Anitha Pillai · Head of Digital · Manufacturing · Mid-market · 180 employees
Scores reflect AI-coding readiness, not the commercial relationship — the deal is paid and healthy. Bumimax has bought roughly 16 AI-coding seats (RM63,000 at RM3,500 + 8% SST) but supplied no engineering telemetry: no tooling inventory, no cycle time, no PR throughput, no defect rate, no agent policy. That makes the two biggest gaps visibility and guardrails rather than enthusiasm; intent is proven by Anitha Pillai's paid commitment, yet Day 2 agentic content has nothing in the org to attach to. Highest-leverage 90-day move: run a 60-minute tooling and licence audit before Day 1, stand up a shared .cursorrules/CLAUDE.md baseline plus an AI PR-review and secrets-handling checklist inside the two workshop days, and instrument cycle time, PRs/dev/week and escaped defects in the same fortnight — so the cohort 2 business case is sold on evidence, not anecdote.
The only concrete data point is deal size: RM63,000 at RM3,500 + 8% SST implies roughly 16 licensed seats, and no Cursor, Claude Code or Copilot inventory, MCP server or IDE rules file was supplied. A 180-person manufacturing group likely runs locked-down Windows images, so Day 1 hands-on labs are the single highest execution risk in the engagement.
Nothing in the supplied context references AI-assisted PR review, prompt hygiene, agent permission scopes or secrets handling — the record contains only commercial fields (stage Won, invoice paid today, tag 'Listed company'). Day 2 assumes an existing code-review habit to bolt guardrails onto, and there is no evidence that habit exists yet, so Day 2 must build it from first principles rather than optimise it.
The single strongest signal is behavioural: a Head of Digital signed and the invoice was paid the same day, which is real executive sponsorship rather than exploratory interest. It is held at at_risk because the sponsor sits in Digital, not Engineering, and no engineering lead, staff engineer or platform owner appears anywhere in the record — so willingness to change developer workflows and blast-radius tolerance for autonomous agents remain unverified.
Zero throughput data was supplied — no cycle time, PRs per developer per week, time-to-merge or deploy frequency appears in the deal record, only commercial fields. That absence is itself the finding: the workshop cannot demonstrate a before/after gap, and a manufacturing IT org without instrumentation typically deploys on a batch cadence, so the baseline has to be captured in week one rather than assumed.
The deal value implies roughly 16 seats against a 180-employee company, so the paid cohort covers a small slice of the engineering population and no phasing plan for further cohorts is visible in the record. With one cohort of ~35 maximum and no named internal champion, adoption risk is concentration: the trained 16 become a silo rather than a standard.
No quality data was supplied at all — escaped defect rate, time-to-feature, rework and developer satisfaction are absent, and the only quality-adjacent signal is the 'Manufacturing' and 'Listed company' tags, which imply a real audit and disclosure bar. That bar is exactly what makes untracked AI-generated code dangerous: without defect attribution, a regression would surface in an audit before it surfaces in a dashboard.