he 7D framework is the operational core of Business-First AI™. Each stage produces a specific output. If an output is skipped, the next stage begins on assumptions — and assumptions compound across the remaining stages until they surface in production as failures.
**D1 — Discover**
The Discover stage produces one primary output: a signed-off problem statement.
The format is: "We need [AI capability] to [business outcome] so that [stakeholder can do X], measured by [metric]."
This is harder than it sounds. Most project briefs arrive as technology requests: "we want to use AI for customer service," "we need a recommendation engine," "can we automate this with ML?" These are not problem statements. A properly structured problem statement names the decision that needs to change, the outcome that depends on it, and the specific metric that will confirm success.
The Discover stage is where most AI projects fail silently. Not because the work is hard — but because the conversation is uncomfortable. Defining the problem precisely forces stakeholders to expose what they actually believe is wrong, which often contradicts the vendor pitch that triggered the initiative.
**D2 — Diagnose**
The Diagnose stage produces an AI Readiness Scorecard — 5 dimensions, each rated 0–5 — and a Prioritised Blocker List.
The five dimensions are: Data Quality, Process Maturity, Team Capability, Leadership Alignment, and Technology Infrastructure. A score below 3 on any dimension is a blocker that should be addressed before D3 begins. Proceeding past a known blocker is a budget risk, not a courage decision.
The most commonly underscored dimension is Leadership Alignment. Many teams have strong data, capable engineers, and well-documented processes — but leadership has never been asked to formally accept that AI outputs are probabilistic, not deterministic. This acceptance needs to happen before the project starts, not when the first model underperforms.
**D3 — Define**
The Define stage produces the PRD for AI Products — a requirements document that includes sections standard PRD templates do not have: data requirements (field-level completeness and freshness), model constraints (what the model is not allowed to do), confidence thresholds and fallback behaviour (what happens when the model is uncertain), an edge case register (at least 10 documented failure scenarios), and monitoring requirements (what gets measured post-launch).
This is the stage most closely aligned with CBAP® business analysis practice. The Define stage is where BA rigour — requirements elicitation, stakeholder alignment, traceability — applies most directly to AI product development.
**D4 — Design**
The Design stage produces an architecture decision record, UX wireframes or interaction design, integration specifications, and a phased implementation plan.
Key decisions made in D4: build vs. buy vs. fine-tune for the model layer; real-time vs. batch for inference; human-in-the-loop vs. fully automated for output delivery; shadow mode vs. A/B vs. full launch for rollout strategy.
Each of these decisions has downstream consequences for cost, compliance, and reliability. Making them in D4 — before engineering begins — is significantly cheaper than revisiting them in D5 or D6.
**D5 — Develop**
The Develop stage is the engineering execution stage. In Business-First AI™, the BA or strategist's role during Develop is requirements clarification, edge case validation, and acceptance testing — not model development. The model is built by engineering against the D3 acceptance criteria.
The most important BA contribution in D5 is catching scope drift: requirements that get quietly changed because they are technically inconvenient. The D3 document is the ground truth; any departure from it should be a formal change request with a business case.
**D6 — Deploy**
The Deploy stage ships the product with governance documentation, monitoring dashboards, rollback procedures, and human escalation paths active from day one. Not as an afterthought.
The Deploy checklist includes: data drift detection active, model performance thresholds configured with alerts, rollback trigger defined (if accuracy drops below X, an incident is raised), human review path tested for edge cases the model cannot handle, and stakeholder communication plan executed.
**D7 — Drive**
Drive closes the loop between the D1 problem statement and real-world outcome. It is the only stage that can formally confirm the project succeeded or failed — because success is defined as the D1 metric moving in the predicted direction, not as the model achieving its accuracy targets.
A 95% accurate model that does not change the business metric has failed D7, even if it passed every D5 acceptance test. Drive is how organisations learn whether their AI investments are working — and it is the stage most organisations skip.