A joiner in Winterthur with 25 staff (an illustrative case) signs up for an AI quoting tool. Estimators draft offers a little faster, the owner mentions it at the next industry event, and the website now says “AI-powered”. Six months later the quoting process looks the same as before: same steps, same people, same margin. Meanwhile a competitor two valleys over rebuilt order intake around a model that reads drawings, flags missing dimensions and proposes a price band. Take that model away and the competitor’s intake stops.
Those two firms would both be called “AI companies” in a sales deck. They are very different businesses, and the gap shows up in a question you can ask yourself this week. If removing the AI breaks how a process runs, the AI may be producing margin there. If removing it only costs you a few features, it is a cost line. This article gives you a six-step way to test that on your own processes, and it is honest about what the evidence does and does not prove. If you want to place the result inside a wider plan, our AI strategy work starts from the same process-first view.
Treat this as a diagnostic to try. It is unproven: no study yet measures margin by removal-test result, and a process that breaks without AI can be a strength or a weakness.
What AI-native and AI-enabled Actually Mean
Nobody owns these terms. No standards body or analyst firm has published a settled definition, and most of the writing comes from vendors and venture investors who benefit from the label. That is worth knowing before you read any ranking of “AI-native companies”.
The usable distinction goes like this. An AI-native company is built around models as the core of its product or delivery model. An AI-enabled company is a conventional business that has added AI features to existing products or workflows. The venture firm CRV puts the test in plain words in its AI Native vs. AI Wrapper piece: deleting the AI layer from an AI-native product stops it cold, while deleting it from software with AI features removes a nice feature and the rest keeps working.
For a Swiss SME the label matters less than the question underneath it. You rarely sell a product built from scratch around a model. You run quoting, dispatch, invoicing and support, and some of those may or may not depend on AI today. The test works at that level, process by process.
What the Evidence Supports
The thesis is attractive, so it needs a sober look. Three findings point in its direction, and none of them proves it.
BCG’s September 2025 survey of 1,250 senior executives put 5% of companies in a “future-built” group, 35% in a scaling group and 60% in a laggard group reporting minimal revenue and cost gains. BCG reports that the future-built group achieves 1.7x the revenue growth and a 1.6x EBIT margin multiple compared with laggards. Two cautions apply. BCG does not use the term “AI-native” anywhere in that comparison. And the groups are defined partly by the value firms report from AI, so some of the gap is built into the grouping.
McKinsey’s 2025 State of AI survey reports that high performers are nearly three times as likely as others to have fundamentally redesigned their workflows. That is self-reported and correlational too, and we quote it from secondary coverage of the report.
Taken together, the sources support a modest claim: redesigning how work flows around AI is associated with better financial results than adding AI to unchanged work. They do not show that “breaks when removed” produces margin. The removal test is our way of spotting whether a redesign has happened, and that link is a hypothesis.
The Six-Step Removal Test for an SME
You need a whiteboard, the person who runs each process, and about half a day. Skip the tool inventory at the start. Begin with the work.
- List three processes that drive revenue or cost. Typical candidates are quoting, order intake, customer support, dispatch and invoicing. Pick the ones where a mistake or delay costs real money.
- Run a switch-off thought experiment for each. Ask: “The AI service is unavailable for two weeks. What happens?” Score the answer using the table in the next section.
- Ask where the AI sits. Does it make a decision that sets price, promise or priority, or does it speed up a task a person would do anyway? A model that proposes a delivery date customers rely on is in a different position from one that tidies email drafts.
- Check whether the process changed. Draw the steps before and after. Who does what now, and what no longer happens by hand? If the answer is “the same steps, faster”, the evidence points to AI-enabled. For the practical side, see mapping a workflow before you automate it.
- Measure cycle time and error rate before and after. Track them over at least two quarters, then connect them to cost or margin per order or job. Licence usage tells you who logs in. It says little about the business.
- Add the dependence check. What would a vendor price doubling or a week of downtime cost you? Which data leaves Switzerland or the EU, and under which contract? Compliance details belong in a separate review.
Step 5 needs a baseline. Many SMEs will discover they never measured the old process, which is a finding in itself, and a useful one.
Scoring Your Processes
Score each process on the switch-off question. The rows below are illustrative examples for a mid-sized trade or service firm. Replace them with your own.
| Process | Switch-off result | Score | What it suggests |
|---|---|---|---|
| Order intake from drawings | Intake queue stops, no one can read drawings at that volume | (a) Stops | Possible redesign. Check dependence |
| Customer support triage | Tickets are sorted by hand, answers take longer | (b) Slower, manual fallback | AI-enabled with real value. Measure time saved |
| Marketing copy drafts | Staff write the drafts themselves | (c) Loses a nice-to-have | A cost line. Keep it only if the saving is visible |
| Invoice coding | Bookkeeper codes by hand, two extra days a month | (b) Slower, manual fallback | Candidate for deeper integration or for dropping |
| Delivery date promise | Dispatcher estimates from experience, accuracy drops | (b) or (a) depending on how customers rely on it | Look at where the decision sits |
Read the (a) rows as the places where AI may already be producing margin, and the (c) rows as the places to question the licence. Most firms find more (c) than they expected. That is useful, because it shows where to stop spending before you decide where to spend more.
One warning on honesty. Owners tend to answer “it would break” about anything that feels essential now. Use the process map from step 4 to check the answer, because gut feel overrates how much has changed.
The Dependence Check: When Breaking Is Bad News
An (a) score is not a trophy. It means the process now relies on something outside your control.
Think through four exposures:
- Price. Model and API pricing has moved quickly. A process whose unit economics only work at today’s price is a bet on that price.
- Availability. Hosted AI services have outages. If your order intake stops for a day, what is the cost per hour?
- Data. Know which provider sees which customer data, under which contract and in which jurisdiction.
- Fallback. Can staff still run the process manually, even slowly, for a week?
A sound answer is a process that breaks without AI and has a documented fallback, a vendor you could replace and a cost you have modelled at double the current price. A fragile answer is a process that breaks, has no fallback and depends on one vendor’s goodwill. Both score (a). Only one of them is a business asset.
Where Swiss SMEs Stand Today
The best local data comes from the 2026 KMU-Arbeitsmarktstudie that Sotomo ran for AXA Switzerland, based on 336 SMEs in German- and French-speaking Switzerland surveyed in March 2026. It found that 35% deliberately integrate AI into their business processes, 39% are testing the technology and 26% do not use AI yet.
AXA’s own press release leads with “three in four Swiss SMEs use AI”. That sum counts testing as use. A fairer reading is that roughly a third have started embedding AI in processes and about as many are still experimenting.
On the business model, 31% of respondents said they expect to adapt it because of AI, and almost two thirds see little or no need. Kathrin Braunwarth, head of digital transformation and innovation at AXA Switzerland, summed it up as AI optimising existing processes without fundamentally challenging the business model. That is the executive’s reading of the survey, hedged with “apparently” in the original, rather than a measured finding. The full picture of Swiss adoption is in Swiss SME AI adoption in 2026.
For the removal test, this means most owners will score their processes (b) and (c) today. That is a normal starting point.
Reading the Result Without Falling for the Label
A word on vendors. Gartner warned in 2025 about “agent washing”, the relabelling of ordinary chatbots and automation as autonomous agents. Gartner’s press release of 25 June 2025 estimates that only about 130 of the thousands of agentic AI vendors are real. Staffing Industry Analysts reported the same point, along with the claim that most agentic propositions lack significant value or return.
The same caution applies to “AI-native”. A tool that calls a model through an API and sells itself as AI-native has not earned the word by that fact alone. Apply the removal test to the vendor’s product as well: what remains if its model provider revokes access tomorrow?
Here is what to do with your scores:
- Mostly (c): you are AI-enabled. Audit the licences, keep what shows a measured saving and stop paying for the rest.
- A few (b): pick the process with the largest cost or delay and measure it for two quarters before changing anything.
- Any (a) with no fallback: write the fallback first, then consider going deeper.
The aim is to know which two or three processes could carry that status, and to confirm with numbers before you commit. For a worked comparison of the building blocks behind such processes, see AI agent vs assistant vs copilot. To check whether your data and teams are ready for a redesign, the AI readiness assessment is the next step. And if part of your process is rule-based work, AI agents vs RPA shows where classic automation is enough.
Frequently asked questions
What is the difference between AI-native and AI-enabled?
AI-native means the product or operating model stops working if you remove the AI, while AI-enabled means a conventional business has AI features that could be switched off without breaking the core. There is no standards-body definition, so treat the labels as a diagnostic question rather than a certification.
Is an AI-native company more profitable than an AI-enabled one?
No study has measured that. BCG links higher EBIT margins to companies that redesign how they work around AI, but its groups are defined partly by capabilities and partly by the value companies report, and the data is correlational. Dependence on AI can lower costs and it can also add cost exposure, so profitability remains an open question.
How do I run the removal test in a small company?
List three processes that drive revenue or cost, ask what happens to each if the AI service is unavailable for two weeks, and score the answer as stops, slower with a manual fallback, or loses a nice-to-have. Then compare the process before and after AI and measure cycle time and error rate over two quarters.
Does a process that breaks without AI mean I am doing well?
Not automatically. Breaking shows dependence, and dependence can mean a real redesign or simply fragility. Check what a vendor price doubling or a week-long outage would cost you, and where your data goes under which contract. A fallback plan is part of a healthy answer.
Should a Swiss SME aim to become AI-native?
Only where a redesigned process earns it. Most SMEs get more from making two or three core processes depend on AI deliberately than from adopting the label for the whole company. Start with the process where a decision on price, promise or priority is currently slow, and test it before widening the scope.