Source of truth: PROCESS.md · E5 internal · not for James
# Bulten UK: what AI was actually for Equals Five internal. For Josh. 16 Sep 2026. Not for James. ## The point of this document Using AI on this job was not "paste a prompt into ChatGPT and get a market pack." It was running structured research that pulls data a single prompt cannot get. That is where AI is genuinely strong on work like this. Where it is genuinely weak is what I want Josh to see clearly below. ## Where AI was strong: structured research The useful chain was: 1. SIC codes for the UK markets Bulten can sell into 2. Every company under those codes (full universe, not a famous-name sample) 3. Group that universe by market sector 4. Filings for those companies 5. TAM built from filings plus explicit assumptions 6. Link TAM / SIC to real company names 7. Scrape and deepen those named accounts 8. Job specs and case evidence: how fasteners are sold, who buys, deal sizes, time to market 9. Validate the market from that evidence 10. Only then run analysis and numbers on that baseline A single prompt skips most of that and invents the middle. Structured research gets you to named companies, filed size, and commercial evidence. That is the floor under any model. ## Where AI was weak (highlight this) ### 1. Forcing market analysis on bad or partial data I kept pushing the models to write market analysis before the research floor was finished. They will happily produce a confident brief on incomplete SIC coverage, missing filings, thin account lists, or half-built TAM. The prose looks fine. The numbers are not grounded. **Rule:** build the research from the ground up first. Do not ask for the analysis chapter until the universe, filings, TAM, and named-account evidence exist. If the data is partial, the honest output is a gap list, not a polished sector narrative. ### 2. Getting distracted from James's two goals James had already framed the job as two things: 1. **Opportunity model** 2. **Growth model** I used the AI to produce side briefs that wandered off those two. Useful-looking pages. Wrong spine. We got lost in the weeds (extra comparisons, packaging, peripheral narratives) while the two models that actually mattered were under-built or late. **Rule:** every session starts by naming which of the two models you are advancing. If the next prompt does not move opportunity or growth, park it. ## How to run the next one 1. Lock James's two goals on a single line: opportunity model + growth model. 2. Run the research chain to a baseline before any "write the market analysis" ask. 3. Use AI hardest on steps that need scale: SIC pull, company lists, filings harvest, account scrape, job/case mining. 4. Use AI lightest on prose until the baseline is real. 5. When something shiny appears (a third comparison page, a new brief shape), ask: does this serve opportunity or growth? If not, stop. ## Delivery hygiene (short) Spell **Bulten**. No model names in client pages. No em dashes. Client body and appendices stay separate. Mirror temporary previews to disk the same day. ## Live artefacts - Client pack: https://bulten-client-pack.vercel.app/ - This briefing: https://bulten-e5-process.articulate-ai.work/ - Appendices: https://bulten-appendices.articulate-ai.work/ - Writer gates: https://bulten-writer-gates.articulate-ai.work/ - ChatGPT exec (recovered): https://bulten-chatgpt-exec.articulate-ai.work/