AS | Ankit Sarawagi|Founder, CFOmatrix·July 2026·10 min read | AI SaaS Special |
- The model is a component, not the company. Customers cannot use a raw model; they buy the product, deployment and customization built on it. That is where defensibility lives.
- A thin wrapper is not a moat. A bare prompt over someone else’s model competes only on price. Depth of product, workflow integration and data is what separates a business from a demo.
- Proprietary data and the feedback loop compound. Data your customers generate inside your product is unique to you, and when it improves the product it becomes a flywheel a competitor cannot copy.
- Investors test the moat with numbers. They scrutinise the gross-margin trend, net revenue retention, revenue concentration and proprietary data, not which model you call.
| 12-30x ARR multiples AI SaaS commands, well above classic SaaS | >100% Net revenue retention: the clearest proof a product is embedded | 4 Layers above the model: product, deployment, workflow, data |
The “Anyone Can Wrap an LLM” Argument, and Why It Misses the Point
The skeptic’s case is simple: the intelligence in an AI SaaS product comes from a model anyone can rent by API, so there is nothing to defend. If a solo developer can rebuild your demo in a weekend, how can you be worth a 20x ARR multiple? It is a fair question, and if a company really is only a prompt and a login sitting on top of someone else’s model, the skeptic is right. That is a feature, not a business.
But the argument quietly assumes the customer wants the model. They do not. A customer cannot run their business on a raw model. A support leader does not want a chat window, they want tickets resolved inside their helpdesk, on their data, within their rules, measured against their SLAs. The model is one ingredient in delivering that. Everything else, the product, the deployment, the customization, the workflow it plugs into and the data it learns from, is the actual company, and none of it is a weekend project.
“The worry that anyone can wrap an LLM misses the point. Customers cannot use a raw model. They buy the product, the deployment and the customization built on top of it. The model is a component; the product is the moat.”
Ankit Sarawagi, from working across AI SaaS startupsThis post sits inside our wider guide to SaaS and AI SaaS finance for founders. Here we go one level deeper on the single question that both founders and investors keep circling: if the model is commoditised, what exactly is the moat?
Thin Wrapper vs Deep Product
Not everything called “AI SaaS” is defensible, and the honest founder should know which side of the line they are on. The difference is not the model, both use the same one, it is the depth of everything around it.
- A prompt and a login over a public model
- Generic, same for every customer
- No integration into how the customer works
- No data that improves the product
- Competes on price; switching cost is near zero
- Solves a full job end to end, reliably
- Deployed and customized to the customer
- Wired into their systems and daily workflow
- Learns from proprietary usage data
- High switching cost; expands inside the account
In the AI SaaS companies I have worked with, serious enterprise buyers rarely want a generic public tool. They want AI deployed in-house or heavily customized: their data, their security, their workflows. Delivering that is exactly the work a thin wrapper cannot do, and it is precisely where the moat is built.
The point is not that wrappers are worthless, some are useful first products, but that a wrapper is a starting line, not a finish line. The defensible companies use the easy model access to get to market fast, then spend everything after that building layers one to three: the product depth, the deployment, and the per-customer economics that make each account both sticky and profitable.
The Proprietary Data Flywheel
Of all the layers, proprietary data is the one that compounds. Public models are trained on public data that every competitor can access, so the base model is a level playing field. What is not level is the data your customers generate inside your product: their tickets, their transactions, their edits, their corrections, their outcomes. No competitor renting the same model has it, and it is what lets your product get measurably better for that customer than any generic tool.
Data gets the attention, but workflow integration is just as powerful. Once your product is wired into a customer’s systems, permissions and daily process, ripping it out means re-plumbing how a whole team works. That switching cost, not the model, is what keeps the account and lets it expand.
What Investors Actually Scrutinise for Defensibility
Here is the reassuring part for founders worried about the wrapper narrative: investors are not checking which model you call. They know the model is rented. They test whether the moat above it is real, and they do it with numbers, because a durable moat leaves financial fingerprints. In practice they do not discount usage or consumption revenue versus committed revenue either, which is why AI SaaS still commands ARR multiples in the 12 to 30 times range. What they scrutinise is whether the business behind that multiple holds up.
| What they check | The moat it proves | Good sign |
| Gross margin trend | Compute is getting more efficient, not runaway | Rising |
| Net revenue retention | Product is embedded; accounts expand | >100% |
| Revenue concentration | Not one or two logos away from collapse | Diversified |
| Proprietary data | A data and workflow asset a wrapper lacks | Compounding |
Read them together and they describe the same moat from four angles. A rising gross margin says your compute is becoming more efficient and you are not just passing model bills to customers. Net revenue retention above 100 percent is the single hardest signal to fake: it means customers are not just staying, they are buying more, which only happens when a product is genuinely embedded. Low concentration shows the product wins broadly, not because of one champion. And proprietary data is the asset that says the moat will still be there after the next round of model price cuts.
“When investors look at an AI SaaS, they are not auditing the model. They look at where gross margin is heading, whether net revenue retention is above 100, how concentrated the revenue is, and whether there is real proprietary data. That is how you prove a moat with numbers.”
Ankit SarawagiThis is also why the wrapper worry fades in a real diligence conversation. A thin wrapper cannot show an expanding NRR or a rising margin, because it has nothing customers get locked into. A deep product shows both, and the numbers do the arguing. If you are heading into a round, the same signals shape how you tell the story from seed to Series A.
Will Falling Model Costs Commoditise AI SaaS?
The final version of the wrapper fear is about the future: models keep getting cheaper and better, so surely margins collapse and everyone converges. The direction of travel is real, but the conclusion is backwards for a genuine product. Falling model costs commoditise the model, which was never your moat. They do not touch the product, the deployment, the workflow or the data, which are.
If your value is the product and the data, cheaper compute flows straight to gross margin. The customer still pays for the outcome, your biggest variable cost falls, and margin rises. That is exactly the trend investors want to see. Cheaper models only hurt businesses whose entire value was reselling model access, which is the definition of a thin wrapper.
So the two ends of the wrapper argument actually cancel out. Yes, anyone can call the model, and yes, the model keeps getting cheaper. For a company that is only a model call, both are fatal. For a company whose moat is the product, the deployment, the workflow lock-in and the proprietary-data flywheel, easy model access got you to market and cheap models make you more profitable. The moat was never the model. It was everything you built around it, customer by customer.
“Nobody buys the model. They buy the product built on it, deployed into their world, learning from their data. Wrapping an LLM is where you start. Building the product around it is the whole business.”
Ankit Sarawagi, CFOmatrixFrequently Asked Questions
Is a thin LLM wrapper defensible?
A thin wrapper, a bare prompt and a login over someone else’s model, is not defensible on its own, because anyone can rebuild it in a weekend and the model provider can ship the same feature. Defensibility comes from everything built around the model: a real product that solves an end-to-end job, deployment inside the customer’s environment, deep customization, workflow integration, and proprietary data that improves with use. A wrapper with none of that competes only on price and gets commoditised.
What makes an AI SaaS defensible?
Customers cannot use a raw model to run their business, so they buy the product, the deployment and the customization built on top of it. The moat is the product and workflow integration (how deeply you are embedded in how the customer works), proprietary data (data the customer generates inside your product that no competitor has), and the feedback loop that turns that data and usage into a better product over time. The model is a component you rent; the product, the workflow lock-in and the data are what you own.
Do investors care about the model or the product?
Investors care about the product, not which model you call. They do not spend diligence checking whether you fine-tuned a model; they scrutinise the signals of a durable business: gross margin and its trend, net revenue retention (do accounts expand, which proves the product is embedded), revenue concentration (how much depends on a few logos), and proprietary data and workflow depth. A strong product on a rented model beats a clever model with no product.
Does proprietary data matter for an AI SaaS moat?
Yes, proprietary data is one of the strongest moats an AI SaaS can build. Public models are trained on public data everyone can access, so they are not a differentiator. The data a customer generates inside your product, their tickets, transactions, corrections and feedback, is unique to you. When that data continuously improves your product, and the improved product attracts more usage which generates more data, you have a flywheel a new entrant cannot copy by wrapping the same model.
How does deployment create a moat?
Enterprises increasingly want AI tools deployed in-house or heavily customized to their environment, security posture and data-residency rules, not a generic public tool. Delivering that, private deployment, integrations into their systems, role and permission models, custom workflows, is hard work a thin wrapper cannot match, and once you are deployed inside the customer’s stack the switching cost is high. Deployment and customization are where a lot of AI SaaS defensibility actually lives.
Will falling model costs commoditise AI SaaS?
Falling model costs commoditise the model, not the product. Cheaper models actually help a real AI SaaS: gross margin rises as compute gets cheaper, and the value the customer buys, the product, workflow, deployment and data, is unchanged. Cheaper models only threaten businesses whose entire value was reselling model access. If your moat is the product and the proprietary data, falling model prices are a tailwind, which is also why investors watch the gross-margin trend rather than the absolute margin today.
This is general educational information for founders, current to mid-2026, drawing on the author’s experience across SaaS and AI SaaS startups, and is not legal, tax or investment advice. Benchmarks such as ARR multiples and net revenue retention are indicative and vary by stage and business model. Verify the current position or consult a professional before acting on a specific matter.
AS | Founder, CFOmatrix | Finance Strategy & Equity Compliance CFOmatrix is a knowledge platform focused on how finance actually works inside growing companies. This guide draws on hands-on experience across SaaS and AI SaaS startups, from defensibility and per-customer margin to pricing, structure, fundraising and the finance function. |