AI-native enterprise

What is an AI-native enterprise?

An AI-native enterprise is a company whose core processes, products and decisions were designed around AI models from the start. The models are not a feature added to the way the company already worked. They are the way the company works. Take them away and the business does not slow down, it stops.

That last sentence is the test most people now use to separate the label from the marketing. The venture firm CRV wrote it down for founders in March 2026: if you remove the AI, does the product cease to function? Not degrade, not lose a nice feature, but stop working entirely. If the answer is yes, the product is AI-native. If the company would carry on much as before, with a few people doing by hand what a copilot did for them, it is AI-enabled.

The test works on a whole company as well as on a product. A law firm that uses a drafting assistant is AI-enabled: switch it off and the lawyers keep drafting, only slower. A firm that takes in a case by having a model read the file, pull the precedents, draft the advice and send only the exceptions to a lawyer has built its intake on the model. Switch it off and there is no intake process left, only a queue.

AI-enabled versus AI-native: the removal test

Sales decks use the two terms as if they were the same thing, so it helps to compare them on one dimension: what happens when the models go away.

AI-enabled

The process existed before the models did, and it still has the same shape. A model was inserted at one step: summarising the mail, suggesting the reply, drafting the report. People still own every hand-over. Remove the model and that step goes back to being manual. The company loses time and keeps working. Practically every established company that uses AI today sits in this group.

AI-native

The process was drawn from scratch with the model as the worker and people as the reviewers. The data was shaped so a model can read it, the steps were defined around what a model does well, and the human role moved from doing the work to checking the exceptions. Remove the model and there is no process, because nobody ever did the work by hand in that shape.

A third phrase, AI-first, sits in between and mostly describes an intention rather than a state. When Shopify's CEO wrote in April 2025 that reflexive AI usage was now a baseline expectation, and that teams had to show why AI could not do the work before asking for more headcount, that was an AI-first policy inside a company that was, and largely still is, AI-enabled. A policy like that pushes a company towards native. It does not make it native on its own.

What an AI-native company looks like in practice

The people who study these companies keep listing the same traits, and none of them is about which model you use.

Data and interfaces designed for agents to read. Product data, customer history and internal APIs are structured so a model can query them without a person copying values between screens. Andreessen Horowitz described the new generation of business software in May 2026 as built for machine readability, with agent orchestration as a first-class feature, instead of a screen for a person with an API added later.

Processes redesigned rather than automated as-is. A BCG team wrote in June 2026 that most organisations introduce AI but leave the operating system of work untouched: they put copilots on unchanged processes and get productivity gains of 10 to 20 percent. The companies that captured far more had redesigned their processes end to end for agents, with people supervising outcomes instead of performing steps.

Small teams with high output. The clearest public signal is revenue per employee. A company built on models needs far fewer people per unit of output than a company where people do the work with a model's help.

Evals and cost per task as operating metrics. Because the model does the work, the quality of its output is an operations number. These companies run an evaluation set against every prompt and model change, the way a factory runs quality control on every batch. Next to it sits cost per task: the model spend, the compute and the human review time needed to complete one unit of work. A process that gets the right answer at ten times the expected cost has a defect, even if the accuracy score looks fine.

AI usage as a hiring criterion. Fluency with AI tools is assessed at hiring and in reviews, the way spreadsheet skills were a generation ago. McKinsey's business-building practice, which interviewed leaders at fifteen AI-native companies, lists hiring for AI fluency among the things these companies do to keep adoption growing. Shopify's memo made the same expectation explicit inside an established company.

Examples from 2025 and 2026

Lovable. The Stockholm company behind the app-building tool of the same name told TechCrunch in March 2026 that it had passed 400 million dollars in annual recurring revenue with 146 full-time employees, around 2.7 million dollars per employee. The product passes the removal test: without models there is nothing to sell.

Cursor. Bloomberg reported in March 2026 that the AI code editor built by Anysphere had passed 2 billion dollars in annualised revenue, double the figure of three months earlier, with about 60 percent coming from corporate customers. The company has not published a headcount we could verify, so we leave the per-employee figure out.

Klarna, as the cautionary case. In February 2024 the payments company said its AI assistant had handled 2.3 million customer chats in its first month, the work of 700 agents. In May 2025 its CEO told Bloomberg that cost had weighed too heavily in that decision and that the result was lower quality, and the company started recruiting people for customer service again. The assistant still handles about two thirds of enquiries. Humans took back the hard ones. The AI did what it was built for. The staff went before the process was redesigned.

The analyst framing. Gartner put AI-native development platforms on its list of strategic technology trends for 2026, announced in October 2025, and predicts that by 2030 around 80 percent of organisations will have turned large software engineering teams into smaller, AI-augmented ones. Note the scope: Gartner uses AI-native for the platforms and the teams that build software, not as a label for whole companies.

What it means for an SME

Most established companies will be AI-enabled for years, and that is fine. You have customers, staff, contracts and systems that were not designed around a model, and rebuilding all of it at once is how you end up where Klarna was in 2024. The useful move is smaller and more radical at the same time: pick one process and rebuild it AI-first, end to end, instead of sprinkling copilots over every department.

A worked example. A wholesaler receives about 200 quote requests a week by mail, half of them with a spreadsheet or a photo of a parts list attached. Today two people read every mail, look up prices in the ERP, check stock and type the quote. The AI-enabled version gives those two people a copilot that summarises the mail. Same process, ten minutes saved per quote.

The AI-native version starts from a different question: what would this process look like if a model did the reading? The price list and stock levels get a clean API so the model can look them up itself. The quote template becomes structured data instead of a Word file. A model reads the request, builds the quote and scores its own confidence. Quotes above the confidence bar and under a set amount go out on their own. The rest land in a review queue with the model's draft and its doubts listed. The two people now review 40 quotes a day instead of typing 40, and spend the rest of their day on the customers who phone. Switch the model off and there is no quoting process to fall back on, which is exactly why you keep the review queue, the evals and a way to pause it.

Three things decide whether such a project works: the data behind it is clean enough for a model to read, there is a measurable outcome such as turnaround time or cost per quote, and someone owns the process after go-live.

What to watch out for with the AI-native label

The label as a rebrand. Adding a chatbot to an existing product and calling the company AI-native is the pattern regulators call AI washing. In March 2024 the US securities regulator fined two investment advisers, Delphia and Global Predictions, a combined 400,000 dollars for claiming AI capabilities they did not have. Apply the removal test to any supplier that uses the label: ask what stops working if the models are switched off.

Staff cuts announced before the process was redesigned. Headcount is the last thing that changes in a genuine AI-native rebuild, because you only learn how many reviewers a process needs once it has run for a while. Announcing the cut first and discovering the quality drop afterwards is the sequence Klarna had to walk back in 2025.

The assumption that AI-native means no humans. Every company in this cohort employs people. They employ fewer people per unit of output, in different roles. The reviewers, the people who write and maintain the evals, and the ones who talk to the customer when the model is unsure are the process. Design them in from day one instead of discovering you need them after the first bad month.