“AI is a tireless intern, not an oracle. Give it the intern's work, not the executive's decisions.”
Think in tasks, not in "AI strategy"
"We need an AI strategy" is how money gets wasted. Value shows up at the level of specific tasks: drafting a first version of a proposal, triaging incoming support tickets, extracting figures from invoices, answering questions against your documentation. Each is a concrete job with a measurable cost today and a measurable cost with AI.
Break a workflow into its tasks and you can see clearly which ones AI fits. The right unit of analysis is never "sales" or "the legal department" — it's "the four hours a week each rep spends summarising calls into the CRM." Specificity is what turns hype into ROI.
What AI is genuinely good at
AI earns its keep on tasks that are mostly about transforming text or other unstructured content, where being approximately right is acceptable and a human can catch the rest. The sweet spot:
- Drafting — first versions of documents, emails, summaries, code. A human edits; the blank page is gone.
- Transforming — translating, reformatting, changing tone or reading level, extracting structure from messy text.
- Triaging — routing, categorising, and prioritising large volumes of incoming work.
- Answering — questions against a body of documents, when you can show the source.
- Assisting experts — making a skilled person faster, not replacing their judgment.
Where it destroys value
The same properties that make AI great at drafting make it dangerous elsewhere. Avoid it where being wrong is expensive and hard to catch, where the task has a deterministic right answer a normal system handles better, or where it would make consequential decisions without a human in the loop.
"Calculate this customer's exact refund," "decide who gets the loan," "file this legally binding document unreviewed" — all bad fits. Either the cost of being wrong is too high, or there's a deterministic tool that's cheaper and more reliable. Using a probabilistic text model where you need a calculator or a rule is how AI projects produce confident, expensive mistakes.
Sizing the opportunity: the value matrix
For each candidate task, plot two things: the business impact if it works, and how feasible it is with today's technology. The matrix tells you what to do.
The discipline this enforces: don't start with the most exciting project, start with the most valuable feasible one. Quick wins in the top-right build credibility, capability, and the eval discipline you'll need before you attempt the hard, high-impact work. The graveyard of AI initiatives is full of ambitious moonshots funded before the team could ship a quick win.
In one line each
- Value lives at the task level, not the "AI strategy" level — break workflows into tasks and match them to what AI does well.
- AI is great at drafting, transforming, triaging, answering, and assisting experts — anywhere approximately right is fine and a human checks.
- It destroys value where being wrong is costly and hard to catch, or where a deterministic tool is cheaper and more reliable.
- Use the impact-vs-feasibility matrix: start with the most valuable feasible task, not the most exciting one.
Where to go next