Most owners we speak to have already done the first round. A few people use a chat assistant, someone connected a model to the helpdesk, and a pilot summarises documents that nobody asked to be summarised. The tools work. The numbers on the P&L barely move.
The firms pulling ahead appear to do three other things. They redesign the process before they add the model, they build their own harness around it, and they feed corrections back so the system improves with use. Buying a better model does none of these for you.
An AI-native company is one whose operating model would break if you took the AI away, and the evidence points to process redesign, a custom harness and feedback loops as the ways it gets there. This article explains each, shows what the research does and does not support, and gives you a test to run on your own business. If you want to map this onto your firm directly, our AI strategy service starts from exactly that exercise.
What “AI-native” actually means
The term has no standard definition. Consultancies, venture investors and vendors use it for two different things: a company that was founded around AI, and a company that operates around AI. Several 2026 commentaries also warn that the label has turned into marketing.
The most defensible definition in the sources is a removal test. Take the AI out. If the product or the operating model collapses, the company is AI-native. If you only lose features, it is AI-enabled. A support team with a drafting assistant is AI-enabled. A support operation where an agent resolves the routine cases, people handle the exceptions and the staffing model assumes both is closer to AI-native.
Two consequences follow. First, age is not the criterion, so an established firm can qualify. Second, most companies sit somewhere in the middle. Even vendors that publish AI maturity models say most firms are still AI-assisted or AI-enabled.
Why bolt-on AI stalls
Adding a model to an unchanged process automates the existing steps, including the ones that only existed because humans were slow or because systems did not talk to each other. You keep the queue, the handoffs and the approval chain, and you pay for the model on top.
The numbers are consistent with this. In McKinsey’s 2025 State of AI survey, about 39% of respondents attributed any EBIT impact to AI, and most of those said under 5%. A small group of high performers, about 6%, attributed 5% or more. McKinsey reports that workflow redesign is one of the strongest contributors to meaningful impact, and that high performers were nearly three times as likely to have fundamentally redesigned workflows. Secondary coverage of the 2026 edition reports a similar picture, with the share of high performers still at about 6%.
BCG’s September 2025 study of 1,250 senior executives found 5% of firms “future-built” and capturing AI value at scale, 35% scaling and 60% reporting minimal gains. BCG sells transformation work, so read the size of the gap with that in mind. The direction still matches McKinsey.
Swiss SMEs fit the same pattern. The AXA and Sotomo study, with fieldwork in March 2026 across 336 SMEs, found that 74% use or test AI, up from 55% in 2024. Its conclusion was that AI optimises existing processes but does not fundamentally call the business model into question. Adoption is broad. Redesign is rare. The full picture of Swiss adoption is in Swiss SME AI adoption in 2026.
Pillar one: redesign the process first
The strongest experimental evidence on this comes from INSEAD. Researchers ran a randomised field experiment on 515 high-growth startups. Half were shown how other firms had reorganised production around AI. Both groups had access to the same tools. The treated group reported 44% more AI use cases, 12% more tasks completed, an 18% higher chance of acquiring paying customers and revenue about 1.9 times that of the control group. Labour demand did not change.
Read the limits as carefully as the headline. The treatment was exposure to peer examples, so the experiment did not directly test redesign, and what treated firms did afterwards varied. The sample consisted of startups rather than established SMEs. It is the best quantitative support for redesign that exists, and it is still only one study. For the practical side, see mapping a workflow before you automate it. For a closer reading of the study, see how companies reorganise work around AI.
Stanford’s Digital Economy Lab reached a similar qualitative conclusion by a different route. Its April 2026 study of 51 successful deployments across 41 organisations found that two thirds of the companies had significant failed attempts before reaching value, and its foreword states that the difference was never the AI model, it was always the organisation. The authors disclose a selection bias toward positive outcomes. Use the study to learn what winners changed. Do not use it to estimate how likely success is. We read the report in detail in what 51 successful AI deployments changed in the organisation.
In practice, redesign means asking questions a tool rollout skips:
- Which steps exist only because of an old constraint, and can they disappear?
- Where does a person decide, and where do they merely forward?
- Who owns the outcome when the agent acts, and who reviews exceptions?
- What do you measure before and after, so a gain is visible?
Our guide to agentic workflows covers how an agent differs from a scripted automation, which is the technical half of this redesign.
Pillar two: build the harness around the model
A model alone is a text generator. Around it sits the harness: the context it receives, the tools it can call, its memory, the guardrails, the evaluations that check its output and the orchestration between steps. Anthropic describes context engineering as curating the optimal set of information available to the model at each step, and names its costs, including context rot and information lost when conversations are compacted too aggressively. The anatomy of an agent harness goes through the six parts one by one.
There is one clean public data point for how much this layer matters. LangChain reported in February 2026 that its coding agent rose from 52.8% to 66.5% on Terminal-Bench 2.0, moving from roughly the top 30 to the top 5, with the model held fixed. The changes were a build-and-verify loop, a verification step before finishing, loop detection and a schedule for reasoning effort. Treat it with the right weight. It is a single vendor’s result on one benchmark, LangChain does not isolate which change mattered, and the authors expect some of the guardrails to become unnecessary as models improve. We work through the numbers and the shelf life of this scaffolding in harness engineering.
That last point is the live debate. One camp argues this layer, built from your data, tools, rules and tests, is where lasting advantage sits. The other argues that vendors absorb those features into their products and each model release retires some scaffolding. No independent study settles it. A sensible reading is that generic scaffolding will commoditise, while the parts tied to your own data, systems, permissions and quality standards stay yours. This is also why the build versus buy decision for AI agents turns on integration depth and differentiation rather than on the model name.
Pillar three: feedback loops that compound
An AI-native system gets better because corrections flow back into it. At the simple end, a person fixes an output and the fix becomes a rule, an example or a test case in the shared context. Thoughtworks’ Rahul Garg describes capturing context, instruction, workflow and failure signals into shared artefacts for exactly this purpose, while admitting that the effect is hard to measure.
At the advanced end, Shopify Engineering described in August 2026 a daily pipeline for its Sidekick assistant. A rubric-based judge scores live conversations, failures are replayed with repair hints, and the team applies reinforcement learning and supervised fine-tuning, tuning the surrounding system before it touches model weights. The post gives serving-cost estimates, about $27 million a year on a frontier model against roughly $1 million for a fine-tuned one, and labels them as estimates. Whether fine-tuning pays at your own volume is a separate calculation, which we run in the economics of fine-tuning. Quality is judged by the team’s own judge, and rubrics and annotation remain manual. To build such a test set from your own support tickets, see AI evals from real tickets.
Two cautions apply. “Self-learning” in a business setting should mean supervised improvement with a person reviewing what changes, never a system that rewrites itself unobserved. And outcome data for human-correction loops at SME scale does not exist yet. The mechanism is sound and widely recommended. The size of the payoff for a firm your size is unmeasured. For what improves on its own and what still needs people, see self-learning AI systems, and for the Swiss legal side of automated decisions, see automated decisions under FADP Art. 21.
What does not hold up
Several statistics circulate in this debate and need care before you put them in front of a board. Our article on what the headline AI statistics actually measure takes the main ones apart.
- Revenue per employee tables. Figures such as a majority of AI-native startups exceeding $400,000 per employee come from a cloud vendor’s survey, and independent trackers disagree widely with each other for the same company. They do not work as benchmarks.
- The MIT “95%” figure. Its definition and sample are reported inconsistently, and one vendor critic argues the failure rate for deliberate production deployments is closer to 25 to 30%. The report itself describes pilots without measurable P&L impact.
- Gartner’s 40%. The June 2025 forecast that over 40% of agentic AI projects will be cancelled by end-2027 is a prediction rather than a measured rate. The same release warns about vendors that relabel chatbots as agents, which we cover in agent washing.
- Layoffs as proof. Block’s 2026 cut of about 4,000 jobs was framed as AI-enabled, while the CEO also cited pandemic overhiring, and Oxford Economics reported AI as the stated reason for only 4.5% of US job losses in the first eleven months of 2025.
- Developer speedups. METR’s 2025 trial found experienced developers 19% slower. Its 2026 update says newer estimates point the other way and calls them unreliable.
The pessimistic and optimistic numbers mostly measure different groups: enterprises and executive expectations on one side, startups, selected successes and single tasks on the other. They do not contradict each other as much as the headlines suggest.
A test for your own company
You do not need a transformation programme to find out where you stand. Run this on one process.
- Removal test. If you switched the AI off tomorrow, what would break? If the answer is “nothing important”, you have a feature.
- Redesign test. Does the process look different from the one you ran before AI, with fewer steps, different roles or a different owner?
- Harness test. Do you own the context, tools, permissions and evaluations, or does a vendor’s default hold all of them?
- Loop test. When a person corrects an output, does the correction change the next one?
- Baseline test. Did you record cost, time and error rate before you started?
Most firms will fail the first one or two on most processes, which is normal. The aim is to pick the process where passing all five is worth the most, then expand. The ROI of AI agents is simpler to calculate on a redesigned process because the baseline and the new flow are both explicit.
Where to start
Start narrow. One process with real volume, a named owner, a measured baseline and a human review step gives you evidence from your own business within a quarter, which is worth more than any survey above. Reuse what you build: the context and tooling, the evaluation set and the correction habit transfer to the next process even when the workflow does not.
The sources agree on the direction and disagree on the size of the prize. That is a good reason to run a small, measured experiment instead of a large, assumed one.
Frequently asked questions
What is an AI-native company?
There is no agreed definition, but the most usable test is structural. If you removed the AI, an AI-native company would lose its operating model, while an AI-enabled company would only lose some features. Company age matters less than whether processes, roles and measurement were built around AI doing real work.
Can an established Swiss SME become AI-native?
Yes, though it happens process by process rather than through a company-wide switch. Pick one workflow with clear volume and cost, redesign it around an agent with human review, measure it against a baseline, and repeat. Most evidence on full redesign comes from startups and large enterprises, so treat SME results as your own to measure.
What is an AI harness and why does it matter?
A harness is everything around the model: the context it receives, the tools it may call, its memory, guardrails, evaluations and the orchestration between steps. It matters because it determines whether results are repeatable. Published vendor results show harness changes alone can shift benchmark scores a long way, though no independent study measures how durable that advantage is.
Are the statistics saying 95% of AI projects fail reliable?
They need context before anyone quotes them. The widely repeated MIT figure is contested on definition and sample, and Gartner's forecast that over 40% of agentic AI projects will be cancelled by end-2027 is a prediction rather than a measured rate. Both describe enterprises. Budget against your own baseline for one process instead.
Do AI-native companies need fewer staff?
The evidence is mixed. In the INSEAD field experiment, treated startups did not change labour demand, and executives surveyed by NBER report essentially no net employment effect so far. Block cut roughly 4,000 jobs in 2026 and credited AI partly, but also cited earlier overhiring. Treat headcount claims as unproven.