01
The work before a model
- Grounded in the source
Map data ownership, quality and method versions. Tie AI output to retrievable evidence.
- Gated by an accountable person
Require a human decision before a regulated action. Record approvals and rejection reasons.
- Observable in operation
Keep audit trails, evaluations, queue health, error handling and release history visible.
- Engineered for operating cost
Budget infrastructure and model usage alongside delivery. Agree how ongoing costs will be measured.
- Measured against a baseline
Define the KPI and measurement period before implementation. Separate a target, a plan and a measured result.
02
Five phases, with review at every hand-off
- 01
Discovery on site
Map workflows, data sources, owners and reliability with the people doing the work. Agree the KPI baseline and acceptance criteria.
- 02
Design with a prototype
Analysts, engineers and approvers use a working prototype before the production scope is fixed. Record gaps and decisions.
- 03
Build in reviewable sprints
Review working software in two-week increments. Test data quality, permissions, evaluations and recovery alongside the features.
- 04
Integrate, accept and go live
Validate integrations, migrate agreed records, train administrators and complete user acceptance before the cut-over.
- 05
Hypercare and measured support
Plan four weeks of close support, then agree an annual maintenance arrangement and SLA. Compare results with the discovery baseline.
03
Why data and risk controls come first
In July 2024, Gartner forecast that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing data quality, risk controls, costs and unclear value. This is a historical forecast, not Fourtix performance data.
04
Start with the Prototype Sprint
Three weeks, a fixed scope and an agreed fee. Use a working prototype and a data map to make a go or no-go decision.




