Running programs with AI in the loop.
Field notes from inside live enterprise transformations — what AI actually changes about governance, reporting, and risk, and what still takes a human with their name on the outcome.
Where is your program quietly drifting?
Eight questions, scored across the five things AI actually changes about delivery — accountability, the record, early warning, capability, operations. Runs entirely in your browser; nothing is sent anywhere.
Run the diagnostic →AI-augmented program delivery
The whole series in one place, organised by the five problems AI actually creates on a live program: accountability, the honesty of the record, signal versus noise, the skills nobody is practising, and keeping the thing alive after launch.
Read the field guide →The handoff nobody verified
An assistant writes the transition brief in four minutes and it reads like certainty. It carries the record, not the program, and the incoming lead starts the job on a map nobody walked.
Read →The tool changed under you
The vendor updates the model behind your assistant, the same prompt returns a different answer, and nobody raises a change request. A tool that helps write the record is part of the configuration, and your trend lines just grew a step that has nothing to do with the program.
Read →The number came first
Estimating used to be an argument that ended in a number. AI hands over the number in ten seconds and the argument never happens, so nobody owns it, nobody found the unknowns, and the first figure in the room anchors everything after it.
Read →Nobody grades their own homework
The same assistant writes the code, the tests, and the results summary, and every check comes back green. That's one opinion repeated five times. A check only counts if it could have come out differently, and AI quietly erased the separation of duties that made checks mean something.
Read →We wrote that down last time.
Every program ends with a lessons-learned document, and the next program repeats most of it anyway. AI makes every lesson instantly retrievable. Retrieval was never the problem — nobody was assigned to change a decision because of what they found.
Read →The program shipped. Did anything change?
The closeout deck is sixty slides, assembled in an afternoon, and every number in it is true. It proves the program happened — not that it was worth doing. AI made evidence of activity almost free and left the cost of proving benefit exactly where it was.
Read →The friction was the filter
Change control worked partly because writing a change request cost somebody two days. AI dropped it to four minutes, and a board sized for five requests a month now sees thirty — every one of them articulate. The gate didn't fail. Its economics did.
Read →Nobody writes a bad proposal anymore
Every bid comes back clean, compliant, and quietly indistinguishable — because the same tools wrote all of them. The written response was a proxy for effort and bench strength, and that proxy has stopped working. Evaluate what a vendor can't generate.
Read →The plan nobody agreed to
AI drafted the schedule in ten minutes and it was better than the one your team would have built. Then three owners admit they never looked at their line. Planning's real product was never the document — it was a room full of people arguing until each of them committed out loud.
Read →One program, five realities
Your CFO, your sponsor, and your PMO lead each asked AI how the program is doing — and got three different answers from the same record. The status report was your common operating picture. Kill it without replacing that function and you get a market of private truths.
Read →The machine that cried wolf
You wired AI in to warn you sooner — now it warns you about everything, and the team has learned to scroll past. Healthcare has a name for this and a decade of scars. Detection is the cheap half; the triage discipline that keeps a human listening is the deliverable.
Read →The recorder in the room
Every meeting on your program quietly became a verbatim, searchable, permanently retrievable record — and no governance body approved that. A transcript tells you what was said. Minutes tell you what was decided. Only one of those is a governance artifact.
Read →Subcontracted judgment
You wrote an AI policy for your employees, then bought most of the program from an integrator whose delivery model changed under you. Governance stops at your badge line. Accountability doesn't — so stop asking vendors whether they use AI and start writing the standard into the contract.
Read →Chain of custody
Six months on, someone asks where a number came from. A human left a trail whether they meant to or not — an agent leaves the output and nothing else. Traceability is what separates a program that can defend its numbers from one that can only apologize.
Read →Judgment atrophy
AI takes the routine work first — and the routine work is where program judgment was built. Aviation ran this experiment already: pilots lost manual proficiency and were the last to know. If the reps are gone, the bench has to be trained on purpose.
Read →The confidence problem
A human who isn't sure sounds like it — and that hedge is information leaders have run on for decades. AI-drafted reports sound equally certain whether they read the record or guessed over a gap. Engineer the doubt back in, or lose your oldest early-warning system.
Read →Mind the seams
Programs don't break inside teams — they break at the seams between them: the handoff nobody owns, the dependency with two owners, the interface that drifted while both sides looked away. AI can map every dependency and still own none of them.
Read →Automation debt
Standing up an automation feels like deleting work. It isn't — you just converted visible manual work into invisible standing work that breaks silently and needs an owner. The build is the cheap part. The upkeep is the program.
Read →Pilot purgatory
Your AI pilot worked — that's the easy 10%. A demo runs on clean data, a willing user, and no consequences. Production runs on the real board, a skeptic, an integration nobody scoped, and a name on the output. That gap is where most AI programs quietly die.
Read →The rubber-stamp review
You kept a human in the loop, so you call the work supervised. But the better the AI gets, the less the reviewer actually reads — and a safeguard that has quietly stopped looking is the dangerous kind, because everyone still trusts it.
Read →The decision backlog
A program rarely dies because the work was too hard. It dies in the gap between when a decision was needed and when someone finally made it. You track the work to the hour — but nobody tracks the decisions, which is exactly where the schedule bleeds out.
Read →Shadow AI is already on your program
Your team already adopted AI — you just didn't approve it. A no-AI policy doesn't stop the usage; it removes your ability to see it. Bring it into the light, or govern nothing.
Read →Garbage in, gospel out
Point an agent at a dishonest Jira board and it won't tell you the board is lying — it will summarize the lie in executive-ready prose. Data hygiene isn't an admin chore anymore. It's the load-bearing wall under every AI capability you want.
Read →When the agent does the work, who signs?
AI drafts the report, flags the risk, moves the ticket. None of that changes who owns the outcome. You can delegate the labor — you cannot delegate the signature. The boundary that decides whether AI makes your program safer or just faster.
Read →The AI-augmented PMO: what actually changes
Gartner says AI could absorb 80% of project management tasks by 2030. That number is real — and the conclusion most people draw from it is wrong. The 20% that's left was always the entire job.
Read →Killing the status-report tax
Teams lose a person-week a month manufacturing reports nobody trusts. AI closes the gap — but only if you build it to tell the truth, not just to be fast. Automating a lie just makes it cheaper.
Read →AI won't save a failing program. It warns you sooner.
The expensive part of a collapse is the three months nobody admitted it. Early detection turned 45 days of warning into $4.7M saved on one program. The rescue is still human. The timing isn't.
Read →The delivery stack behind an AI-augmented program
Not a tool list — a division of labor. The model thinks, the automation moves, a human owns the boundary. Claude, Copilot, Power Automate, Make, n8n, and the governance layer everyone skips.
Read →One note a week, every week since May 2026.
Governance, reporting, program risk, and where the machine stops and judgment starts. Written from live enterprise programs. No pitch, no course, no funnel — just the writing.
Prefer a reader? Subscribe by RSS. Your address is used for the notes and nothing else.