What Does “AI-First” Mean? Look at What Changed.

One rebuilds its product around AI. Another replaces customer support with a chatbot. A third buys Copilot licences for its staff. A fourth announces layoffs and leaves everyone else to work out what “AI-first” now requires of them.

These are not variations of the same strategy. They are different bets, made for different reasons, with different consequences. Calling all of them “AI-first” may conceal more than it explains.

Before judging whether an AI-first strategy succeeded, we need to establish what the company actually committed to doing.

At first glance, “AI-first” looks like an answer. Read it carefully and something is missing. AI has been placed first, but first in what?

Product design? Capital investment? An employee’s working process? Hiring decisions? Customer contact? Cost reduction?

The phrase can accommodate all of them. That flexibility is part of its appeal. Executives can signal urgency. Investors can hear a platform opportunity. Employees can hear a new expectation. Customers can expect a better product. Each audience supplies the part the announcement leaves unsaid.

One way to see how much the meaning has stretched is to return to a time when “AI-first” had a more specific job.

When the phrase pointed to a platform

Google was describing an AI-first future years before generative AI began appearing in workplace mandates.

At Google I/O in 2017, Sundar Pichai compared the shift with the earlier move to mobile. Mobile had forced Google to rethink products around touchscreens and phones. An AI-first world, he argued, would force another redesign around systems that could understand language, images, speech and context.

The announcement was accompanied by Google Lens, a new generation of Tensor Processing Units and the consolidation of the company’s AI work under Google.ai. The phrase pointed towards a product and computing architecture. It described how people might interact with software, what capabilities Google would build and the infrastructure required to support them.

That did not make every subsequent AI product inevitable or every decision correct. It did make the commitment visible. Google was putting research, specialised hardware and AI-shaped product experiences near the centre of its future.

In that version, “first” looked outward. It asked how the product should be reimagined if AI became a foundational computing layer.

The phrase now often looks inward.

When “first” became an instruction

In April 2025, Shopify chief executive Tobi Lütke published an internal memo after it began circulating outside the company. Its headline was unusually direct: “Reflexive AI usage is now a baseline expectation at Shopify.”

Employees were expected to learn how to use AI in their craft. AI usage would appear in performance and peer reviews. Before requesting more people or resources, teams would need to explain why the work could not be done using AI.

This was a different use of “first.” The immediate subject was no longer primarily the design of a product. It was the behaviour of the organisation.

The memo could reasonably be read as several things at once: a learning mandate, a productivity target, a hiring restriction and a declaration about company culture. All four readings came from the same document, but they would not be judged in the same way.

Shopify has since made the operational commitment easier to see. In 2026, its engineering team described an internal hosting platform called Quick, built to let employees publish small tools and prototypes with almost no deployment friction. More than 50,000 internal sites had been created, and over half the company had made at least one. AI was not the original reason for building Quick, but it accelerated adoption by allowing people outside engineering to generate working software.

Shopify also built Aquifer, an internal platform for durable AI agent sessions, with River as one agent working against the company’s real codebase, data warehouse and production signals. This is more substantial than purchasing chatbot subscriptions and declaring the company transformed. It changes the environment in which work can happen.

It also makes the original label harder, not easier, to summarise.

Is Shopify’s AI-first strategy the expectation placed on employees? The infrastructure built for agents? The tools offered to merchants? The pressure applied to headcount? It is all of these, but success in one does not prove success in the others.

A company can build excellent agent infrastructure while applying a poor performance measure. It can make employees faster while failing to improve the product. It can reduce staffing needs while increasing the amount of review and maintenance elsewhere.

The declaration joins these possibilities together. The consequences separate them again.

Duolingo kept the direction and changed the rule

Duolingo’s 2025 AI-first memo used the same historical analogy as Google. The company’s earlier bet on mobile had helped shape its success, and management presented AI as the next platform shift.

But the memo also attached the idea to immediate operating decisions. AI use would be considered in employee performance. Teams would have to examine whether work could be handled by AI before adding people. The company planned to reduce its reliance on contractors for tasks the technology could perform.

The public argument quickly became a referendum on whether Duolingo intended to replace people. That interpretation was understandable, but it obscured a more complicated company underneath the announcement.

Duolingo had already used machine learning in its learning systems for years. Its product strategy included AI-generated course material, conversational features and faster content production. According to the company’s current strategy overview, it published 20,500 new skills across its language courses in the first quarter of 2026, compared with 7,100 per quarter in 2025 and 1,800 per quarter in 2024.

Then one part of the workplace policy failed its own test.

In a May 2026 interview, chief executive Luis von Ahn said employees had begun asking whether they were expected to use AI simply because their usage would be evaluated. Duolingo removed AI usage as a performance criterion. Employees would be judged by their contribution and the quality of their work, with AI used where it actually helped.

That change is often described as a retreat from the company’s AI-first position. It is more revealing than that.

Duolingo did not abandon its AI-supported product features or its attempt to produce learning content at greater scale. It withdrew one interpretation of “first”: that visible use of the technology should itself count as evidence of good performance.

The company kept the direction and changed the rule.

That distinction matters because it exposes a weakness in technology-first declarations. Leaders eventually have to translate the ambition into behaviour that can be encouraged, funded and measured. Tool usage is easy to see. Useful outcomes are harder. Once usage becomes the measure, people have a reason to use the tool even when it does not improve the work.

Duolingo’s revised position is more selective. Von Ahn described its internal rule as using AI for the benefit of learners. He also acknowledged that it did not yet match the company’s best artists and designers, and that scaled content generation could produce material that should never reach users.

So what does AI-first mean at Duolingo now?

It still describes a serious product and production bet. It no longer means that AI must win every decision merely by being AI. The company can keep the label precisely because the label is broad enough to survive the policy change.

The same words can lead to a different scorecard

Klarna gives the phrase another meaning. In May 2025, it said its AI-first strategy was driving exceptional returns.

Its public reporting connects AI directly to operating leverage. In February 2026, the company said revenue had grown 104% since the final quarter of 2022 while operating expenses declined 8%. Over the same period, headcount fell 49%, and revenue per employee rose to $1.24 million. Klarna presented those figures as evidence that AI-enabled productivity was transforming its cost structure.

By the first quarter of 2026, revenue per employee had reached nearly $1.4 million.

Those figures do not establish that AI alone caused the improvement. Headcount, revenue, expenses, product expansion and market conditions can move together for several reasons. What the reporting establishes is the scorecard Klarna wants investors to use. In its version of AI-first, the cost base and the amount of revenue supported by each employee are central evidence.

That is a long way from Google’s 2017 description of a new computing interface. It also differs from Shopify’s emphasis on how employees work and Duolingo’s attempt to scale educational content.

The technology is common. The commitment is not.

This makes broad verdicts about AI-first companies almost useless. One critic may be judging customer service after automation. Another may be examining layoffs. A third may be looking for new products. Management may be pointing to revenue per employee. They can all appear to disagree about the same strategy while evaluating different decisions.

The word “first” becomes meaningful only when priorities conflict.

When speed threatens quality, which one wins? When an AI channel cannot resolve a customer’s problem, is a person still available? When a team wants another employee, what evidence is required? When an AI product draws money and attention, which other work moves down the roadmap?

These choices reveal more than the declaration because a real priority imposes an opportunity cost. Something receives money, authority or attention. Something else waits, shrinks or disappears.

Google’s early AI-first commitment can be examined through the products and infrastructure it chose to build. Shopify’s can be examined through its internal systems, working norms and merchant products. Duolingo’s can be examined through learning quality, content production and the policies it was willing to revise. Klarna’s invites scrutiny of efficiency, service and the relationship between a smaller workforce and a growing business.

The same label leads to different evidence and different questions. It cannot provide a common verdict because it never described a common strategy.

When a company says it is AI-first, the announcement tells us that management wants AI to carry unusual weight. It still leaves the consequential part unresolved.

Sometimes the answer will be a product architecture. Sometimes it will be a staffing rule, a service model or a line in the cost base. Occasionally, it will be little more than a headline.

We cannot know which from “AI-first” alone.