The bottleneck on AI adoption is almost never the model. After two decades in customer research, I am convinced that most AI projects that fail in production were already failing in discovery — the team aimed at the wrong problem, measured success against the wrong baseline, or shipped a tool that solved a brief no end-user had asked for.
At igodemy's Research Lab we sequence every engagement in four phases, in this order. Each phase earns the right to start the next. We refuse to skip them.
- Understand the client — the buying organization as a system.
- Understand the internal users and end customers — the humans the AI will touch.
- Build a scoped proof of concept — narrow, time-boxed, designed to fail informatively.
- Implement in production — reliability, observability, change management, ownership transfer.
Phase 1 — Understand the client
The first two to four weeks go to the operating context. What does this business actually make money on, and what would have to change for that revenue line to grow twenty percent in twelve months? Where does the current process break — what decisions are humans making today, on what evidence, with what error rate? Who has the decision rights to deploy a change, who can quietly veto it, and what has the organization already tried?
The deliverable is a one-page problem statement and a target metric with a current baseline number, signed off by the executive sponsor. If we cannot quantify the baseline, we are not ready to start. Roughly one in four engagements stops here, because the right answer turns out to be a workflow change, a data-quality fix, or a hiring decision — not an AI deployment. We say so plainly.
Phase 2 — Understand the internal users and end customers
The second phase is where most engagements that go wrong have already gone wrong. The internal users are the humans whose work the AI will touch — the advisor whose lead routing will change, the analyst whose first draft will be machine-written, the support agent whose ticket will arrive pre-summarized. The end customers are everyone downstream of those people.
We use the full qualitative toolkit: interviews, contextual inquiries, think-aloud sessions, shadowing on the actual job. The objective is not to validate the client's hypothesis. It is to find where the hypothesis is wrong. We listen for the moments the human currently improvises (that is where the AI will fail), the trust contract with current tooling, the language the user actually speaks, and the edge cases handled silently. Where there is an end-customer dimension, we run segmentation and behavioral factor analysis on real data with proper method.
The deliverable is a research-grounded specification: what the AI must do, what it must absolutely not do, and the success criteria the users themselves agree to.
Phase 3 — Build the proof of concept
By Phase 3 we know what to build, for whom, and how to measure it. Only then do we build. Our POC rules are deliberately strict: scoped narrowly (one workflow, one user group, one outcome metric), time-boxed (four to six weeks, hard stop), and designed to fail informatively — with instrumentation in place for every failure mode we predict.
We measure against the Phase 1 baseline, not a vendor benchmark or a public leaderboard. And we evaluate the POC as the human-plus-AI system the production version will be — not the model in isolation. The right benchmark is "Was the user better at their job?", not "Did the model score higher?" We close the POC with a decision memo — production-ready, redesign-required, or stop — signed by the executive sponsor and a user-group representative. Both signatures matter.
Phase 4 — Implement in production
A POC is theatre. Production is craft. More than half of every engagement budget sits on this side of the line. Production means reliability engineering (latency budgets, fallback paths, regression tests), observability (every decision auditable), change management (training, SOP rewrites, incentive realignment), and ownership transfer — by month six an internal owner is trained and on the on-call rotation.
Measurement continues against the Phase 1 metric on a quarterly cadence, with a written post-mortem when the number moves and when it does not. The output is not just an AI feature. It is a way of working — a way of asking the next question, building the next thing, and retiring the parts of the old workflow the new one made obsolete.
This is not the fastest way to ship AI. It is the way that earns the right to ship more of it next quarter. The teams I have most respected over twenty years of research practice were never the ones with the most sophisticated models — they were the ones who refused to start the build before they had earned the right to. That is the discipline we are building here.