And where it doesn’t
By Steele Consulting
Every business owner is being pitched AI right now. In email. In LinkedIn DMs. In vendor pitches. In the panels at every industry conference. The pitch is usually some version of: “you need AI in your business, or your competitors are going to eat you.”
Some of that is true. Most of it is noise.
At Steele Consulting, we’ve been building custom software for 24 years, which means we’ve been building it before, during, and after every wave of technology that promised to change everything. AI is a real one. It’s also being oversold at the mid-market in ways that will cost growing businesses real money in the next three years — money spent on pilots that don’t ship, tools that don’t get used, and vendors whose ROI story doesn’t survive contact with the client’s actual data.
This post is about how to tell the difference. Where AI genuinely belongs in a growing business. Where it doesn’t. And the three questions to ask before you invest a dollar in an AI initiative.
Why most AI initiatives fail
The problem isn’t the technology. AI works — often extraordinarily well. The problem is the fit.
Most AI initiatives at the mid-market fail for the same three reasons, over and over:
- The problem doesn’t have the volume AI needs to amortize its cost.
- The business doesn’t have the clean, structured data AI needs to learn from.
- The stakes of the decision don’t tolerate the probabilistic answers AI produces.
Any one of these makes an AI project brittle. Two make it a bad investment. All three make it theater.
The pattern we see is almost always the same: a business owner reads about AI, tells the team to “figure out where we can use it,” and the team dutifully picks a workflow that sounds impressive in a demo but fails all three fit tests. Six months later, the pilot is quietly retired.
The businesses that get real value out of AI don’t work that way. They start with the fit test.
The 3 Fit Tests for AI
Fit Test 1: Volume
Does this problem happen thousands of times, or dozens?
AI is expensive to build, expensive to train, expensive to maintain, and expensive to run. It amortizes over volume. When a human does something 10 times a week, hand-coded rules or a checklist are almost always cheaper. When a human does something 10,000 times a week, AI starts to pay back.
If the workflow you’re considering for AI happens rarely, or the human handling it can already do it in five minutes, AI is the wrong tool. It’s over-engineering with a fashionable name.
Fit Test 2: Data
Do you have the historical data, and is it clean?
AI eats data. Without training data — thousands of past examples of the decision you want AI to make — there’s nothing to learn from. And even with data, if it lives in six different systems, has inconsistent labeling, or was never captured in a structured way, the first year of your “AI project” isn’t AI at all. It’s data engineering.
That data engineering work is often the right investment. Just don’t confuse it with AI. If your data isn’t ready, be honest about it, and budget for the pipeline first. Skipping this step is how businesses end up with a $500K AI project that produces less value than a shared spreadsheet would have.
Fit Test 3: Stakes
What’s the cost of being wrong?
AI is probabilistic. It gives you answers with confidence scores, not certainties. That’s fine for a workflow where the wrong answer is inconvenient — a mis-triaged support ticket, a slightly-off content draft, a lead score that ranks one prospect a bit high. It’s not fine for workflows where the wrong answer is expensive, unrecoverable, or regulated.
If a wrong AI decision can lose a customer, trigger a compliance violation, or cost real money before anyone catches it, you either need to keep a human in the loop reviewing every decision (which erases most of the efficiency gain) or find a workflow where AI’s probabilistic nature is actually a fit.
Any one of these tests failing makes an AI project brittle. Two failing makes it a bad investment. All three failing makes it theater.
Where AI actually belongs
When AI passes all three tests, the value tends to be real and measurable. Four categories cover most of the good use cases we see at the mid-market.
1. Triage and prioritization
Sorting incoming support tickets by urgency and topic. Scoring inbound sales leads. Routing incidents to the right responder. These workflows are high-volume, have clean historical data (every ticket, every lead has been categorized before), and the stakes of being wrong are moderate — a mis-routed ticket gets re-routed. AI here saves real hours, at scale, with limited downside.
2. Extraction from unstructured input
Pulling structured data out of invoices, PDFs, emails, contracts, or customer messages. This is one of the most reliable ROI use cases for AI in a growing business. The technology is mature, the accuracy is high, and the alternative — having a human read every document — is often the bulk of a role.
3. Prediction on your own data
Demand forecasting for inventory. Churn prediction for accounts. Revenue projection. These work well when you have real history to train on and when the prediction feeds into a decision the business already makes routinely. They fail when you don’t have the data, when the underlying dynamics change too fast, or when the prediction has to be near-perfect.
4. Drafting and summarization
First-draft content for marketing, HR, internal comms. Meeting-note summaries. Long-document summarization. This is where a lot of the actual AI value in mid-market businesses is quietly happening today, because the humans still review the output before it leaves the building. AI accelerates. Humans decide.
Where AI doesn’t belong
The counter-list is just as important. In these cases, “regular” software, better process, or no software at all is usually the right answer.
- Small-team internal automation. Ten repetitions a week is a rules-and-checklists problem, not an AI problem. The cost of building AI for it will exceed the labor cost of doing it manually for years.
- Regulated or explainability-required decisions. Anything that has to be defended to a regulator, auditor, or customer with “here’s the exact reason we made this call” is a bad fit for AI’s probabilistic answers. Rules engines, decision trees, or human review are usually the right answer.
- Problems you don’t have data for. No data means no model, and building the data pipeline is 80% of the work. Businesses that skip this end up with an AI project that’s really a data project that’s really a data-cleanup project.
- Anything where the existing software already works. AI is not a virtue. It’s a tool. Adding it to a workflow that’s already fine adds cost, complexity, and failure modes for no marginal benefit.
- Customer-critical decisions with no human in the loop. Pricing to a specific customer. Denying a claim. Approving a credit line. The efficiency gain of removing the human is often smaller than the cost of the eventual bad decision, especially at mid-market scale where you can’t absorb reputation damage easily.
What actually happens in an AI project that works
The AI projects we’ve seen deliver real value share a pattern. They start narrow — one workflow, one decision, one dataset. They have a human in the loop for at least the first six months. They measure the wrong-answer rate honestly. And they’re built into an existing system, not sold as a standalone AI product.
The AI projects that fail also share a pattern. They start with the technology and look for a problem. They involve consultants who charge by the pilot. And they never quite ship into production because the last 20% of the accuracy gap turns out to require another six months of work that wasn’t scoped.
How this connects to the other decisions you’re making
The three fit tests slot in naturally alongside our other decision frameworks. If AI passes the fit test, the next question is Build, Buy, or Bend — which vendor tools do most of what you need, and where does custom software have to fill the gap? If AI fails the fit test, the question often becomes whether the underlying problem is really a Process Problem or a Software Problem — because a lot of “we need AI” complaints are actually workflow problems in disguise, and no amount of technology fixes those.
How we approach this at Steele Consulting
We build AI into custom software when the three fit tests all say yes, and we tell clients when they don’t. The projects where we’ve delivered the highest AI ROI weren’t the ones where the client walked in wanting AI. They were the ones where the client walked in with a specific, high-volume, data-rich problem, and we recommended AI as the right tool for it. That’s a very different conversation from “help us figure out where AI fits in our business.”
If you’re being pitched AI right now and want a second set of eyes on whether it actually fits your situation, that’s the conversation we’re happy to have. Reach out and we’ll walk through the three tests with you.