When Your AI Problem Is Actually a Process Problem

TL;DR. We turned down four AI projects this quarter. Three of them were process problems wearing an AI costume. The fourth was a knowledge-capture problem. None of them needed an AI agent. If your quoting takes three days, your inventory is wrong, or your best estimator is retiring, the AI question is downstream of a more basic one: do you actually have a process, or do you have four people doing the same thing four different ways? Fix that first.

A VP of Operations at a $60M HVAC contractor called us last month asking for a quoting agent. Twenty minutes into the call we figured out the real problem. Three of his five senior estimators were using different markup formulas. Two of them weren't writing anything down. The job costs in his ERP didn't match the job costs in his quotes. Whatever we built him would have automated chaos, not solved it. We told him to spend the next quarter standardizing the markup logic and pulling the actual job-cost history out of his ERP. Then call us back.

That call wasn't unusual. We've turned down four AI builds in the last three months because the prospect didn't need AI. They needed to do operations work first.

The four AI problems that aren't AI problems

Most of the "we need AI" calls we take fall into one of four buckets. Three of them don't need AI at all. The fourth needs AI but is also a process problem that has to be solved first.

The standardization problem. Different people in the same role doing the same task different ways. Five estimators with five markup approaches. Three claims processors who classify the same edge cases differently. Two service managers who write up the same callback with different root-cause codes. An AI agent trained on this data learns five contradictory rules, then makes the variance worse because it now produces output at higher volume. The first job here is to pick the right way and get everyone doing it. That's a manager-with-a-spreadsheet job, not an AI job.

The integration problem. The data you need exists somewhere in the company, but it lives in three systems that don't talk to each other. Job costs in your accounting system, time tracking in a separate field tool, material costs in a spreadsheet your office manager maintains. You don't have an AI problem here, you have a plumbing problem. Until those systems share data, no AI agent can give you the answer you want, because the answer doesn't exist yet in any one place. The fix is integration work, usually a few weeks of pulling APIs together or building a middleware layer.

The documentation problem. Your senior people know how the work should be done. They just haven't written it down. The new hire takes a year to ramp because the institutional knowledge lives in three people's heads. An AI agent built on undocumented tribal knowledge produces confident-looking garbage, because the model has no source of truth to learn from. The job here is to get the knowledge out of the heads and into a place where it can be referenced, tested, and improved. See How to Capture Tribal Knowledge Before Key People Leave for what that actually looks like.

The workflow problem. Handoffs that drop work. Approvals that take six days because nobody knows who has to approve. Customers calling because they don't know what's happening with their job. An AI agent in this environment becomes another box in the handoff chain, often making the lag worse because it adds a system to check. The fix is workflow design: who does what, in what order, with what handoff to whom. Most mid-market operators have never actually mapped this. Doing so reveals four to six places where a human could be replaced by a button, before any AI is involved.

Why this matters more than vendors will tell you

The data on this is brutal. The RAND Corporation's 2024 study on why AI projects fail found that more than 80% of AI initiatives fail to deliver their intended value, roughly double the failure rate of non-AI IT projects. The top root cause, according to the 65 senior data scientists and engineers RAND interviewed, was business leadership misunderstanding the problem they were trying to solve.

McKinsey's State of AI 2025 survey reached a similar conclusion from a different direction. Eighty-eight percent of organizations now use AI in at least one function, but only about 6% qualify as high performers achieving meaningful financial impact. The biggest predictor of which companies got real EBIT out of AI? They redesigned end-to-end workflows before selecting the AI technology. McKinsey's framing in the report: AI is 20% algorithms and 80% organizational rewiring.

Translate that out of consulting language. The companies getting value from AI are the ones who fix the process first, then automate. The companies getting nothing are the ones who tried to skip the process work and let the AI compensate. It can't.

The five-minute diagnostic

When somebody on your team says "we need AI for X," ask three questions before you go look for a vendor.

1. If I gave this to your best person, could they do it correctly every time?

If the answer is no, you don't have an AI problem yet. You have a process problem. Until your best person knows the right answer, no AI model can be trained to produce it. The work is to define what right looks like, get it on paper, and pressure-test it on real cases. Then revisit the AI question.

2. Is the data the AI would need actually accessible in one place?

If the answer is no, you have an integration problem. The AI can't learn from data that doesn't exist in a queryable form. You can sometimes solve this with the AI build itself by including the integration work in scope, but it adds weeks and cost. The honest version is to do the integration first as its own project, prove the data is clean and accessible, then decide whether AI is the right layer on top. We've seen $50M operators spend $200K on AI that delivered nothing because the underlying data was a swamp.

3. Does the workflow this AI would live in actually exist?

If the workflow is "Sarah handles it however she sees fit," there's nothing for the AI to plug into. AI agents work best as a step in a defined process, not as a replacement for having a process. If you can't draw the workflow on a whiteboard in fifteen minutes, the AI agent will live in a vacuum and produce results nobody knows what to do with.

If you can answer yes to all three, you may have an actual AI problem, and we can have a useful conversation. If you can't, the next ninety days should not be about AI vendors. They should be about a manager with a spreadsheet, a half-built integration, or a documentation push.

When it actually is an AI problem

For the avoidance of doubt: there are real AI problems at mid-market operations companies. We build them every month. The pattern that genuinely needs AI looks like this:

• A high-volume, repeatable decision that a person with training can make correctly almost every time, but the volume is too high for the people you have.
• Data lives in one or two accessible systems that the AI can query.
• A defined workflow already exists, with clear handoffs in and out of the AI step.
• There's an obvious failure mode (the AI gets it wrong) and a defined human review path for that case.

When a $40M insurance administrator wants to summarize 5,000 incoming claims a week into a structured intake record, that's a real AI problem. The decision is repeatable, the data is in one system, the workflow already routes claims through an intake queue, and there's a human reviewer downstream who catches errors. Build the agent.

When a $55M distributor wants an AI agent to "tell them what's wrong with their inventory," that's not a real AI problem yet. They have eight people doing inventory differently across three warehouses. Fix that first.

What this means for your AI budget

The vendors quoting you $200K to $400K for AI projects need you to believe the problem is sophisticated. The actual mid-market AI projects we ship for four-week fixed prices look more like this: pick a workflow where you've already done the process work, scope it down to the smallest meaningful slice, ship it, and measure whether the metric moves. If yes, expand. If no, kill it and try the next slice.

If you haven't done the process work, no vendor quote saves you. The $400K vendor will deliver something polished that doesn't move the metric. The $50K vendor will deliver something cheaper that also doesn't move the metric. You don't have a vendor problem. You have a problem definition problem.

We've written about why AI pilots stall at the same point every time: it's almost always at the boundary where the AI output meets a workflow that wasn't designed for it. The pilots that don't stall are the ones where somebody did the workflow work before the pilot started. The ones that stall are the ones where the workflow was supposed to "emerge" from the AI deployment. Workflows don't emerge. They get designed.

If you've read this far and you're starting to suspect your AI initiative is really a process initiative, that's the right instinct. The same is often true of the quoting workflow problem most contractors describe as a tooling problem. The right next step is usually not to call a vendor. It's to spend a Friday afternoon with your three best operators and a whiteboard, mapping the actual process. Whatever you find is the real first project. AI is the layer on top, if it's the answer at all.

If you want a partner who'll tell you when AI isn't the answer rather than sell you one anyway, book 30 minutes with us. We've turned down enough projects to recognize the shape of the ones worth doing.

---

Keep Reading

• What AI Still Can't Do for Mid-Market Operations. The honest limits of current AI for operations work, and where the technology genuinely fits.
• Why AI Pilots Stall at the Same Point Every Time. The boundary where AI output meets undesigned workflow, and how to avoid running into it.