Where AI Genuinely Helps (and Where It Doesn't)

Every team has a list of jobs they hope AI will take over. About two in eleven survive contact with reality — and that is good news, because the two are usually worth more than the eleven combined. Here is the map we use, refined over years of telling people both answers.

Where It Shines

• Drafting from source material. First drafts of reports, summaries, and responses grounded in your documents. A person still reviews — but the blank page is gone.
• Triage and routing. Sorting incoming requests, tickets, and applications by content. High volume, clear categories, human confirms the edge cases.
• Retrieval across messy knowledge. Finding the right paragraph across thousands of pages in seconds. This alone justifies most of our engagements.
• Repetitive transformation. Reformatting, extracting, reconciling — work that is rules-shaped but too varied to script.

Where It Doesn't

• Final judgment calls. Hiring, lending, medical, legal outcomes — anything where a wrong answer harms a person. AI can prepare the file; a person decides.
• Negotiations and relationships. Trust is built by people showing up. Sending a model signals the opposite.
• Thin or forbidden data. Fewer than a few hundred examples, or data you should not have in the first place. Both end badly.
• Work nobody understands. If no person can describe how the task is done well, there is nothing to teach the system.

Scoring Candidates

List every candidate job and score it on volume, pain, data readiness, error tolerance, and reversibility. Eleven candidates typically compress to two worth building, three worth deferring, and six worth declining with thanks. The declines are the valuable part — each one is money and months given back.

Running an Honest Trial

For the two survivors, run a short trial on real inputs with real users. Track results where everyone involved can see them — a shared project board keeps the trial honest. Lock down access from day one: trial data deserves the same credential discipline as production data, because it becomes production data the moment the trial succeeds. Judge at the end by the numbers you wrote down at the start, not by how impressive the demo felt.