Asia PacificNew-business exploration in one app, with a record of every decision and AI that drafts but never decides
The division's new-business team kept its candidate list, research and meeting notes in personal spreadsheets and email, so the reasons behind each decision were lost. We framed the exploration loop, designed one app to run it from long-list to lean canvas to field notes, and proved it in a Craft Sprint: AI drafts and proposes, and nothing changes until a person approves. It is now in trial and moving to a production build.
changes applied by AI without human approval
From first specification to a live demo for the division's executives
- Client
- Industrial products division of a Japanese manufacturer
- Users
- New-business exploration team
- Scope
- Craft Sprint prototype · Trial build · Production build
- Timeline
- Jul 2026 sprint → Sep 2026 trial → production build
- Team
- Consultants, full-stack engineers
Screens covering the loop, from long-list and voting to portfolio and decision log
Evidence levels (fact, counterpart opinion, hypothesis, hearsay); only facts can mark a hypothesis verified
AI functions: theme drafts, consumables scoring, Go/Pivot/Kill judgement, next-action advice, revenue hypotheses, field-note diffs
For AI to draft a new theme from a name and a one-line note
Good ideas were found, but the evidence and the reasons got lost
The division looks for new businesses in a fixed cycle. It builds a long-list of themes, agrees a short-list, tests each one on a lean canvas and goes out to meet universities, potential customers and partners. Each step lived in a different file.
What the team learned on visits rarely made it back into the hypotheses in a structured way. Six months later, nobody could say why a theme had been dropped.
Scattered information
Candidate lists, research notes and meeting minutes sat in personal spreadsheets and email, so no one had the full picture.
Inconvenient evidence buried
The more a team wanted a theme to work, the easier it was to skip negative customer feedback or competitor news.
Slow to start
Background research on each new candidate took days, so the long-list stayed short.
Frame the loop, prove it in a sprint, then harden it with real users
One rule shaped every design decision: AI never overwrites data. Everything it produces is a draft or a proposed diff. A person approves, edits or rejects each item, one by one, and there is no bulk-approve button. Evidence that is not a fact cannot promote a hypothesis to verified. The server enforces this when an item is decided and again when it is applied.
Craft Sprint
- Lean canvas tool turned into a local app with AI judgement
- Long-list, field notes and AI diff proposals added
- Tested on real meeting notes, then a live demo to executives
Trial build
- Database-backed app hosted in Japan with sign-in and admin-issued accounts
- Voting, short-list, portfolio, signals and decision log
- Architecture and security overview for the client's information-security review
Trial
- Users try it freely on sample data that can be reset at any time
- Per-user usage tracking to guide the next stage
- Feedback on screens and functions goes straight into the backlog
Production build
- Migration to Cloudflare Workers
- Real themes and minutes; the reset function switched off
- Security hardening as required, such as IP restriction and MFA
The exploration cycle now runs in one place, with a record of every decision
AI can draft and propose in seconds. What the team needed was a tool that keeps judgement with people and still shows them the evidence they would rather not see.
From trial to a production platform
After the trial, the app moves to a production build on Cloudflare Workers and the team starts entering real themes and minutes. The same rules carry over: AI drafts and proposes, people approve, and every decision is logged.
Real data
Switch off the sandbox reset, issue accounts in stages and start building up real themes and field notes.
Production platform
Rebuild the runtime on Cloudflare Workers, with security controls such as IP restriction and MFA to match the client's requirements.
From Go to project
Connect themes judged Go to the division's process for turning them into funded projects.
How we can help with a similar problem
Business Applications
Applications for processes that still run on spreadsheets and email, such as certification, warranty and approvals, built once and shared across the region.
ExploreAI Assistants
Generative AI built into one specific workflow, from an assistant that answers with its sources to an application where AI drafts and people decide.
ExploreRelated work
Next case →
Facing a similar challenge?
Tell us about it in a few lines. We will come back with a practical first step, usually within a week.