Skip to content
Chat with AI Agent
Business & governance

How Companies Reorganise Work Around AI: What 515 Startups Gained from Seeing Peers Do It

Orange ITS — AI engineering team 9 min read

Picture the owner of a 45-person engineering firm who has paid for AI licences for a year. Drafting is quicker, a few people use the tools daily, and nothing about the way work moves through the firm has changed. The owner suspects there is a bigger version of this, and has no picture of what it looks like in a company of similar size.

That missing picture is the problem, and it is cheaper to fix than most owners assume. If you are weighing where AI belongs in your operations, our AI strategy work begins by looking at how comparable firms organise their work before anyone chooses a tool.

The most useful input to an AI operating model redesign may be a well-chosen example of how a peer reorganised its work, and a field experiment with 515 startups is one of the few randomised tests of why. This article covers what that experiment did, what it cannot tell an SME, and a practical way to collect and read peer examples in Switzerland.

Scope note: for the argument on why a bolted-on tool plateaus, see Why Bolting AI onto Old Processes Plateaus. For the mapping exercise itself, see AI workflow mapping.


What the INSEAD Experiment Actually Did

The study is a working paper by Hyunjin Kim, Dahyeon Kim and Rembrand Koning, titled “Mapping AI into Production: A Field Experiment on Firm Performance” and listed by INSEAD as working paper 2026/20/STR. According to the INSEAD abstract, 515 high-growth startups took part.

The authors frame the central difficulty as a mapping problem: finding where AI creates value in a firm’s production is a key bottleneck. Their treatment targeted that bottleneck directly. Treated firms learned how other firms had reorganised production around AI, and that prompted them to search for use cases across more functions.

The design, according to the HBS AI Institute summary

The design details below come from the HBS AI Institute summary of the study rather than from the full paper. According to that summary, both groups received the same API credits, access to frontier models and technical training. The only difference was what else each group saw.

  • Treatment group: case studies of AI-native firms reorganising their workflows, teams and business models.
  • Control group: workshops on general entrepreneurship.

The same summary describes a programme of about 10 weeks.

Why does this design matter? Because the control group was not left without AI. Everyone had the tools and the training. What differed was exposure to examples of reorganised work, which is exactly the variable an owner can try to reproduce.


What Changed for the Firms That Saw the Examples

The INSEAD abstract reports the following differences between treated and control firms:

OutcomeReported difference
AI use cases identified44% more, concentrated in product development and strategy
Tasks completed12% more
Likelihood of acquiring paying customers18% higher
Revenue1.9 times higher
Demand for external capital39.5% lower
Labour demandUnchanged

Two details in this table deserve more weight than the headline numbers. First, the extra use cases appeared in product development and strategy, which are functions where a founder can reshape the work rather than only speed up a task. Second, labour demand did not change. The firms did not hire fewer people. They got more from the same headcount, which is a milder claim than the “replace staff” framing that dominates AI headlines.

The effect sits in the upper tail

According to the abstract, the revenue and investment gains were largest at the 90th percentile and above. In plain terms, the average hides a spread. A small number of firms moved a long way, and many moved little.

That matches a pattern worth keeping in mind. Seeing an example gives a firm an idea. Whether the idea turns into revenue depends on the team, the market and the follow-through. For how to read distributions and definitions in AI evidence generally, see Six AI Statistics Your Board Quotes, which applies the same questions (who was measured, and how success was defined) to the better-known headlines.


What the Trial Does Not Show an SME Owner

The study is careful work, and it still answers a narrower question than the one an owner is asking. Five limits matter.

  1. The sample is startups. High-growth startups have little organisational inertia. They have no legacy ERP, no works council, no 20-year-old approval chain. A Swiss SME has all of these.
  2. The treatment was exposure. The experiment tested what happens when firms see how others reorganised. It did not test whether the firms themselves carried out a redesign, nor whether a redesign works.
  3. The bundle cannot be unpacked. The HBS AI Institute summary describes the cases as one treatment, so the study cannot say which case or which idea drove the effect.
  4. The context is short and structured. A programme of roughly 10 weeks is a long way from a three-year operational change in a manufacturing firm.
  5. The gains are concentrated. Revenue and investment effects were largest at the 90th percentile and above. An owner should not expect the average outcome to be the typical one.

Nothing in the trial shows that an established Swiss SME would see similar gains. What it supports is a modest, testable claim: learning how peers organised their work changes what a firm looks for. That is an argument for gathering examples early, and a reason to keep judging the outcome by your own before and after numbers.


Seven Ways a Swiss SME Can Get Peer Examples

None of these require a consultancy engagement. They differ in cost, in how comparable the peers are, and in how honest the answers tend to be.

Programmes and funding

The Canton of Zurich programme. The Canton of Zurich ran an AI innovation programme for SMEs together with the Swiss Data Science Center, over one year starting in July 2025. According to the announcement on ahead-zh.ch, it combined peer-to-peer learning with prototyping, and it was open to SMEs headquartered in the canton or with significant operations there. A next edition has not been confirmed, so ask the organiser before planning around it.

The Innosuisse innovation cheque. The cheque is worth up to CHF 15,000 and is open to SMEs with fewer than 250 FTE. It funds feasibility studies, it requires a Swiss research partner, and Innosuisse pays the research partner rather than your firm. That makes it a way to buy a structured, outside look at one process. It is not a window into other companies, and it needs a clear question before you apply. Check the current conditions on the Innosuisse page.

Networks you may already have

Industry associations. Trade bodies run working groups and conferences where members compare notes. The useful sessions are the small ones where someone describes a process that changed, with enough detail to be questioned.

Supplier reference customers. When a software or AI supplier proposes a solution, ask for two customers of your size, in a similar sector, who have run it for at least six months. Then speak to them without the supplier in the call. A supplier who cannot name such customers has told you something.

Owner round tables. Five or six owners from non-competing firms meeting quarterly, each bringing one process that changed or failed to change, is cheap and effective. The rule that keeps it honest is to share what went wrong as well as what worked.

Structured visits. An hour in a peer’s office or workshop, watching how a handoff actually happens, tells you more than a slide deck. Send questions in advance so that the visit does not turn into a tour.

Your own firm

A one-page before and after case. Write down one change you have already made, even a small one, in the format shown below. It forces you to find out what really changed, and it gives the rest of your team a concrete example from a company they recognise. Our AI workflow mapping guide describes how to build the map that a good case starts from.


How to Read an Example: Ask About Roles, Handoffs and Decision Rights

Most owners ask a peer the wrong first question: which tool did you use? The answer ages within months and says little about whether the approach transfers to your firm. A better first question is what changed in the way work is divided between people and software.

Four questions travel well:

  • Roles. Who does the work now, and what did that person do before? Did a job description change, or only the toolset?
  • Handoffs. Where did a case used to wait, and where does it wait today? Which handoff disappeared, and which new one appeared?
  • Decision rights. Which approvals moved, shrank or vanished? Who can now decide something that used to need a manager?
  • Measures. What did the team count before the change, and what did they count afterwards? Cycle time, error rate and cost per case are the usual candidates.

Then ask the question that keeps the exercise honest: what would you do differently if you started again? Peers who have been through a redesign usually have a clear answer, and it is often the most valuable minute of the conversation.

A one-page template for your own case

If you write your own before and after page, keep it short enough that colleagues will read it. A format that works:

FieldWhat to write
ProcessThe name of the process and the team that owns it
BeforeSteps, roles, where cases waited, one baseline number
What changedRoles, handoffs, decision rights, data at entry
AfterThe same baseline number, measured the same way
What still needs a personThe judgement steps that stayed human
What we would changeOne honest correction

Written once for each change, these pages become a small internal library. That is worth having for a practical reason. Owners tend to trust examples from their own building more than anything from a conference stage, and a library of five or six pages gives a new project a starting point that no external study can supply.


Where Peer Examples Mislead

Examples are a cheap input, and cheap inputs have failure modes.

Survivorship. Peers tend to share what worked. Ask explicitly about the pilot that was dropped. Any set of cases is a curated sample.

Context mismatch. A change that works in a 12-person firm with one owner-manager may fail in a 90-person firm where approvals run through three departments. Ask what in the peer’s setup made the change possible, such as a short chain of command, clean data or a single system of record.

Copying the surface. Adopting a peer’s tool without the change in roles and handoffs reproduces the plateau described in our piece on bolt-on AI. The example is useful for what it changed in the work.

Reading too much into the numbers. A peer’s figure of hours saved is self-reported and unaudited. Treat it as a pointer to where to look in your own data.


What the Evidence Supports

The INSEAD trial supports a narrow claim and a practical habit. The claim is that startups which saw how others reorganised work around AI went looking for more places to use it, and the best of them reported markedly better results. The habit is to collect concrete examples of reorganised work early, from peers you can question, before committing budget to a tool or a programme.

For an SME, the first step costs an afternoon: write one before and after page from your own firm, then ask two peers for theirs. If the comparison shows that no one around you has changed roles, handoffs or decision rights in any process, that is useful information too. It tells you the question of AI operating model redesign is still open in your sector, and that an early, measured change can be a real advantage.

Frequently asked questions

What did the INSEAD AI field experiment test?

It tested whether showing startups how other firms had reorganised production around AI changes what they do with AI. In the trial of 515 high-growth startups, treated firms found 44% more AI use cases, completed 12% more tasks and were 18% more likely to win paying customers. The treatment was exposure to peer examples. The trial did not test a redesign method.

Does the INSEAD study prove that reorganising work around AI pays off?

No. It shows that learning how other firms reorganised work led startups to search for more AI use cases and to report better outcomes, mainly among the top performers. It does not isolate redesign itself, and the sample was high-growth startups. For an established SME, treat it as a reason to look at peer examples early, and verify any gain in your own numbers.

Where can a Swiss SME find examples of AI redesign at peer companies?

Start with supplier reference customers, industry associations and owner round tables, where peers share what changed in their daily work. Formal routes exist too. The Canton of Zurich ran an SME AI programme with the Swiss Data Science Center from July 2025, and the Innosuisse cheque funds a feasibility study with a research partner. Check each organiser for the current edition and conditions.

What should I ask a peer company about its AI project?

Ask what changed in roles, handoffs and decision rights, rather than which tool they bought. Useful questions are who does the work now, which approvals moved or disappeared, what data the process needs at entry, and what the team measured before and after. Tool names age within months, while the way work was divided between people and software tends to transfer.

Is the Innosuisse innovation cheque a way to learn from peers?

Only indirectly. The cheque is worth up to CHF 15,000 for SMEs under 250 FTE and funds a feasibility study carried out by a Swiss research partner, which Innosuisse pays directly. It gives you outside expertise and a structured look at your own process rather than a view into other companies. Confirm the current conditions with Innosuisse before planning around it.

Insights

Put these ideas to work

A 30-minute call is enough to find out whether an AI agent fits your workflow — and what it would return.