used to catch it now out of the loop. That is worse than where you started, and it will fail you confidently.
It's the same wrong answer, but this time at machine speed.
Accountability, not capability, is the scarce thing now.
So where do you start
You already know which process it is.
Everybody has one. The thing that eats a week every month, or the approval that goes missing, or the number the sites stopped believing two years ago. You do not need an AI strategy to begin. You need two weeks and one honest look at the worst of them.
Find out what is actually broken
Two weeks
Which decisions are costing you, what data they lean on, where the definitions disagree, and what is worth automating and what is not. You get a plan for the next 90 days instead of a general worry about AI.
Then you choose: pilot or not. With me or with anybody else.
Run one decision in shadow
Six to eight weeks
One decision, one bounded scope, running alongside how you do it today so we can compare. A gate on anything consequential, and a record of every call it makes.
Then you choose: go or stop. If the numbers do not hold up, I will tell you to stop.
Hand it to your team
Until you do not need me
Widen what the pilot proved, a wave at a time. Monitoring, training, written procedures, and a handover your people can run without me in the room.
The point is that I leave.
What those two weeks are actually for
Same tools. Same company. Two sequences.
Nobody sets out to build the top row. It happens because the tool is available today and agreeing on definitions is slow, political work that nobody gets promoted for. That is exactly why it is the step that gets skipped, and exactly why skipping it is expensive.
Automate first, sort it out later
Agree, prove, then automate
Step one is the one everybody wants to skip. It is also the one that decides whether the other two hold, and it is why I do it with your team in the room rather than in a document I hand over at the end.
Two things you have to believe before any of this matters
That I have run the thing, and that I will not hand it to a machine and walk away.
Separate claims. Most people selling AI governance can only make the second one, because they have never carried a number. Open any line to see how it was checked.
One. What I have built and had to answer for
- Source
- Derived from the resume experience bullet's "ten-plus ERP systems" (docs/resume/resume.yaml) and the person.json bio's account of consolidating fragmented manufacturing and supply chain systems into one reporting framework; framed for the Home mockup's proof column per jbroberg-handoff-complete/00-MASTER-HANDOFF.md section 5 ("confirmed facts for the build"), 2026-07-22.
- Checked
- 2026-07-22
- Why it is here first
- It is the meeting at the top of this page, solved. Three answers to one question becomes one answer with an owner, and the definitions become the standard rather than the argument.
- Source
- Confirmed directly by Julia, 2026-07-22 (relayed through the rebuild orchestrator), superseding the 2026-07-18 resume-adjudicated figure of five analysts across four international sites. Also stated in jbroberg-handoff-complete/00-MASTER-HANDOFF.md section 5, "confirmed facts for the build": "Team is six analysts, working globally."
- Checked
- 2026-07-22
- Why it is here
- Standing a function up from nothing is a different skill from running one you inherited, and it is the one this work needs, because I am building something your team has to own after I leave.
- Source
- Same underlying figure as automation-savings-ratio (10:1, cleared 2026-07-20, resume-derived) - this entry exists only to carry the Home mockup's own claim wording (mapping the work / deleting the waste) verbatim, per the rebuild's rule that ported copy is final as written in the mockup. Same figure, two framings; automation-savings-ratio is left untouched and still resolves the resume/About/PDF surfaces.
- Checked
- 2026-07-22
- The distinction
- Automating waste makes waste permanent. Map the value stream with the operators first, remove the steps that should not exist, then automate what is left. The savings come from the removal.
- Why it matters to you
- It is also why the change survived after I moved on. People do not defend a process they helped delete.
- Source
- Aggregates lead-time-reduction (17 days to 1, cleared 2026-07-18) and backlog-dollar-reduction ($17M to $7M, cleared 2026-07-18) into the single figure the Home mockup's tile needs, the same aggregation pattern already used by the governed-systems-tests entry above (id governed-systems-tests aggregates jbos-tests + bridge-tests). Needed because a Home proof tile resolves exactly one metric_ref.
- Checked
- 2026-07-22
I have been the one who gets the call at five in the morning when the number is wrong. Most people selling this have not.
the one who gets the call at five in the morning when the number is wrong
Two. How I build so somebody can still answer for it
- Source
- Vendor Brain runs this way in production every day inside a NASDAQ-listed manufacturer, and the same gates run in my own operating system and memory layer.
- How it works
- The gate is in the architecture, not in a settings menu, so it cannot be switched off by somebody in a hurry on a Friday.
- Where it runs
- Vendor Brain runs this way in production every day inside a NASDAQ-listed manufacturer, and the same gates run in my own operating system and memory layer. See it running.
- Source
- Vendor Brain reconciles suppliers in production but writes nothing back on its own. Every merge a person confirms.
- Why this and not a rollback promise
- Recovering from a bad automated write is a worse position than never allowing one. Nothing writes without a dry run first, and the bridge to the system of record is read-only by construction.
- Where it runs
- Vendor Brain reconciles suppliers in production but writes nothing back on its own. Every merge a person confirms. See it.
- Source
- An append-only audit trail on every consequential action.
- Why it matters
- An agent nobody can explain is an agent nobody can defend. The first time it is wrong in front of a customer, that is the only thing that matters.
- Source
- Cemri et al., Why Do Multi-Agent LLM Systems Fail, NeurIPS 2025, across more than 1,600 execution traces from seven frameworks. arxiv.org/pdf/2503.13657
- Checked
- 2026-07-22
- How checked
- The blueprint is revision-stamped REV A, dated 2026-07-17. A commit-level artifact separating the spec from the build is still to be produced.
- What it means for you
- A better model will not fix a management problem. The design document is the control, so I publish it before anything gets built.
An agent with no evaluation suite is not a production system. It is a demo that has been left switched on.
The small step
Tell me the decision that keeps going wrong.
Plants, warehouses, payables, claims, month-end, onboarding, procurement. The words change and the shape does not. Systems that disagree, a definition nobody ever locked, one person holding it together by hand, and now something automated sitting on top of all of it.
Send it over
Four fields. No phone number, no budget question, no "how did you hear about us."
If this is not something I should take on, I will say so and point you in a different direction. Rather talk it through? Grab 15 minutes.