An AI moat is whatever stops competitors, or a free AI tool, from copying your business now that AI makes features cheap to build. Moats that still hold include proprietary data, being embedded in customer workflows, regulatory and domain know-how, distribution to an existing customer base, community, and a cost-to-serve that protects your gross margin. A thin layer over someone else's model is not a moat.
Key takeaways
- AI drains moats built on content libraries, simple features and per-seat pricing.
- Proprietary data, embedded workflows, domain rules and distribution are getting more valuable, not less.
- For AI products, gross margin is a moat: wrappers without a cost lever lose every price war.
- Charging for the work done (resolutions, transactions) protects revenue when AI makes customers more productive.
- Ask: if AI did my core deliverable free tomorrow, what would customers still pay me for?
Chegg had what looked like a wide moat.
Millions of students paid monthly for homework help. Its library of step-by-step answers was enormous, and Google sent a steady stream of students straight to it. Content plus traffic. Hard to copy.
Then generative AI started giving students personalized homework help for free. Google started answering questions directly in search with AI Overviews.
In the first quarter of 2025, Chegg's subscribers fell 31% to 3.2 million. In October 2025 it cut about 45% of its staff. Its CEO at the time, Nathan Schultz, said Chegg "would not need to review strategic alternatives if Google hadn't launched AI Overviews." By the second quarter of 2026, revenue was down 51% on a year earlier.
Two moats, both drained in about two years.
That's what this page is about. When AI makes features nearly free to build, which moats still hold? And how do you find out whether yours is one of them?
Why AI changes the moat question
Two things are happening at once.
Copying got cheaper. A small team with AI coding tools can rebuild a product's features in weeks. Content that took years to write can be generated on demand.
Intelligence got cheaper. Epoch AI found that the price of reaching a given level of AI performance fell by about 40 times a year, with a wide range depending on the task. Whatever you charge for "smart" output, a cheaper version is coming.
So any moat built on what your product can do is shallower than it was. Moats built on what only you have, where only you sit and what only you know are getting deeper.
Which moats does AI erode?
| Moat | Why AI weakens it | Example |
|---|---|---|
| A content or answer library | AI generates answers on demand | Chegg |
| Search traffic as your main channel | AI answers the question before the click | Chegg's Google traffic |
| Simple features | Faster and cheaper to copy than ever | Any single-feature tool |
| Per-seat pricing | AI lets customers do the same work with fewer people | WiseTech's reason for scrapping seats |
| Being a thin layer on someone else's model | The model vendor can undercut you, and owns your costs | Low-margin AI wrappers |
Which moats can AI not copy easily?
Proprietary data
Not public data. Every model has already read the internet. Data only you collect: your customers' history, outcomes you've measured, signals your users generate. Arena built a free AI leaderboard on more than 10 million user evaluations, then sold the resulting data to AI labs. Tomorrow.io builds its own weather satellites.
Embedded rules and workflows
WiseTech put it bluntly in 2026: its embedded customs and regulatory rules, built over years, "cannot be replicated by prompting an AI model." Looker's semantic layer, the part of the product that once seemed "too technical," is now how Google lets AI agents query a company's data with trusted definitions. The hard, boring know-how is the moat.
Distribution to an installed base
The incumbents who already sit in front of the customer can bundle AI and raise prices. Slack put AI into Business+ and lifted the price from $12.50 to $15 per user. Salesforce reported Agentforce annual recurring revenue above $1.5 billion in August 2026. Canva sells AI to more than 265 million monthly users. A great AI feature with no distribution is a demo.
Community and trust
People don't switch communities because a competitor's software is 10% better. And in regulated work (health, legal, financial), customers buy accountability, not just output.
Where I'm placing my own bets
This is the thinking behind Nagaris, the AI platform I founded for regulated industries. The models will keep getting better and cheaper for everyone. In regulated work, what's hard to copy is the compliance, the data sovereignty and the trust that a firm's clients' information is handled properly. I'd rather build there than try to out-model the model companies.
The AI wrapper problem: margin is a moat too
In the book, high gross margin is one of the eleven moat factors. In AI, it's become one of the most important.
Traditional software costs almost nothing to deliver to one more customer. AI products don't. Every query, every agent run, every generated image costs real money in compute. Bessemer reported that its fastest-growing AI companies averaged about 25% gross margin, "often negative," against about 60% for steadier ones.
The extreme case is the wrapper: a product that resells another company's model. Cursor reportedly ran a negative gross margin in the quarter to January 2026, and responded by building its own model and routing work to cheaper ones. Wix said its app builder Base44 started 2026 at a gross margin of "nearly zero" and targeted about 60% after moving to its own model.
The lesson: if you don't control your cost to serve (your own model, smart routing, caching, or pricing that passes heavy usage through), the model vendor effectively sets your margin. Read how to price an AI wrapper without losing money and AI gross margins.
Check whether your AI product has a margin moat
Model token costs, heavy users and gross margin with the free AI Gross Margin Calculator before a price war finds the weak spot for you.
Pricing as a moat: charge for the work, not the seat
Here's a moat few people talk about. Your pricing model can protect you, or expose you.
If you charge per user, and AI lets your customer do the same work with half the team, your revenue halves while you deliver more value. WiseTech saw this coming. In December 2025 it launched CargoWise Value Packs, removing seat fees entirely and charging on transactions. By February 2026 about 95% of customers were on the new model. Its words: "our revenue is not diluted by customers becoming more productive."
Intercom made the same bet from the other side. Its Fin AI agent charges $0.99 per resolution, billed only when the customer's problem is solved. Every improvement in Fin's resolution rate earns Intercom more money, so its whole team is aligned with the customer's result. Read how Intercom bet the company on 99 cents per resolution and the guide to outcome-based pricing.
Credit Pack DesignerTurn your cost per action into a credit rate card with healthy margins.Open the free tool →The MOATS framework, updated for AI
| Letter | Classic version | AI-era version |
|---|---|---|
| M Marketing supremacy | Cheap acquisition, strong brand | Being the name AI assistants recommend; owning an audience instead of renting search traffic |
| O One | No single point of failure | Not dependent on one model vendor or one channel |
| A Audience | Low concentration, network effects, ongoing usage | Data network effects: every user makes the product smarter |
| T Toughness | Advantage, margin, scale | Cost to serve under control; domain rules competitors can't prompt their way to |
| S Switching | Switching costs, embedding | Workflows, proprietary data and integrations customers would have to rebuild |
For the full framework and a scoring rubric, see the MOATS framework.
The AI moat test for your business
Answer these honestly:
- If AI did my core deliverable for free tomorrow, what would customers still pay me for? Write the answer down. That's your moat. If the answer is "nothing," move up to outcomes, accountability or done-for-you service now.
- What data do I have that a competitor can't buy or scrape?
- Where am I embedded in my customers' systems, workflows or routines?
- Do I depend on one model vendor, one platform or one channel?
- Does my pricing grow or shrink when my customers get more productive?
- For AI features: do I know my cost per customer, including my heaviest users?
What to do next
- Run the test above and write a one-paragraph answer to question 1.
- Move up the value chain if your answer is weak: from answers to outcomes, from drafts to done-for-you.
- Start capturing proprietary data from every customer interaction, with permission.
- Check your pricing metric. If it's seats, model what happens if your customers' teams halve.
- Model your AI margins before you scale any AI feature, using the AI Gross Margin Calculator.
- Score your moat on all five MOATS letters and re-score every year. In the AI era, once every two years is too slow.
Frequently asked questions
Is data still a moat when AI models are trained on the whole internet?
Public data isn't. Everyone's model has already read it. Proprietary data is: your customers' transaction history, outcomes you've measured, sensor or satellite data you collect, evaluations only your users generate. The test is whether a competitor could buy or scrape the same data. If they can't, and it makes your product better, it's a moat.
Can an AI wrapper ever have a moat?
Yes, but rarely from the model. A wrapper can build a moat from distribution, a workflow customers live inside, proprietary data from usage, or a niche brand. What it can't do is compete on price with the model vendor it depends on. Several fast-growing AI app companies have reported negative or very low gross margins for this reason.
Does AI make small service businesses less defensible?
It removes the moat of doing routine work faster than the client could themselves. It strengthens the moats of trust, accountability, local relationships and knowing a client's situation. An accountant who uses AI to do compliance work faster and spends the saved time advising is more defensible than before, not less.
Should I build my own AI model to have a moat?
Almost never as a small or mid-sized business. Model capability keeps getting cheaper and the big providers improve every few months. Build your moat in the layers around the model: your data, your workflow, your distribution and your pricing. Larger AI companies do build or tune their own models, but mainly to cut cost, not to look different.
How do I know if AI is a threat to my subscription business?
Look at what customers pay you for. If the answer is information, first drafts, answers or simple analysis, AI is a direct substitute and you should move up to outcomes, accountability or done-for-you service. If customers pay for access, relationships, physical delivery or being embedded in their systems, AI is more likely a tool you can use.
Get both books free
The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.
Robert Coorey, MBA
Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.
Sources
- Higher Ed Dive, Chegg slashes nearly half its workforce as AI eats into its business (Oct 2025)
- Chegg, Q2 2026 results (Aug 2026)
- WiseTech Global, 1H26 results briefing (Feb 2026)
- Epoch AI, LLM inference price trends (Mar 2025)
- Bessemer Venture Partners, The State of AI 2025
- Intercom, Fin pricing (accessed Oct 2026)
- Google Cloud, Looker product page (accessed Oct 2026)
Facts and prices checked 27 October 2026. Spotted something out of date? Tell us.