Our take: strategy is a list of things you will not do with AI
Why an AI strategy is defined by its exclusions, what a useful list of refusals contains, and how it makes the rest of the decisions easy.
The short answer
Most documents called an AI strategy are a list of things the business would like to do, which is a wish list with a cover page. The half that makes it a strategy is the list of things it will not do, and why. Exclusions are what make every subsequent decision fast: when a vendor pitches, when a manager brings an idea, when an article describes what a competitor announced, the answer is usually a line that was written down calmly months earlier rather than a fresh debate under pressure. A useful exclusion list has four parts: uses you will not touch on principle, data that will not leave your approved environment, decisions that stay human regardless of how good automation becomes, and the much longer list of things that are simply not this year’s priority. Writing it is also the fastest way for a leadership team to find out whether it actually agrees, because everyone nods at ambitions and people differ sharply on refusals.
The four kinds of exclusion
| Kind | Examples | Why it is written down |
|---|---|---|
| Not on principle | Automated decisions about people; AI that pretends to be human; generated content published unreviewed; surveillance of staff | Encodes your values before the convenient moment arrives |
| Data that will not move | Customer records, health or financial details, client documents, source code, outside approved tools and terms | Turns a judgement call into a rule anyone can follow |
| Decisions that stay human | Hiring, pricing exceptions, refunds above a threshold, service refusals, anything affecting someone’s rights | Protects both the person affected and the business |
| Not this year | Personalisation, agents, voice, predictive maintenance, custom models, a product | Focus; and a date to revisit rather than a permanent no |
Writing yours
- Start with the principled refusals, and make sure the leadership team genuinely agrees rather than assents.
- Write the data rules in a form any employee can apply without interpreting.
- List the decisions that stay human, by name, with the reason.
- Choose this year’s two initiatives and write everything else on the not-this-year list with a review date.
- Circulate it so people can use it; an exclusion list nobody has read prevents nothing.
- Revisit quarterly, moving items off the not-this-year list deliberately rather than by drift.
- Add exclusions when you learn something, especially after a near-miss.
What the list does for the business
It answers vendors in one line. It lets staff act without asking. It gives the board a clear statement of position rather than a list of hopes. It protects the two initiatives that matter from the eight that would dilute them. And when something goes wrong in the industry, a phishing incident through an AI tool, a discrimination case, a data breach at a provider, the business can point to the line it wrote before it happened, which is worth more than any reaction afterwards.
What this means for you
Write the half of your AI strategy that says no: the uses you refuse on principle, the data that will not move, the decisions that stay human, and the things that are not this year with a date to revisit. Circulate it, use it as the standing answer to pitches and ideas, and review it quarterly. It focuses the two initiatives that matter, encodes your position before you need it, and is the part of the document you will actually use.
Frequently asked questions
Is a list of refusals not just negativity?
It is focus. A small business can pursue perhaps two AI initiatives well in a year. A list of what it will do has ten items and delivers two by accident; a list that names the two and explicitly excludes the rest delivers them on purpose. The exclusions also encode your risk position, which is what protects you when a plausible idea arrives with momentum behind it.
What belongs on the list?
Four kinds of exclusion. Uses you will not touch on principle, such as automated decisions about people. Data that will not leave your environment or your approved tools. Decisions that stay human regardless of how good the automation gets. And things that are simply not this year's priority, which is the largest category and the one most often left unwritten.
How often should it change?
The principled exclusions rarely, because they reflect your values and your legal position. The not-this-year list every quarter, as capabilities and priorities change. Reviewing it is also the moment to notice whether something you excluded last year has become obvious, which is a healthy sign rather than an inconsistency.