Moat Atlas / AI Era
English editionVI
⚡ Strategy report · 2026

When AI can do the engineering,
you win on how you run it.

Two people today do what twenty people did yesterday. The things that used to win — a long-standing brand, sales channels, customers who hate switching — still have value, but anyone can get them now.

This piece answers four simple questions: what is losing value, what still holds, what only exists because of AI, and where a small company should start.

5 things
Losing value
because anyone can do them
8 things
Still hold
even with AI around
6 things
Brand new
they exist only with AI
6 steps
To start now
for a small company
Part 1

Five things losing value

Still useful, but no longer what sets you apart — because with enough time and money, anyone can get there.

01↓ Falling

A long-standing brand

A name built over years, customers who know you. Still worth something — but a startup with a good product now earns trust far faster.

02↓ Falling

Lots of users

The more users, the better the product. Still true — but AI is useful on day one, without waiting to gather a crowd.

03↓ Falling

Sales channels

Resellers, relationships, your own distribution. Still matters — but AI is taking over most of the work of finding and serving customers.

04↓ Falling

Customers hate switching

Changing vendors is a pain, so people stay. Still real — but AI has made moving data and connecting systems much easier than before.

05↓↓ Going fast

A strong engineering team

Losing value fastest. AI writing code lets two people do what used to take a team of twenty.

Part 2

Eight things that still hold

Not everything old is dead. AI makes these eight stronger rather than wiping them out.

How to measure · An advantage is durable only if all four answers are yes
01Will customers pay because of it?
02Is it rare, or does everyone have it?
03Could a rival copy it within a year or two?
04Is there something else that replaces it?

Fail one question and it is only a temporary edge. The durability label on each card below shows how well it answers all four.

These four questions are a plain-language version of two classic frameworks: VRIN (Barney, 1991) and 7 Powers (Helmer, 2016).
01Durability: very high

Your own data

Proprietary data

Data gathered over years of real work, not downloadable from the internet. What matters is clean data labelled correctly by people who know the field — not sheer volume.

Why it still holdsThe better AI gets, the more your own data is worth. Public models lift everyone equally, so they help no one win in particular.
ExamplesBloomberg (30 years of price data) · Epic (hospital medical records).
02Durability: fairly high

Servers and chips

Compute

Owning chips, data centers and the underlying models. Needs enormous capital and only pays off at very large scale.

Why it still holdsThe capital barrier is so high that almost nobody can push in. But this is not a small company's game.
ExamplesNVIDIA · cloud providers AWS, Azure, Google Cloud.
03Durability: fairly high

Customers already paying you

Distribution

You already have the customers. Launch a new AI product and selling it to them costs almost nothing in acquisition.

Why it still holdsAI speeds up selling, but it does not conjure a base of people who already trust you and already pay.
ExamplesMicrosoft bundling Copilot with Office 365 · Apple putting AI straight into the iPhone.
04Durability: fairly high

Customers hate switching

Switching cost

Moving to a rival means migrating data, reconnecting systems, retraining staff. Expensive, slow, and risky.

Why it still holdsAI makes the technical part easier, but business risk and the effort of retraining people do not shrink with it.
ExamplesSAP · Salesforce · ServiceNow.
05Durability: very high
strongest for small firms

Deep knowledge of one industry

Domain expertise

The things only years in the trade teach you: the exceptions, the unwritten rules, how to handle the hard cases. This kind of knowledge is hard to write down, which makes it hard to copy.

Why it still holdsAI can answer the question, but knowing which question to ask still belongs to the person in the trade. This is the one place a small company beats a big one.
ExamplesHarvey (law) · Hippocratic (healthcare) · Cadence (chip design).
06Durability: very high

Licences and certifications

Compliance

SOC 2, ISO, the EU AI Act, medical data standards. You cannot sell to large customers without them, and getting them takes years.

Why it still holdsTime is the one thing AI cannot compress. An audit period still runs its full months no matter how fast you ship.
ExamplesPalantir (defence) · Epic (healthcare) · licensed fintech companies.
07Durability: fairly high

Being trusted when things go wrong

Brand & trust

Customers choose you because they believe you will take responsibility when something breaks. It matters more now that AI can be confidently wrong or leak data.

Why it still holdsAI can write anything, so the question "who answers for it when it is wrong" is worth more every year.
ExamplesThe big audit firms rolling out AI · IBM in enterprise.
08Durability: very high

The more users, the stronger

Network effects

More users make the product more valuable. With AI it runs three ways: tool builders ↔ the platform ↔ end users.

Why it still holdsAI is useful on day one without a crowd. But by year three, whoever has the crowd keeps the market.
ExamplesHugging Face · GitHub Copilot and its extension marketplace.
Part 3

Six things that are new because of AI

You cannot buy them or copy them quickly. They are built day by day — and the earlier you start, the wider the gap.

Type 1

Made stronger by AI

Advantages that already existed; AI just makes them bigger. Take AI away and they are still there.

Your own data · Existing customers · Switching costs · Network effects
→ Most of the eight in Part 2.
Type 2

Only possible with AI

Without AI they do not exist. They come from what only AI can do: learn from every use, handle complex work on its own, run like a real employee.

The data loop · Wiring AI into a system · Control · A new way to charge
→ The six below.
01
⟶ The loop

The data loop

Data flywheel

Not a warehouse of data sitting still, but a loop: the product attracts users → every use generates your own data → your AI gets smarter → the product gets better → more users. The loop speeds up as it runs, so being a year ahead is very hard to close.

02
⟶ Capability

Wiring AI into a system that works

AI orchestration

Not "knowing how to use ChatGPT". It is getting several AIs to work together inside a real system: handling it when they get things wrong, checking outputs before use, putting a human approver at the important steps. It is a new craft — buying the same model does not give it to you.

03
⟶ Control

Control and a paper trail

Governance & audit

Once AI acts on the company's behalf, large customers ask immediately: who approved it, is it logged, how far is the AI allowed to go. Whoever can answer wins the contract — even against a better model.

04
⟶ How you charge

A way to charge only AI makes possible

Outcome pricing

Charging for results instead of per user. Or serving customers too small to be profitable with human staff but profitable with AI. Incumbents struggle to copy it because it breaks how they make money today.

05
⟶ People

A team used to working with AI

Human-AI teaming

The whole company knows how to bring AI into decisions without losing human judgment. It is why two companies with identical tools get completely different results. You cannot buy it; it takes years.

06
⟶ Time

Being early on AI regulation

AI compliance

The rules made specifically for AI are still being written — Europe, the US and the international standards bodies all published theirs only in the last few years. Move early and you gain both a time lead and a voice in the rules.

→ Part 4 · What to do

Where should a small
company start?

You do not have a conglomerate's money or an incumbent's customer base. But you have three things they lack: speed, deep knowledge of one industry, and the ability to put everything into one place. Six steps, in order.

Step 1 — pick your ground  ·  Steps 2–5 — build capability  ·  Step 6 — accumulate over time
01

Pick one genuinely narrow niche

Do not build "AI for every industry". Pick the single industry you know best — accounting, law, healthcare, domestic logistics — and solve the hard cases that general AI usually gets wrong. The more exceptions and regulation a field has, the harder it is for others to push in.

02

Make the product generate data from day one

Design it so every use gives you something back: what they corrected, which option they chose, where the AI got it wrong. This asset compounds — every day you start earlier is a day of lead.

03

Invest in wiring AI together, not in training your own model

A small company almost never has the resources to train its own model. The win sits one layer up: getting AI to run reliably, recover from errors, check its outputs, and stop for a human at the critical step. That is what large customers actually pay for.

04

Build the control layer into the very first version

Log who did what, set permissions, require approval before the AI acts, filter sensitive content. This is not a feature to add later — without it you cannot sign a large contract. Shipping it early puts you roughly a year ahead.

05

Charge for results, not per user

Customers do not care how many minutes the AI ran; they care whether the job got done. Charge per completed task, per cost saved, per revenue generated.

06

Get certified before customers ask

SOC 2, ISO 27001 and ISO 42001 take nine to eighteen months. Expensive and slow, but they are the ticket into serious customers. A year's head start opens accounts your rivals cannot touch yet.

Stack all three layers = an advantage
rivals need years to break

Any single layer can be copied. But having all three inside the same narrow niche is very hard to unpick — a rival has to rebuild all three from scratch while you keep moving.

Deep industry knowledgeLayer 1
Your own data loopLayer 2
AI wired into a systemLayer 3

Before, you won on what you owned.
Now you win on running AI
better than the rest. — In one specific industry. Over years, not over a quarter.

For a small company this is both bad news and good news. Bad news: being good at engineering is no longer enough to win. Good news: for the first time, a small company can build a real advantage within a few years — if it picks the right niche and starts collecting data early enough.

Sources

Barney, J. (1991). Firm Resources and Sustained Competitive Advantage. Journal of Management, 17(1).

Helmer, H. (2016). 7 Powers: The Foundations of Business Strategy.

Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press.

Porter, M. E. (1980). Competitive Strategy. Free Press.

Iansiti, M. & Lakhani, K. R. (2020). Competing in the Age of AI. HBR Press.

Katz, M. & Shapiro, C. (1985). Network Externalities and Compatibility. American Economic Review, 75(3).

Klemperer, P. (1995). Competition when Consumers have Switching Costs. Review of Economic Studies, 62.

Daugherty, P. & Wilson, H. J. (2018). Human + Machine. HBR Press.

Bai, Y. et al. (2022). Constitutional AI. Anthropic Technical Report.

Park, J. S. et al. (2023). Generative Agents. UIST '23.

NIST (2023). AI Risk Management Framework 1.0.

European Parliament (2024). EU AI Act — Regulation 2024/1689.

← All writings