Future of Work
What 'AI-first' companies actually change about how people work
Every second company now claims to be "AI-first". Strip out the ones that mean "we bought licences and sent a memo", and the genuine article is doing something more interesting: rewiring who decides, who checks and who ships.
The memos started arriving in 2025. Shopify's chief executive told staff that using AI was now a baseline expectation and that teams must show a job can't be done by AI before asking for headcount; Duolingo declared itself "AI-first" and caught a week of headlines for it. Since then the phrase has spread through earnings calls like knotweed. As a consultant, I've now seen the inside of both kinds of company: the ones where AI-first is a slide, and the ones where it's a different way of running the place. The difference is not the tools. It's four structural changes.
1. The bottleneck moves to review — so review gets redesigned
When drafting becomes nearly free — first versions of code, copy, contracts, campaign plans arriving in minutes — the constraint shifts to checking. Firms that don't notice this simply drown their senior people in review work and wonder why quality slipped. The genuinely AI-first ones redesign the checking layer: explicit sign-off gates for anything touching money, customers or the law; tiered review, where routine output gets sampled rather than read line-by-line; and named human owners for every AI-produced artefact. The quiet rule that separates the grown-ups: the machine can do the work, but it can never own the work.
2. Documentation stops being admin and becomes fuel
AI systems are only as good as the context they're given, so the tidy-minded suddenly inherit the earth. Style guides, process documents, decision logs, well-labelled data — assets that used to moulder in wikis — now directly determine output quality, because they're what gets fed to the models. AI-first firms treat prompt libraries and context documents as maintained infrastructure with owners and version control. It is the great revenge of the people who always wrote things down.
The machine can do the work, but it can never own the work.
3. Teams get wider and shallower
When each person plus an AI toolkit covers more ground, the arithmetic of team design changes. The visible symptom is smaller project teams with wider spans — one marketer running what used to need four, a two-person "team" shipping a product feature with agents handling the connective tissue. The controversial symptom is at the entry level: if AI does the work juniors used to learn on, where do seniors come from? The honest firms are rebuilding junior roles around supervision, verification and client contact rather than production — an apprenticeship in judgement instead of an apprenticeship in typing. The dishonest ones are just hiring fewer juniors and calling it efficiency. That bill arrives in five years.
4. Decision rights move — explicitly
The subtlest change is to who may decide what. AI-first operating models write down which decisions the machine can take alone (reorder the stock, route the ticket), which need a human glance (send the quote), and which demand a named human decision regardless of what the model recommends (hire, fire, price, apologise). Committing that to paper does two things: it stops the quiet creep of unreviewed automation, and it forces a conversation most firms have never had — what, exactly, is our judgement for? Interestingly, the same discipline governs the deployment decisions in physical automation, where the winning robots are the ones given narrow, explicit authority.
What the sceptics get right
Two things, as of early 2026. First, the measured productivity gains, while real — controlled studies keep landing in the tens of per cent for drafting-heavy tasks — are uneven, and self-reported time savings routinely exceed what output data shows; some of the saved time evaporates into checking the AI's work. Second, "AI-first" has already been used as air cover for ordinary cost-cutting, which poisons trust and makes honest adoption harder everywhere else. If the strategy only ever produces redundancies and never produces redesigned work, staff can be forgiven for reading it as a euphemism. The productivity evidence deserves its own sober look — we've given it one in our piece on the four-day week and AI.
What to copy, whatever your size
You don't need a memo or a rebrand. Pick one process, and make four changes: let AI produce the first draft; define who reviews it and how thoroughly; write down the context documents the AI needs and keep them current; and record which decisions stay human. That's AI-first in miniature — and it's worth doing before your competitors' version of it stops being a slide.