If AI usage becomes mandatory, what measures whether it helped?

Tobi Lütke shared an internal Shopify memo in April 2025^{[1]April 2025} making habitual AI use a baseline expectation. Shopify later described^{[2]Shopify later described} how experimentation, model access and sharing had become part of its working culture.

The message is persuasive as an invitation to learn. Refusing to experiment with powerful new tools is a poor way to understand them. The harder question appears when the instruction becomes a performance requirement.

If everyone must show that they use AI, most people will find a task on which to use it. That produces an adoption number. It doesn't establish that a customer problem was solved faster, that the output improved or that the employee spent less time checking mistakes.

We can separate three measurements: adoption, successful task completion and business value. An internal policy that measures only the first can reward activity instead of results.

Consider an analyst preparing a weekly report. A model saves twenty minutes drafting text but introduces an incorrect number that takes forty minutes to find. Tool usage increased; net productivity fell. In another context, the same tool might help an analyst discover a problem that was previously invisible. The sign of the effect depends on the full task.

A company needs room for playful tests, where people can learn without affecting customer records. It also needs production processes for finance, security, permissions and public information. Both can exist within a strong AI culture.

Our interest in Bonfÿre is to make those contexts explicit: models and agents can prepare work, while the receiving application owns the records and permitted actions.

A useful practical rule is to encourage people to try tools, share what worked, document failures and measure outcomes. That is an extension of Shopify's experimental spirit, not an argument against it.

The baseline worth defending is curiosity plus evidence about real work.