About a third of the diagnostic engagements we run recommend AI for the wrong workflow — and conventional software, process redesign, or hiring for the workflow the client asked us about. Three patterns where AI is the expensive way to solve the problem.
About two-thirds of the diagnostic engagements we ship recommend AI work somewhere in the operation. The other third don't — at least, not for the workflow the client originally called us about. The other third get a recommendation that some combination of conventional software, process redesign, or hiring would solve their actual problem at a fraction of the cost.
This post is about those engagements. Three patterns recur. Each looks like an AI problem on the surface and resolves to something else under the diagnostic.
The most expensive AI engagements we've seen called off mid-diagnostic were attempts to AI-ify a workflow that wasn't broken. The team had heard about AI from a peer, a vendor pitch, or an article. They wanted to do AI somewhere. The workflow they picked was the one that was easy to point at — usually customer support intake, sometimes marketing copywriting, occasionally finance reconciliation.
The mistake is that easy to point at and high-leverage to improve are different attributes. A workflow that's running smoothly at its current cost and throughput, with no major bottlenecks or quality complaints, is not the place to start. The improvement headroom is small. The disruption cost is high. The team owning the workflow doesn't want the change. The AI implementation lands as a solution in search of a problem, and the problem it eventually finds is the team's morale.
The diagnostic question that catches this: what would happen if we did nothing about this workflow for the next twelve months. If the answer is nothing important, the workflow probably isn't the AI investment.
The right move, when the workflow is fine: don't build. Recommend that the AI budget go to a different workflow, or to the operational instrumentation that surfaces which workflows actually have headroom.
A second pattern: the workflow's quality issues trace to specific human decisions — a manager who's not delegating, a process where two teams disagree on ownership, a hiring gap that hasn't been filled. The client believes AI will route around the people problem. AI won't.
A specific example we walked recently: a 200-person services firm wanted an AI agent to triage incoming client requests, because their current triage process was inconsistent. The diagnostic surfaced the actual cause — the firm had hired two intake coordinators who had been given conflicting verbal direction by two different partners about how to prioritize work. The inconsistency was structural, not technological. Adding an AI to the same triage workflow would have produced inconsistent AI outputs trained on inconsistent human inputs.
The right move: don't build AI. Fix the decision rights. Once the partners aligned on triage priority and the coordinators had a written rubric, the inconsistency resolved without any technology investment.
The diagnostic question that catches this: who currently decides this, and what would they say the decision criteria are. If two people give two different answers, AI is the wrong tool. Get them to one answer, then revisit.
The third pattern is the one that turns into the most expensive false start when missed. The client wants AI to make a decision — score leads, route tickets, predict churn, classify documents. The model would work fine, if the data were clean. The data isn't clean.
We've watched companies spend six-figure budgets training models on data that the company itself wouldn't trust to make the same decision manually. The model produces outputs that are consistently wrong, which is sometimes worse than randomly wrong — the consistency creates the illusion of a working system. The leadership team, looking at the model's confident output, believes the decision quality has improved. The downstream teams, who can see the misclassifications, lose trust in the system and either work around it or work in parallel.
The fix is upstream. Before the model: data quality, data dictionary, data ownership, data hygiene. Boring infrastructure that pays back when the model arrives. Without it, the model is performing on a foundation that wasn't built to hold it.
The diagnostic question that catches this: if a person were to make this same decision manually, against the same data, what would their accuracy be. If the answer is not great, the data isn't reliable, the problem isn't a model problem. It's a data problem. Fix that first.
The Floburn diagnostic is built to surface all three patterns before any build begins. The deliverable distinguishes AI is the right tool here from AI is not the right tool here in writing. We bill the diagnostic either way. Some clients are surprised that we'd write a recommendation that doesn't include a build; the surprise is its own data about how most AI consulting works.
The shorter version: the question what AI work should we do is the wrong starting question for most companies. The right starting question is which of our workflows have the most operational headroom right now, and what's actually constraining each one. AI is one of several answers. Sometimes it's the best answer. Often, in practice, it isn't.
If you'd like the discovery call, bring the workflow you'd been considering for AI work. We'll walk the three diagnostic questions against it. The conversation is free; the answers it produces are the work we'd otherwise charge for.