The build · Real
baxie
The AI system I run under my own name, guardrails and limits included.
See the build →Composite case · AI Program Diagnostic
Three stalled pilots.A plan in two weeks.
A B2B industrial distributor with a new COO and a board deadline. Three AI pilots in flight, none clearly working. Two weeks later the COO had a workflow map, a 90-day sequence with an owner, a baseline, and a build type on every line, and a kill list.
Illustrative composite, not a real client. The shape of the work is drawn from real patterns. The company is not a named client and the numbers are illustrative. For work that is real and checkable, see baxie.
01 · The situation
New COO.Inherited pilots.
30 hrs/wkRFQ re-entry by hand, baseline
6%RFQ error rate, baseline
23 of 38AP exception types that follow a rule
5 of 14Paid tools on the kill list
Day 10Readout, board-ready plan
This composite is a B2B industrial distributor with about 85 people. A new COO inherited three AI pilots, none clearly working, and a board that wanted an AI story before the next quarterly review.
The ops team had become the company's integration layer by default. Much of the AI and automation spend was unused or duplicated. Nobody had written down how long anything took before the pilots started, so nobody could say whether they worked.
02 · The problem
Where the timeactually went.
RFQ quotes retyped by hand
Four ops people spent about 30 hours a week re-entering supplier RFQ quotes into the ERP, from 35 different PDF and spreadsheet templates. A 6% error rate drove costly misquoted reorders.
AP exceptions nobody wrote down
Accounts payable exception routing followed 23 repeatable decision patterns that lived only in the AP lead's head. It ate about 9 hours of that person's week.
Ops as the integration layer
When the prior head of ops left, three undocumented vendor workflows broke for two weeks. The knowledge left with that person.
Tool sprawl
Fourteen paid AI and automation tools, the bulk of them unused or doing the same job as another one.
Pilots with no baseline
Three AI pilots in flight, none measured against a starting number, and a board review on the calendar.
03 · The two weeks
Map first.Then rank.
The diagnostic buys the plan. It does not build anything. That was the point here: the COO needed to know what to fund before anyone wrote more code.
- Day 1, a 60-minute kickoff. The COO, the ops lead, and the AP lead. We agreed on the access list and who I would interview.
- Days 2 to 9, the workflow map. ERP and inbox data for the RFQ and AP workflows, plus interviews with 8 people who do the work. Every handoff got an owner, hands-on time against elapsed time, and what breaks downstream when it slips.
- Sorting what is AI work. RFQ entry across 35 templates turned out to be deterministic: same fields every time. Plain extraction code handles it while the estimator reviews instead of retyping. The judgment stays human.
- Day 10, the readout. Sixty minutes with the COO and whoever had to sign off. The plan went to the board as a one-pager.
04 · The plan
What day 10handed back.
The top of the 90-day sequence and the kill list. Every line carries a named owner, a measured baseline, and a build type. The quick win is something the team could ship before any build started.
| # | Workflow | Owner | Baseline today | Build type |
|---|---|---|---|---|
| 01 | AP exception routing Quick win | AP lead | 9 hrs a week. 23 of 38 exception types follow a repeatable rule | Plain code for the 23. The other 15 stay human |
| 02 | Supplier RFQ quote entry | Ops lead | 30 hrs a week across 4 people, 35 templates, 6% error rate | Extraction code, estimator reviews every quote |
| 03 | Vendor workflow capture | COO | 3 workflows undocumented. The last departure broke them for two weeks | Human review. Written into the layer the team and the AI both read from |
| × | 5 overlapping AI and automation tools Kill | COO | Unused or duplicating another paid tool | Stop them |
Each of the three stalled pilots landed either on a line with a baseline or on the kill list with its reason: no baseline, blocked upstream, or not enough history for an agent to learn from.
05 · After the readout
A plan the COOcould defend.
The COO walked into the board review with a ranked plan instead of another pilot. Every line had a number to beat, so the build could be measured against where the company started.
The plan is the client's to take anywhere. If they build with me, the build is priced per workflow shipped and fixed in writing at the readout, and the $2,500 comes off it when they move within 30 days.
06 · Further reading
The othertwo examples.
Keep me on · Composite
Healthcare SaaS
The AI lead quit and three pilots drifted. Triage, the claims workflow, and a permanent owner hired.
Read the case study →07 · FAQ
Questions.
Our AI pilot works in a demo but we cannot get it into production. What is actually blocking it?
Almost always a handoff nobody wrote down: a data access approval, a system of record that will not expose an API, or a decision that needs an executive who has not been asked. The diagnostic maps the workflow before it recommends anything, because a pilot that stalls twice usually stalls on the same unmapped thing.
How do you turn stalled AI pilots into a 90-day plan?
Map the workflow each pilot touches, measure where it stands today, and name the owner of each piece. Rank the lines by the handoffs each one removes. Pilots with no baseline, a blocked upstream, or not enough history for an agent go on the kill list with the reason. Every line that survives gets a named owner, a measured baseline, and a build type.
How long does an AI Program Diagnostic take for a 100-person company?
Two weeks. Day 1 is a 60-minute kickoff, days 2 to 9 I map the work, and day 10 is the readout. Larger organizations do not take longer. They take more interviews inside the same two weeks.
How do you evaluate AI vendors when nobody internally can tell the difference?
By writing the evaluation criteria before looking at any vendor, tied to the workflow you are trying to change. Then a bake-off on your own data, with a defined pass bar. The failure mode is letting the demo set the criteria.
What does the diagnostic actually deliver?
The workflow map, with owners, handoffs, hands-on time against elapsed time, and what breaks downstream. The 90-day sequence, where every line carries a named owner, a measured baseline, and a build type. The kill list, with a reason for each item. And a 60-minute readout with whoever has to sign off. All of it is yours to keep updating after I leave.
Where does AI actually help in quoting?
AI can pull quantities and specs out of an RFQ, flag missing information before a quote goes out, and draft the quote documents. Someone who knows the work still sets the price and catches the exceptions. The code does the reading and the typing. Your estimator does the judging and reviews instead of retyping.
GAT.
Stalled pilots need a map. Start with the plan.
The AI Program Diagnostic. Credited in full if you build within 30 days. Next opening October 6.
Book 30 minutes with EdwinWant the full offer first? See what the plan includes.