How to identify AI opportunities in your company (method, not hype)
"Where do I put AI?" is the wrong question. The right one: which work in your operation has the signals AI handles well? The method, without the hype.
The question "where do I put AI in my company?" is the wrong question — and it's the one behind AI projects that go nowhere. The right question is: which work in my operation has the signals that AI handles well? This article gives you the method to find it: map, spot the signals, classify and prioritize. No hype.
Start with the work, not the technology
- 1Map the processes that consume the most hours of your team
- 2Spot the opportunity signals in them (below)
- 3Classify each opportunity: rules, AI, or agent?
- 4Prioritize by impact and feasibility — not by novelty
The signals that there's an AI opportunity
- Unstructured text at volume: emails, tickets, comments, documents that someone reads and classifies by hand
- Repetitive decisions with nuance: approve/reject, prioritize, assign — similar but not identical cases (if they were identical, rules would be enough)
- Bottlenecked knowledge: questions only two people on the team know how to answer, and everyone asks them
- Manual information lookup: "let me check the system/the drive/my email" as a recurring customer-service step
- Data that exists but nobody uses: sales history, conversations, logs that could anticipate problems and today just take up space
Classify before you get a quote
| The work is… | The likely solution is… |
|---|---|
| Repetitive with fixed rules | Traditional automation (cheaper, more predictable) |
| Interpreting text/documents/nuance | AI applied to the process |
| Conversing and answering questions with your information | Connected chatbot |
| Completing tasks across several systems | AI agent |
| A messy process | Fix it first — AI doesn't fix the mess, it speeds it up |
Prioritize with two questions
For each classified opportunity: how much is solving it worth? (hours, errors, lost sales — estimate honestly) and how feasible is it? (does the data exist?, is the process stable?, is the risk of error tolerable?). High on both → pilot. High value, low feasibility → fix feasibility first (almost always: data). Low value → drop it without guilt, even if it's the "coolest" idea.
The right pilot
A serious AI pilot: one process, one measurable outcome, weeks not months, and success criteria defined BEFORE you start. If it works, you scale it; if it doesn't, you learned cheaply. Be wary of any proposal that starts with "end-to-end transformation".
What you may also be wondering
Do I need a lot of data to use AI?
It depends on the case. To automate text interpretation (emails, tickets), current models work well without custom training. For predictions about your business you do need history. That's why feasibility is assessed per opportunity, not in general.
Where does it make sense to start evaluating?
Two common candidates to evaluate are customer service — questions and follow-up using your own information — and administrative processes that involve classifying, extracting or recording information. But the right starting point comes from YOUR process map, not from a statistic.