1. What a Diagnostic Actually Is
A diagnostic is a paid, fixed-scope engagement that answers one question: what should you build, buy, fix, or leave alone — and in what order?
It is not a discovery call. A discovery call is a sales activity, correctly and openly, and it produces a proposal. A diagnostic produces an architecture, a sequenced plan, and a cost model that stands on its own and is usable with any vendor, including the ones who did not write it.
The distinction is structural rather than semantic. A discovery call is free because its cost is recovered in the engagement that follows, which means it must produce an engagement to have been worth doing. A diagnostic is paid, which means it can conclude "do not build this" without anyone losing money on the outcome.
That difference determines what recommendations are available. Everything else follows from it.
2. The Independence Test
Ask one question of anyone offering to scope your project: is this engagement priced and delivered separately from whatever you would build afterwards?
If the answer is no — if it is free, or credited against a future build, or contingent on proceeding — then the recommendation "you should not build this" is expensive for the person making it. They may make it anyway. Many good people do. But you are relying on their character rather than the structure, and structure is more reliable than character under commercial pressure.
This is not an accusation of bad faith. It is an observation about incentives. The same firm, with the same engineers, will produce a different recommendation depending on how the engagement is priced — not because anyone is being dishonest, but because ambiguous evidence resolves in the direction the incentives point. That is how humans work, and designing around it is cheaper than hoping.
A well-structured diagnostic makes "buy the off-the-shelf tool," "fix your process first," and "do nothing this year" available as conclusions. If those outcomes would be commercially painful for your advisor, discount their build recommendation accordingly.
3. What It Should Cover
A diagnostic that only looks at code is missing most of the problem. The scope that produces useful answers:
The operation as it actually runs. Not the documented process — the real one, including the spreadsheets, the workarounds, and the person everyone asks when something unusual happens. This is where the actual requirements live.
The systems and their seams. What exists, what it costs, what talks to what, and where humans are acting as the integration layer. Every manual handoff is a candidate and should be costed in hours per week.
The data model. Where the authoritative version of each entity lives, where it disagrees with itself, and what it would take to establish a single source of truth. Almost every ambitious project is gated on this and almost none of them budget for it.
The codebase, if one exists. Architecture, dependency health, test coverage, security posture, and an honest read on what is worth keeping. Rewrite-versus-refactor is a finding, not an assumption.
Security and access. Auth model, secrets handling, permission scopes, data exposure. Findings here are often the most urgent thing in the report and rarely what the client engaged you for.
Constraints that are not technical. Budget, timeline, internal capability, appetite for change, and who has to be convinced. A technically optimal plan that the organisation cannot execute is not a plan.
4. The Deliverables That Matter
Four things, and you should be able to hand all of them to a third party who has never spoken to the author:
A findings report. What is working, what is not, what it is costing — prioritised by impact rather than by how interesting the problem is. Ordered so that reading only the first page is still useful.
A target architecture. How the systems should fit together, what is bought, what is built, and where the boundaries sit. Specific enough to quote against.
A sequenced roadmap. What happens first, what it unblocks, and why that order. With honest cost ranges — ranges, not a single number implying precision nobody has at this stage.
An explicit list of what not to do. The features considered and rejected, and the reasoning. This is the most-skipped deliverable and frequently the most valuable one, because it prevents the same debate recurring quarterly.
If the output is a slide deck that ends in a proposal, you received a pitch. The test is portability: could you take this document to a different vendor and get a comparable quote from it? If not, it was not a diagnostic.
5. Red Flags
A quote before an investigation. A fixed price for a build, given after one call, is a number chosen for the sale rather than derived from the work. It will be revised, and the revision will not be downward.
No access requested. A diagnostic that never asks to see the codebase, the data, or the actual workflow is producing generic advice with your company name on it.
Every recommendation is a build. Sometimes true, statistically improbable as a pattern. If an advisor has never recommended buying, they are not evaluating.
Timeline certainty that exceeds the evidence. "Twelve weeks" stated with confidence before anyone has looked at the data model is a guess presented as a plan.
No named engineer. If you cannot find out who will actually do the work, you are buying a brand and receiving whoever was available.
A diagnostic typically runs one to three weeks and costs a small fraction of the build it precedes. Against the cost of building the wrong system — which is the entire build, plus the delay, plus the rebuild — it is the cheapest insurance available in software procurement. And the version of it worth buying is the one that can tell you to stop.
Run the readiness assessment
Eight questions on your data, process, and tooling. Three minutes, no system access, no sales call.