For each one: the result, the before and after, and how the steps divide between rules, AI, and people.
Across these four builds, 34 of 58 steps run on rules. AI reasoning handles 14. One uses machine learning. 9 stay with a person. Shares are by step count, not by time or cost.
| Build | Steps | Rules | AI reasoning | Machine learning | A person |
|---|---|---|---|---|---|
| 01 Sales call to costed proposal | 13 | 6 (46%) | 4 (31%) | 1 (8%) | 2 (15%) |
| 02 Candidate search and shortlisting | 16 | 8 (50%) | 3 (19%) | 0 (0%) | 5 (31%) |
| 03 Published analysis from a private dataset | 13 | 9 (69%) | 3 (23%) | 0 (0%) | 1 (8%) |
| 04 Structuring public records in an opaque credit market | 16 | 11 (69%) | 4 (25%) | 0 (0%) | 1 (6%) |
| All four | 58 | 34 (59%) | 14 (24%) | 1 (2%) | 9 (15%) |
13 steps: 6 rules, 4 AI, 1 machine learning, 2 human
Someone went back through notes or the recording, wrote up what the prospect needed, priced ideas from memory, and drafted the follow-up. Quality depended on who took the notes.
A recorded call goes in. Pain points, costed proposal ideas, and a follow-up draft come out, ready for a person to review and send.
Why you can trust it: every proposed idea carries a word-for-word quote from the call, or it is left out. The draft goes to the team, never straight to the prospect.
| 1 Select | It runs after every sales call, with one owner and the same three outputs each time. |
| 2 Diagnose | The constraint was the write-up: turning an hour of conversation into evidence someone can act on. |
| 3 Design | Rules for intake, signature checks, transcript parsing, and price tiers. AI reads the call in four passes. A person sends. |
| 4 Prove | The bar: every proposed idea carries a word-for-word quote from the call, or it is left out. Every score states its reason. |
| 5 Pilot | The draft goes to the team, never straight to the prospect. A person refines it and presses send. |
| 6 Decide | Each call is scored on five signals, so follow-up effort goes where the intent is. |
16 steps: 8 rules, 3 AI, 5 human
Resumes sat on two boards and in a folder of documents. Answering a brief meant opening them one by one and building the shortlist by hand.
Resumes become structured profiles, a brief becomes a ranked shortlist, and the client responds on one link.
Why you can trust it: if the model fails, the rule-based ranking stands. The recruiter picks the shortlist and sends it; AI never contacts the client.
| 1 Select | The same search repeats for every client brief, and the data already sat in one system. |
| 2 Diagnose | The constraint was reading, not searching. The facts were locked inside documents. |
| 3 Design | Rules for sync, duplicates, query parsing, and first-pass scoring. AI structures each resume once and re-ranks the top 20. The recruiter picks. |
| 4 Prove | If the model fails, the rule ranking stands. The stronger model runs once per resume, the cheaper one on each search. |
| 5 Pilot | The recruiter builds the shortlist and sends it. AI drafts the email and never contacts the client. |
| 6 Decide | Clients request introductions on the shared link, so each shortlist has a visible result. |
13 steps: 9 rules, 3 AI, 1 human
Each market note was written by hand: pull the numbers, build the charts, write the story, check the figures again. Most findings in the data were never published.
The system picks the story, computes the numbers, drafts the narrative, checks every claim, and emails a person to approve.
Why you can trust it: every figure in the draft is recomputed from the data and checked again by a separate review. Nothing is published without a person’s click.
| 1 Select | A repeatable output on a fixed schedule, from data already structured and owned. |
| 2 Diagnose | The constraint was trust, not writing speed. A published figure that is wrong costs more than no report. |
| 3 Design | Rules pick the topic, compute every figure, build the charts, and render the page. AI writes the narrative and runs a pre-publication review. A person approves. |
| 4 Prove | The draft’s figures are recomputed from the data and checked again by a separate review. A blocking finding stops the run. |
| 5 Pilot | The draft stays hidden. The approver gets an email with the draft and one button. No click, no publication. |
| 6 Decide | A story is not repeated within three weeks, so each report has to say something new. |
16 steps: 11 rules, 4 AI, 1 human
130K+ loans and 1.4 trillion dollars of commercial real estate debt extracted from public filings, matched across sources, and confidence-scored.
The information was public but unusable: scattered across thousands of documents in different formats, with borrowers behind layers of entities. Analysts pieced deals together one at a time.
A pipeline collects the records, extracts the loans, matches them across sources, and resolves who the borrower really is.
Why you can trust it: a record is promoted only when a second source corroborates it. Several agents gather evidence, and none may score its own confidence. Uncertain cases go to a person.
| 1 Select | High value and a hard problem: the records are public, but nobody had structured them. |
| 2 Diagnose | The constraint was trust in each link, not extraction. A wrong match is worse than a missing one. |
| 3 Design | Rules parse, normalise, match within set tolerances, and decide promotion. AI reads scanned pages and proposes links. Several agents gather evidence, and none may score its own confidence. |
| 4 Prove | A record is promoted only when every blocker is clear, including a second corroborating source. |
| 5 Pilot | Uncertain cases go to a review queue for a person to adjudicate. |
| 6 Decide | Every record carries one of four confidence tags, so an analyst knows what is confirmed and what is inferred. |
Half an hour with Andrew. If AI is not the answer, you will hear that too.
