In late 2025 we counted the AI copilots in our LinkedIn feed for one week. We stopped at 41. Most of them did the same thing: a chat interface, a prompt template library, an integration with Slack or Notion, and a 14-day free trial. The kind of product that gets compared to ChatGPT in week two and loses that comparison every time. Meanwhile, the products we ship at the studio (Carriva, PrepareMesCours, Creaclases) keep getting paid by people who would not pay a horizontal copilot for the same problem. This is our argument for vertical AI SaaS in 2026, with the math, the failure modes, and the parts we got wrong.
What "vertical" actually means in 2026
A horizontal copilot is "AI for everyone": Gmail summary, meeting notes, code completion, document drafting. Useful, often delightful, almost always undefended. The marginal cost of switching is roughly zero. The marginal cost of OpenAI or Anthropic launching a free version is approximately one keynote.
A vertical AI SaaS is "AI for one job in one industry, with the data and workflow baked in": auditing a French RIS for a retirement advisor, generating a Spanish-language lesson plan that respects the curriculum of one of ten Latin American countries, drafting a tutor profile that matches the credentialing rules of a specific market.
The line is not "do you use an LLM?" Both do. The line is "is the LLM the product, or is the LLM the engine inside a product that does work the LLM cannot do alone?"
The four moats horizontal copilots cannot copy
We picked our four products around four moat patterns. They are not perfect. They are what we found by accident and now look for on purpose.
1. Proprietary data shape
Carriva is the cleanest example. The RIS is a PDF that a human emails to an advisor. There is no API for it. We had to build the parser. Six months later, our parser handles 14 RIS layouts (yes, the official document has variants) and a horizontal copilot would need to do that work from scratch and care enough to maintain it as the document changes.
A horizontal AI is a powerful generalist who has not read your specific files. A vertical AI has read 4,000 of them and remembered the edge cases.
2. Workflow that is not a chat
Teachers using PrepareMesCours do not want to chat with an assistant. They want a séance ready in 8 minutes that respects their academy's pedagogical framework, their level, their duration, and their classroom constraints. The product is a form. Then a generation. Then a polished PDF that goes into a binder. The chat is a 5% surface, not the product.
Most horizontal copilots are 95% chat. The minute the customer's job is "produce this artifact in this format", chat is the wrong primary surface.
3. Regulatory or domain trust
Carriva audits an official document. The output influences advice that influences a person's retirement income. We cannot hallucinate. We had to build a retrieval system over authoritative pension-law documents and architect the prompt so the LLM cites and quotes rather than invents. We wrote about running RAG in regulated industries because the constraints there are different from horizontal use cases.
A horizontal copilot is allowed to be wrong sometimes. A vertical AI in a regulated space is not. That is a moat, because most generalists will not engineer for it.
4. Distribution into a real channel
Vertical SaaS lives or dies on whether you can reach the niche. With CGPs we can. There are directories, associations, conferences, and 534 prospects we can name. With teachers we can: they share lesson plans on dedicated forums and our French-language SEO ranks for terms a horizontal copilot will never optimize for.
Distribution into a horizontal market is the most expensive game on Earth. Distribution into a vertical is just the work of caring more than anyone else.
The 2026 sharpening: why this gets better, not worse
In 2024 we worried that LLMs would close the moat. Anthropic and OpenAI release ever-more-capable models. Surely the gap shrinks?
The opposite happened. In 2026 the moats sharpened.
When the model is a commodity, the work above the model becomes the entire product.
Three forces drive that:
- Cost-per-call is collapsing. A generation that cost us $0.40 in early 2024 costs us roughly $0.07 today on the same quality. That money does not show up as our margin. It shows up as our ability to ship features that would have been unaffordable two years ago.
- Long-context windows reduce the value of pure prompt-engineering tricks. A 1M-context window means you can throw the regulation, the customer's history, and the example outputs into one call. The skill is no longer the prompt. The skill is the data plumbing.
- Buyers are sophisticated. A French CGP in 2026 has tried ChatGPT. They know what a generalist gives them. They are now asking the right question: "does your product know my domain and my workflow, or are you a wrapper?" Wrappers lose that conversation.
If you are evaluating an AI SaaS idea in 2026, the question is no longer "can I add AI to a workflow?" The question is "can I add a workflow to AI?" That is what vertical means.
What the math looks like for a small studio
We run a one-person studio (with help from contractors when it counts). Here is roughly the cost shape for a vertical AI SaaS we ship:
- LLM cost per active customer: $1.20 to $4 per month at current pricing, depending on how chatty the workflow is. We kept this under control by writing about pricing AI features so the unit economics work.
- Hosting cost per product: we self-host on Proxmox behind Coolify (around $4 per month for a small VM). Each app ships as a Next.js 15 standalone Docker container on that stack.
- Analytics cost: Umami self-hosted, $0 marginal. PostHog Cloud free tier (1M events/month) for one product. We deliberately moved off self-hosted PostHog Hobby because it was eating 12 GB of RAM for a single project.
- Marketing cost: SEO and direct outreach. Roughly $0 in paid acquisition right now. We are deliberately not turning on ad spend until we cross 100 daily organic clicks per product.
Net: a vertical AI SaaS at 50 paying customers (which is small) at 50 EUR per month is 2,500 EUR per month gross with maybe 200 EUR in variable costs. That math is not glamorous, but it stacks across 4 products and it does not need a Series A to exist.
A horizontal copilot at the same scale spends 5 to 10 times as much on customer acquisition because it competes with everybody. The unit economics are not the same product.
The mistakes we keep making
We pretend we are not making them. We are.
We keep over-investing in features that thrill us and not the customer. The temptation to ship a beautiful prompt-engineering layer that nobody asked for is real. We caught ourselves three times in the last quarter.
We sometimes confuse "shipping a feature" with "selling the feature". A vertical AI SaaS is a sales-and-product company. The product needs to ship, then it needs to be sold. We are still mediocre at the second part. Our recovery is to do customer development one call per week, every week.
We have been late to differentiated SEO. For the first 9 months of Carriva we ranked for the same terms as Factorielles and got buried. Our cleanest growth came from rewriting our angle (audit, not simulation) and writing 30 metier-specific landing pages. That is not glamorous AI work. It is just work.
How to pick your vertical
Three filters we use:
- Can you name 50 individual practitioners in the niche by Tuesday? If you cannot, distribution will eat you.
- Does the work produce an artifact (PDF, plan, diagnosis, document) that a human would print or send? If yes, AI can compress the time to artifact and the customer will pay.
- Is there a domain expert who will sit with you for 90 minutes and care if you build it well? If no, you are building from imagination, and imagination loses to specificity every time.
If a prospective vertical AI SaaS fails any of those three, we walk away. If it passes all three, we have shipped to production in 6 weeks. We documented our best example in the post on how PrepareMesCours went from idea to production in 6 weeks.
The "wrapper" objection, addressed
A founder building horizontal AI will sometimes tell us "you are just a wrapper around an LLM". They mean it as a critique. We hear it as a misunderstanding.
A wrapper around an LLM is a chat UI with a prompt template. That is genuinely fragile. The model provider can ship the same thing tomorrow.
A vertical AI SaaS that uses an LLM is a wrapper around a workflow, with the LLM as one ingredient. The workflow includes data ingestion, document parsing, retrieval over a curated corpus, structured output validation, audit logging, multi-tenant access controls, billing, customer support, and the institutional knowledge of having sat with 50 practitioners in the niche.
The model provider cannot ship our workflow tomorrow because they do not know our niche. They are not going to read 4,000 RIS documents. They are not going to spend three weekends with a CGP in Grenoble. Their incentive is to be horizontal. Our incentive is to be specific.
Every time we hear the wrapper critique we ask the critic: "if this is a wrapper, why has the model provider not shipped it?" The answer is always silence or "they will eventually". Eventually is a weak threat to a product that is shipping today and accumulating customer-specific value every week.
What "ship the niche" looks like in practice
A small thing we do that compounds: every Friday afternoon, we spend 30 minutes adding to the studio's "niche knowledge" notes. New regulation we noticed. New competitor move. A vocabulary term we mis-used in an email. A specific advisor's complaint about a feature.
These notes feed into our prompts. They feed into our content. They feed into our customer-development calls. After a year, the notes are an asset that no horizontal copilot can replicate without the same 52 Friday afternoons.
The moat in vertical AI is the accumulated, specific knowledge that takes a year to build and an hour per week to maintain.
This is the part founders most often skip. They want the moat without the year. There is no shortcut. There is no growth-hack version of "spend 90 minutes a week with a domain expert for 18 months". The work is the work.
What is next
For us in 2026: more vertical AI SaaS, smaller scope per product, faster customer-development loops, and a deliberate refusal to build the next horizontal copilot. The world has 41 of those this week. It does not need our 42nd.
If you are sitting on an AI idea right now, the most useful thing you can do is name a practitioner you can email today, send them one honest sentence about a problem you think they have, and see whether they reply in a way that suggests they would pay you to solve it. That is the entire test. The model will be cheaper next quarter. The vertical will not be.
