How we deliver AI
A method built to surface risk early
AI projects fail in predictable ways: unmeasured baselines, untested assumptions about data, late security reviews and launches without monitoring. Our method addresses each one at the point where it is cheapest to fix.
/ 01Phase 1
Discover
Decide on evidence whether and how to proceed.
Activities
- Interviews with the people who do the work
- Representative data and document samples under an agreed handling arrangement
- Baseline measurement: volume, time, errors, cost
- Feasibility spikes on the riskiest assumptions
- Risk classification and early involvement of security, privacy and legal
Outputs you keep
- Scorecard of success metrics agreed with sponsors
- Evaluation set, first version
- Architecture options with trade-offs
- Go or no-go recommendation and a scoped plan
/ 02Phase 2
Design & Build
Ship the narrowest valuable slice, measured.
Activities
- Threat model and data flow design
- Iterative sprints with working software at each demo
- Evaluation run on every change; results shared each sprint
- Approval and escalation interfaces designed with users
- Cost and latency budgets tracked from the start
Outputs you keep
- Production-grade code in your repositories
- Evaluation and regression suites
- Infrastructure as code
- Security and privacy documentation
/ 03Phase 3
Deploy
Release safely and prove value with real users.
Activities
- Shadow mode against live work where possible
- Pilot group rollout with feedback loops
- Monitoring, guardrails, alerting and rollback in place before release
- Training for users and approvers
Outputs you keep
- Pilot results against the agreed scorecard
- Operations runbook and incident playbook
- Rollout plan for wider release
/ 04Phase 4
Run
Keep it reliable, improve it, or hand it over cleanly.
Activities
- Quality, cost and incident reporting
- Model and dependency upgrades with regression testing
- Content and index maintenance
- Improvement backlog driven by production evidence
Outputs you keep
- Monthly service reports
- Release notes with evaluation results
- Complete handover pack if you take it in-house
Throughout
Practices that never switch off
- Evaluation before every release
- Human approval for consequential actions
- Least-privilege access to data and tools
- Documentation as a deliverable, not an afterthought
- Plain reporting of risks, assumptions and limits
Start with discovery
A short, fixed-scope discovery is the lowest-risk way to find out what AI can do for a specific workflow.