Most teams pick their first AI automation project the wrong way. They start with whatever is loud — the process someone complained about in a meeting, or the demo that looked impressive on LinkedIn. That's how you end up six months in with a chatbot nobody uses and a budget nobody can defend.
The good AI automation opportunities are usually boring. They're the high-volume, unglamorous processes buried in operations, finance, and support. Here is the framework we use to find them, the scoring rubric to rank them, and the anti-patterns that look automatable but will burn your first quarter.
The five signals of a real opportunity
A process is worth automating when it has all five of these. Not three. Not four. All five.
- High volume. The task happens hundreds or thousands of times a month. Automating a thing that runs twice a quarter saves you nothing and costs you maintenance forever.
- High error cost. When a human gets it wrong, something expensive happens — a missed SLA, a compliance flag, a refund, a churned customer. Low-volume plus high-error-cost can still qualify (think contract review). High-volume plus zero-error-cost usually doesn't.
- Stable rules. The logic doesn't change every month. If the policy behind the process is rewritten quarterly, you're automating a moving target and you'll spend more time patching than you saved.
- Clear owner. One person or team is accountable for the outcome and can say "yes, that's correct" or "no, fix it." No owner means no one to validate the output, approve edge cases, or defend the project when someone asks why it exists.
- Measurable outcome. You can state the before-number and the after-number. Hours saved, error rate, turnaround time, cost per transaction. If you can't measure it, you can't prove it worked, and you won't get funded for the second one.
A scoring rubric you can actually use
Score each candidate process 1 to 5 on the five signals, then multiply nothing — just add, and weight the two that matter most. Volume and error cost carry double weight because they drive the return; the other three are feasibility gates.
- Volume (×2): 1 = a few times a month, 5 = thousands.
- Error cost (×2): 1 = a typo nobody notices, 5 = regulatory or revenue impact.
- Rule stability (×1): 1 = changes constantly, 5 = hasn't changed in a year.
- Owner clarity (×1): 1 = nobody owns it, 5 = named accountable owner.
- Measurability (×1): 1 = purely subjective, 5 = already tracked in a dashboard.
Max score is 35. Our rule of thumb: anything above 24 is a strong candidate, 18 to 24 is worth a discovery conversation, and below 18 you skip it — no matter how exciting the demo looks. The value of the rubric isn't the exact number. It's that it forces the argument out into the open. When a stakeholder pushes their pet project, you can point at the two feasibility gates it fails.
Where the value actually hides
Nobody nominates their best opportunities, because the best ones don't feel like problems — they feel like "just how the work gets done." Go look for them directly.
- The copy-paste tax. Anywhere a human moves data between two systems by hand — CRM to invoicing, email to spreadsheet, PDF to ERP. This is high volume, stable, and measurable almost by definition.
- The triage queue. Inbound requests that a person reads and routes: support tickets, claims, applications, vendor emails. Classification and routing is one of the most reliable wins in AI automation today.
- The "waiting for approval" gap. Processes that stall because they need a summary, a first-pass draft, or a data pull before a human decides. You don't automate the decision — you automate everything up to it.
- The month-end scramble. Anything that only happens in a painful burst — reconciliations, reporting, renewals. High error cost, clear owner, and a deadline that makes the ROI obvious.
Anti-patterns: shiny but worthless
These pass the sniff test and fail in production. We've watched all of them fail.
- The unstable process. It looks perfect — until you learn the rules changed three times last year and are about to change again. Fix the process first, then automate it. Automating chaos gives you faster chaos.
- The no-owner orphan. Everyone agrees it should be automated; nobody will sign off on the output. Without an owner, edge cases pile up unresolved and the project quietly dies. No owner, no build.
- The judgment call in disguise. It looks rules-based but every case secretly depends on context, relationships, or unwritten exceptions. Underwriting a nonstandard deal, handling a strategic customer complaint. AI can assist here, but full automation will make confident, wrong decisions.
- The vanity chatbot. A conversational interface bolted onto a problem that was never a conversation. If users don't currently ask questions in natural language to get this done, a chatbot adds friction, not speed.
- The low-volume perfectionist. A genuinely hard, genuinely interesting problem that happens eleven times a year. The engineering is fun. The ROI is imaginary. Save it for later.
Sequence for momentum, not just score
Your highest-scoring candidate isn't always your first build. The first project has a second job: proving the model works so you get to do the next one. So bias the first pick toward something with a clear owner who wants it, data that already exists, and a result you can show in a slide within 60 days. A 26 that ships and gets adopted beats a 32 that's still in a data-cleanup swamp two quarters later. Score to decide what's worth doing; sequence to decide what to do first.
The framework above gets you a ranked list you can defend. Turning that list into a real roadmap — with effort estimates, data readiness checks, and the honest "don't build this yet" calls — is exactly what an AI Readiness Assessment is for. If you've got a shortlist and want a second set of eyes before you commit budget, that's the conversation to have.