In short
An AI-native company is designed around AI in two places at once: in how its people work, and in what it sells. The AI is not a feature added to an existing process. It is the process. If you took the AI out, the company would not go back to an older way of working. It would stop working.
The phrase is used loosely. Many companies that call themselves AI-native are AI-enabled: they bought tools and added them to work that was designed for people without those tools. The difference is not the amount of AI in the building. It is whether the work was rebuilt around it.
This article gives a definition you can test and the research that measures it. It then gives two public examples of companies rewriting their operating rules, and the reasons most attempts fall short. xvertical does not work for any of the companies named here. They are public cases a reader can check.
How does AI shape an AI-native company? Two channels
An AI-native company runs AI through two channels: a process channel, where AI changes how people work, and a product channel, where AI is built into what the firm sells. That is the frame of a June 2026 working paper, AI-Native Firms, by Hyunjin Kim of INSEAD and Rembrand Koning of Harvard Business School. They study “firms built around AI capabilities” and describe “a process channel, in which AI changes how people work inside the firm, and a product channel, in which AI capabilities are built into what the firm sells.”
The process channel is AI in the operating model. The paper’s examples are AI agents that write code, AI-assisted customer service and automated sales operations. The product channel is AI in the thing a customer pays for. The paper’s examples are products that let customers generate slide decks, build software or run their accounts receivable.
In our view, the process channel comes first for an operating business. A hospitality group sells meals. A publisher sells magazines. A manufacturer sells parts. The product does not change. What changes is how the functions behind it run: how a booking becomes a forecast, how a support call becomes a resolved ticket, how a quality check becomes a decision.
How do you tell AI-native from AI-enabled? The removal test
The removal test asks one question: if you remove the AI, does the product stop working? The venture firm CRV put it that way in its guide to AI-native companies of March 2026, where AI “isn’t a feature or an enhancement” but “the architectural foundation on which the entire product depends.” CRV’s wording of the question is: “if you remove the AI, does the product cease to function?”
The same test works on a function instead of a product. Take sales. If the AI disappeared tomorrow, would the team go back to the spreadsheet, or is there no spreadsheet left to go back to? If the first, the function is AI-enabled. If the second, it is AI-native.
Three signs tell the two apart in practice.
- There is no manual version of the process waiting in a drawer. The AI-native version is the only version.
- The data the function runs on is collected for the AI to use, not exported from one tool into another for a weekly report.
- Roles are defined by what a person decides and checks, not by what a person types.
What the research measures
AI-native startups are 25% smaller than non-AI startups in the same industry and cohort, according to the Kim and Koning paper. The paper draws on Y Combinator batches from winter 2020 to fall 2024 and on US venture-backed startups whose first financing closed between 2020 and 2024, and links each firm’s AI-native status to workforce data on team size, function, seniority and hierarchy.
The abstract gives the numbers: “Relative to non-AI startups in the same industry-cohort, AI-native firms are 25% smaller. Their share of engineers is 13% greater, and the shares of entry-level workers and managers are each roughly 15% lower. Their hierarchies are half a seniority level flatter—yet valuations are comparable, suggesting higher value created per employee.”
Three caveats. The sample is startups, which carry no legacy processes. The paper is a working paper in draft form, not a peer-reviewed publication. And a comparable valuation is not a comparable profit.
What the paper does show is the shape of the organisation. Fewer layers. Fewer people doing entry-level work. More people building. The authors’ reading is that embedding AI into the product, “beyond layering on AI tools into existing workflows,” is “a primary way” these startups scale knowledge work without large teams of knowledge workers. Layering tools on existing workflows is the AI-enabled pattern. Rebuilding the workflow is the AI-native one.
What does the process channel look like in an existing company?
Two public memos from 2025 show the process channel inside companies that already existed. Neither announces a tool. Both change the rules by which the company decides who does what.
In April 2025, Shopify’s CEO Tobi Lütke posted an internal memo publicly to pre-empt a leak. As reported by Workplace Journal, it made reflexive AI use a “baseline expectation” for staff, a “fundamental expectation of everyone at Shopify.” AI use became part of performance and peer reviews. “Learning to use AI well is an unobvious skill,” Lütke wrote, and nobody should “opt out of learning the skill of applying AI in your own craft.”
The memo’s hiring rule is the part that changes the operating model. Before asking for more headcount and resources, “teams must demonstrate why they cannot get what they want done using AI.”
Later the same month, Duolingo’s CEO Luis von Ahn posted a company memo on LinkedIn, covered by Pure AI. “We’re making a similar call now, and this time the platform shift is AI,” he wrote. The company would gradually stop using contractors for work that AI can handle, and new headcount would be approved only where a team could not automate more of its work.
The line that matters for this article is about process: “Making minor tweaks to systems designed for humans won’t get us there. In many cases, we’ll need to start from scratch.” The memo drew public criticism. Von Ahn wrote that the change was “about removing bottlenecks,” not about replacing staff.
Agree or disagree with either memo, notice what they change: the headcount rule, the review rule, and the instruction to start from scratch rather than tweak. That is the process channel at company scale.
Why most companies stay AI-enabled
Most companies stay AI-enabled because the pilot is laid on top of an unchanged process, and the people and data it needs are not there. Two widely reported studies from 2025 and 2026 measure that.
In August 2025, MIT’s NANDA initiative published a report on the state of AI in business, summarised by Fortune. It drew on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments. It found that “about 5% of AI pilot programs achieve rapid revenue acceleration,” with the rest “delivering little to no measurable impact on P&L,” and that “for 95% of companies in the dataset, generative AI implementation is falling short.” The core problem, in Fortune’s summary, is a “learning gap” on both sides, not the quality of the models. Generic tools do not adapt to the way a particular company works.
In April 2026, Gartner published a survey of 782 infrastructure and operations leaders, conducted in November and December 2025 and reported by The Register. Gartner found that only 28% of AI use cases fully succeeded and returned on the investment, that one in five failed outright, and that 57% of the leaders had seen at least one failure. Gartner’s research director Melanie Freeze gave the reason: “The 20 percent failure rate is largely driven by AI initiatives that are either overly ambitious or poorly scoped.” Among those who had seen failures, 38% said persistent skill gaps still hamper their AI work, and 38% named poor data quality or limited data availability as a direct cause of failure.
Read the two studies together and the pattern is the AI-enabled one. A pilot is laid on top of an unchanged process. The people who run the process have not learned the tool, and the data the tool needs sits in exports and spreadsheets. The pilot cannot change the process, so it cannot deliver, so it is cancelled. Skills and data are function-level problems. They can be solved inside a function, by the people who run it, with the workflow rebuilt around what the AI can now do.
AI-native and vertical integration
Our view is that an AI-native function is usually an owned function. Vertical integration is the decision to own a stage of the work instead of buying it. When the AI sits inside the workflow and the data of a function, the company has to own that workflow and that data. A supplier’s tool with the company’s data exported into it is the AI-enabled pattern again, under a different name.
Owning the function does not mean building everything. An AI-native function buys the model and owns the workflow, the data and the decisions. Which capabilities to own and which to keep buying is its own question. The cost of ownership has fallen, which is why it is now within reach of mid-sized companies and not only of the startups in the Kim and Koning sample.
How does an existing company become AI-native?
An existing company becomes AI-native one function at a time. The companies in the research were born this way. An operating business has to get there in steps, and the order follows from the failures above.
- List the functions and mark the ones the business depends on. Sales, service, operations, finance, whichever ones a bad week shows up in.
- Pick one function with an outcome you can measure before and after. A support line has a resolution rate. A reporting process has a number of hours.
- Rebuild how the work is done with the people who do it. The AI is not added to the process. The process is redrawn around what the AI can now carry, and the people decide and check.
- Write the rules the memos wrote: who approves what, when a person steps in, how the team’s use of the AI is reviewed.
- Measure, then move to the next function.
A chatbot on an old process is AI-enabled. A support line that no longer has a manual version is AI-native. The second one is harder, and it is the one that changes the company.
Key takeaways
- An AI-native company is designed around AI in how it works and in what it sells. If you remove the AI, it stops working rather than reverting to an older process.
- AI-enabled companies add AI to processes that would still run without it. The removal test tells them apart, for a product and for a single function.
- A 2026 Harvard Business School and INSEAD working paper found AI-native startups 25% smaller, with 13% more engineers, about 15% fewer entry-level staff and managers, flatter hierarchies and comparable valuations.
- Shopify’s and Duolingo’s 2025 memos changed hiring, review and design rules, not just tools. That is what the process channel looks like inside an existing company.
- MIT found that for 95% of companies, generative AI pilots delivered little to no measurable profit effect. Gartner found 28% of AI use cases fully succeeded. The causes are scope, skills and data, all of which can be solved inside a function.
- An AI-native function is usually an owned function. Buy the model, own the workflow and the data.
Sources
Each quote above is linked where it appears. This is the full list.
- Hyunjin Kim and Rembrand Koning, “AI-Native Firms”, Harvard Business School Working Paper 26-090, 9 June 2026 (PDF). The definition, the two channels, the sample and the figures on size, hierarchy and valuation.
- CRV, “What Is AI-Native? The Founder’s Guide (2026)”, 31 March 2026. The removal test and the distinction from AI-enabled products.
- Jessica Bird, Workplace Journal, “Shopify CEO calls for reflexive AI use as ‘baseline expectation,’ no headcount growth until options explored”, 8 April 2025. The lines quoted from Tobi Lütke’s memo.
- John K. Waters, Pure AI, “Duolingo Leans Into AI, Phases Out Contract Roles in Push for Scalable Learning”, 29 April 2025. The lines quoted from Luis von Ahn’s memo, which he posted on LinkedIn.
- Sheryl Estrada, Fortune, “MIT report: 95% of generative AI pilots at companies are failing”, 18 August 2025. The MIT NANDA figures, method and the “learning gap”.
- Lindsay Clark, The Register, “Only 28% of AI infrastructure projects fully pay off, survey finds”, 7 April 2026. The Gartner survey of 782 infrastructure and operations leaders, with the causes of failure and the quote from Melanie Freeze. The underlying release is Gartner’s press release of 7 April 2026.
Questions this post answers
What is an AI-native company?
An AI-native company is one whose work and product are designed around AI from the start, rather than one that added AI tools to the way it already operated. Take the AI away and the company does not go back to an older way of working. It stops working.
What is the difference between AI-native and AI-enabled?
An AI-enabled company has added AI features or tools to processes that existed before and would still run without them. An AI-native company has rebuilt the process or the product around AI, so there is no manual version left to fall back on.
Are AI-native companies smaller?
In a June 2026 working paper from Harvard Business School and INSEAD, AI-native startups were 25% smaller than non-AI startups of the same industry and cohort, with a larger share of engineers, fewer entry-level staff and managers, and comparable valuations.
Can an existing company become AI-native?
Yes, one function at a time. It starts with the functions the business depends on, rebuilds how that work is done together with the people who do it, and measures the result before moving to the next function. Adding a chatbot to an unchanged process does not count.
Why do most AI projects not deliver?
Gartner's 2026 survey of 782 infrastructure and operations leaders found that 28% of AI use cases fully succeeded and paid back, while 20% failed outright, mostly because they were over-ambitious or poorly scoped. MIT's 2025 study found that for 95% of companies, generative AI pilots delivered little to no measurable effect on profit.