How to Tell If Your Problem Is Actually a Good Fit for AI

A framework built from the inside out, from shipping a real AI product in legaltech.

Most conversations about AI start in the wrong place. Companies ask: What can AI do? They watch demos, read case studies, and try to reverse-engineer which of their problems might fit. It's a little like watching someone else's highlight reel and trying to figure out what sport you should be playing. The better question is: What does a great AI opportunity actually look like? I've spent the last year helping build and launch an AI-native product in the legaltech space. Specifically AI-powered legal bill review. It's a niche vertical, but it turned out to be a near-perfect petri dish for understanding where AI creates real value and where it doesn't. Here's the framework I've distilled from that work. Ten criteria that signal a strong AI opportunity. The more boxes you check, the better your fit.

1. Your inputs are unstructured and high-variance

If your data lives in clean rows and columns with consistent formats, you probably don't need AI. In those cases a well-built rules engine will do the job faster and cheaper. But if meaning depends on context, phrasing, and nuance. If the same concept gets expressed fifty different ways depending on who's writing, that's where rules-based systems start to break. Humans handle this intuitively. We pattern-match. LLMs can now do this at scale. Legal invoices are a textbook example. "Reviewed documents" could mean five minutes or five hours. "Conference with client" tells you nothing about what was actually discussed. The meaning is embedded in the phrasing, the context, and the history. That's not a bug in the data, it's the nature of the domain.

2. You can't write clean rules for it, but humans can do it by feel

There's a category of decisions that resist codification. You can't write a flowchart for them. But put an experienced person in front of the problem and they'll recognize the answer almost immediately. That intuition, that sense of "this looks off", is exactly what AI is now able to approximate. The task isn't "apply rule X, get outcome Y." It's "does this look like something problematic?" That requires clustering, grouping, and analogy. It requires recognizing that two things that look different on the surface are actually the same problem. Experienced billing reviewers can spot block billing, catch duplicate entries, and flag vague descriptions — not because they're following a rulebook, but because they've seen the pattern before. AI can now develop that same pattern library.

3. The volume is high, but each decision is low-to-mid stakes

This is a sweet spot that's easy to overlook. Full automation is a very high bar — the stakes need to be low enough to accept errors. Pure human review is too slow and too expensive at scale. But the middle ground: high volume, moderate stakes, with humans remaining in the loop, is where AI consistently earns its keep. No one loses sleep over a single $50 line item on a legal invoice. But across a portfolio of matters, across thousands of invoices, across millions of line items, the aggregate impact is significant. AI can work the surface area that humans physically can't.

4. Expert judgment exists, but it's applied inconsistently

Here's one of the more underrated signals. If the "right answer" is knowable, if domain experts broadly agree on what good looks like but that judgment isn't being applied consistently across people, time, or workload, AI can act as a standardization layer. Humans get tired. They apply different standards on different days. They disagree with each other in ways they can't fully articulate. AI doesn't have a Monday morning and a Friday afternoon. It applies the same judgment, the same way, every time. In legal billing, two reviewers looking at the same invoice will often reach different conclusions — not because one is wrong, but because the standards are fuzzy and human. AI makes the fuzziness explicit and the standards consistent.

5. You have feedback loops that can teach the system

This one is easy to underestimate early, and impossible to ignore later. AI without feedback loops stalls. AI with strong feedback loops compounds. Every approved or rejected decision is a data point. Every correction is a signal. If your workflow generates a steady stream of "good call / bad call" information (and you can capture and use that information) your AI system gets smarter over time rather than staying static. Build for the feedback loop from day one, even before you know exactly how you'll use it.

6. One party knows more than the other

Information asymmetry is everywhere in business. One side of a transaction has more expertise, more context, or more control over how information is presented. The other side is flying partially blind. AI can act as an intelligence equalizer. Not by giving the disadvantaged party access to what they don't have but by helping them make better sense of what they do have. In legal billing, law firms write the invoices. They know what the work involved. Clients often don't have the internal expertise to evaluate whether a charge is reasonable. An AI layer doesn't close that gap entirely, but it gives clients a fighting chance.

7. There's hidden structure in the mess

Your data looks like chaos. But buried inside that chaos is structure — categories, patterns, themes — that just hasn't been made explicit yet. This is one of AI's underrated superpowers. Not just analyzing structure, but creating it. Taking a wall of unstructured text and turning it into categories, groupings, trends, and themes that a human brain couldn't hold simultaneously. Turning invoice line items into a coherent picture of how a legal matter was staffed, how time was distributed, which tasks were routine and which were unusual, that's structure extraction. And it changes what's possible in terms of analysis, negotiation, and decision-making.

8. The work is cognitively repetitive, not just physically repetitive

This is the distinction that separates AI's lane from traditional automation. RPA (robotic process automation) is great at repetitive clicks. The same sequence of actions, over and over, with no variation. That's not AI's wheelhouse. What AI handles is repeated thinking. The same type of cognitive task, applied again and again to different inputs. It's too cognitive for rules-based automation. It's too repetitive for humans to sustain attention and quality. That's the sweet spot. Reviewing thousands of invoice entries that are all slightly different, all requiring the same kind of judgment — that's cognitively repetitive work. Humans burn out on it. AI doesn't.

9. You can design for "assistive" rather than "autonomous"

Full automation is a much harder bar to clear, and for most enterprise use cases right now, it's the wrong goal. If you can design a workflow where AI makes recommendations and humans make final calls, you dramatically lower the stakes of any individual error and dramatically increase the speed of adoption. The wedge isn't "replace the human." It's "make the human faster, more consistent, and less fatigued." The model we built — AI flags issues, humans decide what to do about them isn't a halfway measure. It's the right architecture for where we are. And it's exactly what risk-averse organizations can actually get comfortable with.

10. Small improvements have outsized economic impact

AI doesn't need to be perfect. It needs to be better than the baseline. If your current process catches a certain percentage of billing violations, and AI catches a few percentage points more — what's that worth in dollars? If the aggregate financial surface is large enough, the math gets interesting very quickly. The legal billing market runs through billions of dollars a year in invoice spend. Catching an additional 5% of non-compliant charges isn't a rounding error. That's the kind of economic leverage that makes AI investment easy to justify, even when the technology isn't flawless.

The Pattern Behind the Pattern

Looking across all ten criteria, there's a common thread: AI works best when the problem is too nuanced for rules, too large for humans, and too important to ignore.

Legal bill review happened to hit almost every one of these signals, which is part of why the project worked. But I've seen the same pattern show up in compliance review, customer support triage, contract analysis, and half a dozen other domains.

Before you chase the AI use case, check the checklist. The technology is only as good as the fit.