What Unilever and PepsiCo forbid AI to do (and why it concerns you)

While the market pushes to automate everything, two of the world's largest advertisers do the opposite: they write down, in black and white, what AI is not allowed to do in their house.

When you suggest to an SMB leader that they write AI usage rules, the reaction is usually the same: that is bureaucracy, we will see later. The argument that lands is showing that the biggest players already do it.

Unilever: tools per use case

Unilever did not roll out a general AI across the board. The group has dedicated tools per use case: customer sentiment analysis, product page generation in the brand tone. Each tool has a scope, and the result is measured: 90% less time processing customer emails.

The lesson is counter-intuitive: the measurable gain comes from restricting scope, not from opening the floodgates.

PepsiCo: explicit bans

PepsiCo goes further with an ethics framework co-developed with Stanford, which explicitly forbids AI for two uses: recruiting and individual consumer targeting.

This is not timidity. It is the recognition that a brand is defined as much by what it refuses to do as by what it produces.

A brand is defined as much by its bans as by its output. AI does not change that rule, it makes it urgent.

What it validates for everyone else

If groups of that size, with those resources, take time to document written limits, then formalized framing is not a studio whim: it is standard practice. That is the credibility argument to use with a client who finds the exercise pointless.

Concretely, it means the deliverable of a brand foundation must contain two lists: what AI may do in your context, and what it will never do. The second is often the hardest to write, and the most useful.

---

Frequently asked questions

Isn't writing bans limiting yourself?

No. Unilever gets 90% less time precisely because each tool has a defined scope. The frame produces the result, it does not cap it.

What bans should an SMB set?

It depends on the business, but recurring ones are: no AI on decisions about people, no publishing without human review, no unverified claims.

Do you need a full ethics framework?

Not necessarily. One page of clear rules, known by the team and actually applied, beats a thirty-page document nobody reads.