AI-first means treating AI as the default option you consider before you add a person, a process or a tool. It describes how decisions get made, not what a team can already do. That distinction is worth holding on to, because most of the companies that made the phrase famous ended up measuring something else entirely: how much AI their staff used.
On 2 September this year, Meta told its engineers it would not use AI adoption dashboards or token counts to evaluate impact. The memo, reported by The Information, put AI-assisted code changes at 93% and said adoption was not the goal. Five months earlier, Duolingo had dropped AI use from its performance reviews.
Neither company has gone cold on AI. Both abandoned the same specific thing. That pattern is the most useful thing a marketing leader can know about AI-first at the moment, and it has nothing to do with picking the right label for your team.
Sundar Pichai, April 2016. In that year's Founders' Letter he wrote that Google would move from a mobile-first to an AI-first world, and everything around the line is about products: Search, Photos, voice, translation. There is nothing in it about how employees should work.
The employee meaning arrived nine years later, over about three weeks.
On 7 April 2025 Shopify's Tobi Lütke published his own internal memo making reflexive AI use a baseline expectation, with teams required to show AI could not do the job before asking for headcount. It never uses the phrase “AI-first”. On 28 April, Duolingo's all-hands email did, and attached a near-identical set of rules.
So the word and the machinery came from two different companies in the same month and have been welded together ever since. The fusion is about eighteen months old, not a decade.
It depends who you ask, and the gap is the interesting part. Gartner, which has more influence over how your board uses this vocabulary than anyone, defines AI-first as making AI a default consideration when you address a business challenge. It also says plainly that forced adoption of AI is not the answer.
That is not what the term has meant in practice. Every documented corporate adoption since April 2025 arrived with enforcement attached: headcount gates, contractor freezes, AI use written into performance and peer review, onboarding deadlines with consequences.
So when someone announces that the company is AI-first, they might mean a decision rule, or they might mean a set of rules about you. Before anyone argues about the label, it is worth finding out which one is actually on the table.
They stopped. Duolingo dropped AI usage as a formal performance criterion in April 2026, roughly a year after introducing it. Fortune reported that staff had begun asking whether they were meant to use AI for AI's sake, and Luis von Ahn said the company had been pushing something that in some cases did not fit.
Meta's version is fresher and less flattering. Its November 2025 policy made AI-driven impact a core expectation for 2026, an internal leaderboard ranked employees by the tokens they consumed, and people did the obvious thing with it.
By September the guidance had been rewritten around work quality and impact instead. Two of the best-documented AI-first mandates anywhere have now dropped the same mechanism, five months apart, for the same stated reason.
There is evidence, and it is better than the anecdotes. A 2026 working paper by Jie Gong and colleagues studied a medical-device company that required around 5,000 employees to submit 200 AI queries a month, then cut the target to 100 through a staggered rollout across branches.
They had the complete query logs. Under the 200-query target, 32% of queries were either repeats or nothing to do with work. Halving the target cut query volume by 30%, improved the quality of what was left, and increased sales by 7% among sales employees.
Read that twice, because it cuts both ways. The mandate did increase first-time adoption, which is a genuine result. It was the level of the target that turned a tool into a scoreboard.
Worth noting the limits: this is a working paper, the company is anonymous, and nobody has run the equivalent study inside a marketing team.
Because it is one of the oldest findings in management research, and AI has handed it a very clean modern example. V.F. Ridgway wrote it up in Administrative Science Quarterly in 1956: install a quantitative measure and people reorganise around the measure rather than the purpose behind it.
His illustration could have been written for this. Public employment interviewers judged on the number of interviews they conducted ran fast interviews and placed almost nobody. Steven Kerr made the same argument in 1975, and Holmström and Milgrom later formalised it: where some parts of a job are easy to measure and others are not, strong incentives on the easy parts pull effort away from the hard ones.
Marketing is close to a worst case. Assets produced, emails sent, posts shipped and prompts run are all trivially countable. Distinctiveness, judgement and taste are not, and those are the parts that decide whether any of it works.
That was my instinct, and it does not survive contact with the definitions. AI-native is the bigger claim, not the safer one.
IBM defines it as designed from the ground up with AI at the core rather than bolted on later. Harvard Business School Online draws the same line: the thirty-year-old company systematically adding AI is AI-first, and the startup built around AI from day one is AI-native.
A marketing team that started using AI in 2024 does not clear that bar. It is not a neutral alternative either, because AI-native is the aspirational word the mandate companies use about themselves. Meta's own performance memo framed its change as moving toward an AI-native future.
There is a practical reason to be careful too. AI-first describes an internal policy, while AI-native describes a capability, and US regulators have spent the past two years bringing cases against companies whose AI capability claims ran ahead of what their systems did. If you want the longer version of that vocabulary, we have covered it in Explaining AI-Native Marketer and What Is an AI-Native Marketing Team?.
The gate, not the metric. Look at which mechanisms have survived.
Duolingo dropped the review weighting and kept the rule that headcount is only approved where a team cannot automate more of its own work. Shopify's requirement that teams show AI cannot do the job before asking for more people has never been withdrawn. I could not find a single company that has publicly reversed a gate.
The difference is structural. A gate asks one budget-holder one answerable question at the moment they want something. A metric asks every individual, continuously, to look like they are using AI.
If I were installing one thing in a marketing team this quarter, it would be the question rather than the dashboard. Before we brief an agency, hire a contractor or buy another tool: what would this look like if we built it with AI first?
Answer it honestly. Sometimes the answer is that AI cannot do it, and that is a good answer.
Three things, and none of them is weak. Gartner forecasts that organisations sustaining an AI-first strategy will do measurably better than their competitors by 2028, so a marketing team that publicly refuses the vocabulary is declining a category its board has been told to aim at.
Second, mandates demonstrably work on adoption. Coinbase compressed an adoption curve projected to take quarters into a single week by setting a deadline with consequences attached, and Gong's study found the same first-time effect.
Third, the null results people quote against AI, including the Danish study of 25,000 workers that rules out effects larger than 2% on earnings and hours, measure payroll records rather than output quality. Better marketing would not show up in that data anyway. A serious academic defence of proxy measures exists, too: when the real outcome is slow or genuinely unmeasurable, rewarding something adjacent may be the least bad option rather than a category error.
I still think the balance of evidence points the other way. But “mandates do not work” is not what the research says, and I would rather be accurate than tidy in a piece about overclaiming.
Stop arguing about which word goes on the team. Labels are free, which is exactly why they prove nothing, and the AI marketing vocabulary has multiplied faster than anyone's ability to evidence it.
McKinsey's most durable finding across two survey waves is that the small group of companies getting real financial value out of AI are not the ones using the most of it. They are the ones who redesigned how the work happens, and they are around twice as likely to have a defined process for measuring what their AI initiatives actually did.
That is the whole thing. Put the AI question at the gate where budget decisions get made, measure the work rather than the tool use, and then call the team whatever you like.
AI-first means treating AI as the default option you consider before adding a person, a process or a tool. Gartner defines it as making AI a default consideration when you address a business challenge, and says explicitly that forced adoption is not the answer. It describes how decisions get made rather than what a team can already do.
AI-first is a decision posture an existing organisation can adopt. AI-native is an origin claim: IBM and Harvard Business School Online both define it as being built around AI from the start rather than adding it later. On those definitions AI-native is the larger claim, so a team that adopted AI recently cannot really make it.
No, and the difference is worth keeping. AI-first describes how a team decides what to do, while agentic describes software with enough autonomy to plan and act within set guardrails, plus the workflows built around it. A team can be AI-first without running a single agent, and it can run agents without any company-wide AI-first policy.
The evidence says no. Duolingo removed AI use from its reviews in April 2026 and Meta dropped adoption dashboards and token counts in September 2026, both after staff started using AI to satisfy the measure. A 2026 field study of around 5,000 employees found that 32% of queries under a monthly usage target were repeats or unrelated to work.
Measure the work rather than the tool. Put the AI question at the point where budget decisions get made, so anyone asking for headcount, an agency or a new tool has to say what the AI version would look like first. Then judge the output on the measures the team already answers for.
SaaStrix is where B2B marketers become agentic marketing leaders, and you can try it free for five days.
.svg){kind=icon}