Worked example, synthetic data
What a board pack looks like when every number has to prove where it came from
The usual objection to putting AI in front of a sponsor is not that it cannot write. It is that nobody can tell which sentence is grounded and which one the model invented. So the pack gets re-checked by hand, and the hours you were meant to save come straight back.
This is the same monthly narrative with one rule applied: cite the source or cut the claim. Below is a worked example for a fictional PE-backed precision manufacturer.
Monthly operating review, Q2
Kestrel Machining Group
Synthetic example. Precision components, approximately 47M revenue, sponsor-owned since 2024. Not a real company.
Traced findings
- Gross margin fell 310 basis points quarter on quarter, from 34.2 percent to 31.1 percent.GL_Q2_2026.xlsx, tab Margin, row 44
- The principal driver is alloy input cost, up 18 percent against the Q1 average unit price.PO_ledger_Q2.csv, p.3
- Volume was not the cause. Units shipped rose 4 percent over the same period.Shipments_Q2.xlsx, row 118
- Receivables over 90 days reached 11.4 percent of the balance, against 6.8 percent at Q1 close.AR_aging_2026-06-30.xlsx, tab Summary
Contradiction, unresolved
Two sources disagree on the input-cost increase and both are cited rather than reconciled.
Purchase ledger shows plus 18 percentPO_ledger_Q2.csv, p.3
Operations deck states plus 12 percentOps_review_Q2.pptx, slide 9
The pack does not pick a winner. A 6 point spread on the single largest margin driver is a question for the operator, and it is put in front of the board as a question.
Cut, no traceable source
Headcount efficiency improved across the second quarter.
Present in last quarter's narrative, carried forward by habit. No file in the warehouse supports it, so it does not appear in the pack. It is shown here only to make the deletion visible.
The four rules underneath it
Every figure carries its source
Each number in the narrative is written with the file, tab and row it was read from. A sponsor who wants to check one does not have to ask anybody; the trail is on the page.
Disagreements are surfaced, not smoothed
Where two sources give different numbers for the same thing, both are quoted and the conflict is flagged unresolved. Nothing gets averaged into a comfortable middle to make the pack read cleanly.
Unverifiable claims are cut
A statement that cannot be traced to a source does not get softened or hedged. It is removed, and the removal is shown, so the pack is honest about what it does not know.
A human signs it off
The output is a draft for the person who owns the number, not an automated send. Review sits between generation and the board, and cost caps sit around the whole thing.
Built for the firm that already owns the data foundation
If you run the discovery, the data flow, the connections and the visualisation for PE-backed manufacturing, aerospace or distribution companies, your clients have already paid for the hard part. Trustworthy numbers exist. That is exactly the point at which an AI layer starts paying off instead of embarrassing everybody.
I work as the delivery layer on top of what you build, white label. You own the client relationship. I build the AI and it carries its sources. How the partnership works.
Frequently asked questions
Is this a real company?
No. Kestrel Machining Group is synthetic, invented for this page. The figures, files and contradictions are illustrative. No client data appears anywhere on this site, and no measured client outcome is claimed.
Does this replace our data foundation work?
No, and it does not work without it. This is the layer that sits on top of a clean, connected warehouse. If the underlying numbers are wrong, citing them faithfully just documents the wrong answer faster. Foundation first, AI second.
What happens when the AI cannot verify something?
It is dropped and the drop is shown, as in the cut section below. That is the whole discipline: cite the source or cut the claim. A pack that quietly hedges is more dangerous than one that admits a gap.
How would this start with our firm?
One real client use case, sized, with a fixed quote back. It folds into the discovery you already run rather than being a separate blueprint. From there: a working pilot on one real dataset, then a fixed-scope production build.
Send one real scenario
Describe one client situation, even loosely, and you get back what it would do, how it would be built, and what it would take. Everything in writing, nothing to sit through.