The Five-Phase Engagement Model structures AI consulting work as five sequential phases — Conversation, Strategy, Discovery, Build, and Handoff — each with a defined deliverable and a legitimate exit point. The client can stop after any phase with something complete in hand. In A2 Digital’s experience, about a third of clients stop after Strategy, because the strategy work surfaces a cheaper path than the one they walked in with.
By A2 Digital, an AI strategy and implementation firm. The Five-Phase Engagement Model is how we structure our own flat-fee engagements; the exit-after-every-phase rule is what keeps the model honest.
Why do AI engagements need a phase structure at all?
Because the alternative — one long scoped-at-the-start project — forces the client to commit to a build before anyone knows whether the build is the right answer. The most expensive AI mistake isn’t picking the wrong model. It’s picking the wrong problem and building competently against it for nine months.
A phase structure with real exits converts one large irreversible decision into five small reversible ones. Each phase produces something complete: a clear answer, a roadmap, a data foundation, a running system, an independent team. If the engagement stops early, the client keeps a finished artifact rather than a fraction of an unfinished one. And a structure where the firm can lose the later phases is the only structure in which the earlier phases can be honest — a firm that has already sold the build cannot be trusted to tell you the build is unnecessary.
What are the five phases?
| Phase | Typical duration | Deliverable | Legitimate exit looks like |
|---|---|---|---|
| 1. Conversation | 30 minutes | A straight answer on whether an engagement is warranted, and which kind | “You don’t need us. Here’s what to do first.” |
| 2. Strategy | ~6 weeks | A phased roadmap a board can read: build / buy / don’t-do per workflow, costs, risks, sequencing | The roadmap reveals a cheaper path — an off-the-shelf product, a process fix, or “not yet” |
| 3. Discovery | 4–8 weeks | A data foundation: documented workflows, surfaced tacit knowledge, lineage, labeling protocols | The foundation goes to an internal team or a later build — it is complete on its own |
| 4. Build | Scoped per engagement | A production system built with the client’s team in every working session, documentation co-written | The system ships and the client’s team already knows it, because they built it |
| 5. Handoff | ~2 weeks + defined on-call window | Operational transition: runbooks, the team running the system, the firm on call then gone | The client doesn’t need the firm. That is the deliverable, not a side effect. |
What happens in the Conversation phase?
The Conversation is a 30-minute working session whose only job is to determine whether the organization needs an engagement at all — and if so, which phase to enter at. It is free, because charging for it would create pressure to find a problem.
Some organizations arrive needing strategy; some arrive with a solid roadmap and need discovery or build; some arrive needing a leadership hire before any program work makes sense — in which case the right referral is to an executive search, not a consulting engagement. (A2’s sister firm, Talent Echo Advisory Group, runs retained search for Chief AI Officers and AI governance leaders; the two firms share an address and an operating philosophy, and the sequencing question — leader first or program first — is a conversation-phase question.) And some organizations arrive not needing anything yet. Saying so, in minute twenty, is the phase working as designed.
What happens in the Strategy phase?
Strategy is a fixed-length, flat-fee engagement — about six weeks — that answers what to build, why, in what order, and what not to build at all. The work is concrete: sessions with the people who own the workflows AI would touch, data lineage mapping, a pressure-tested business case, a risk register with regulatory mapping, and a phased roadmap with one-, three-, and twelve-month horizons that a board can read and a CFO can budget.
In A2’s experience, about a third of clients stop here — not because the engagement failed, but because it succeeded: the strategy work surfaced a cheaper path than the one they walked in with. An off-the-shelf product covers the need. A process change removes it. Or the honest sequencing answer is “not yet.” A strategy phase that never produces those answers isn’t analysis; it’s a sales document for the build.
What happens in the Discovery phase?
Discovery is where AI projects are actually won or lost, because most fail at the data layer rather than the model layer. The phase surfaces the knowledge the system will depend on — much of which is not in any database but in the heads of the people who have done the work for a decade: the rules they apply, the exceptions they recognize, the judgment no documentation has captured.
The method is craft, not survey: structured interviews with the people who do the work, direct workflow observation, asking why three times in a row, and documenting the answers in a structure AI systems can consume and the team can read back to confirm. The output is a data foundation — clean, labeled, lineaged, trusted — that is complete on its own. It can feed a build phase, an internal team, or a project two years from now. Skipping this phase doesn’t save its cost; it moves the cost downstream and multiplies it.
What happens in the Build phase?
Build is implementation under the co-builder discipline: the client’s team in every working session, sessions co-led, documentation co-written as the work happens, architecture on the stack the client already owns, and governance artifacts — bias evaluations, model cards, decision logs, audit trails — produced from the first sprint rather than bolted on before the audit. The full argument for why that discipline is the difference between a system and a dependency is in Why do AI consulting projects fail?
The honest trade-off: this is slower than a sealed delivery team in the first month, and dramatically faster across the lifetime of the system, because no one ever has to call the firm to ask how it works.
What happens in the Handoff phase?
Handoff is a structured two-week transition in which the client’s team takes over operations, with the firm on call for a defined post-engagement window — and then genuinely gone. Runbooks are finalized (not written; they were written all along), operational ownership moves, and the on-call window has an end date in the contract.
The test of the phase is blunt: after the window closes, does the client need the firm? If yes, the handoff failed — and so, in a meaningful sense, did the engagement, whatever the system does. Independence is the deliverable the other four phases were building toward.
Do all engagements go through all five phases?
No — and a model that forced them to would be a funnel, not a framework. Some engagements skip phases: an organization with a rigorous internal roadmap may enter directly at Discovery or Build. Some loop back: a build that surfaces an unmapped workflow returns to Discovery for it. And by design, many engagements end early — at the Conversation with “not yet,” or after Strategy with a roadmap pointing somewhere cheaper.
The sequence is real, though, which is why the phases are numbered: strategy before build, discovery before code, handoff planned from kickoff rather than improvised at the end. The model’s value is not that every client walks the whole path. It’s that at every point on the path, stopping is a legitimate outcome with a complete deliverable — which is the property that lets you trust the advice you get inside it.