First AI Startup? Why the Model Is the Easy Part

Why most first AI companies stall before they start

Picture a founder six months into their first AI startup. The demo is sharp. The model is the best money can buy. The pilot client loved the pitch. And yet nothing is live, nobody is using it, and the renewal conversation is getting quieter by the week.

That founder is not unusual.

MIT's Project NANDA, in its report The GenAI Divide: State of AI in Business 2025, found that 95% of generative AI pilots stall early and never reach scaled adoption, and its lead author put the failures down to how organisations use the models rather than the models themselves.

Gartner goes further for agents specifically, predicting that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls.

At Novosapien we have spent two years building AI Workforces for enterprise and SME clients, and our co-founder George Westbrook has watched this pattern from inside the build. George shapes what we build and how it lands with clients, from first idea to live deployment. When we sat him down for our co-founder interview, Stop Fixing Your Data. Start Fixing Your Requirements., he did not blame the model once. He blamed systems, collection and people. In his words: "You can build the best solution in the world and still fail at this step."

If you are about to start your first AI company, or your first agentic AI product inside an existing one, here are the three lessons George would give you before you write a line of code.

Lesson 1: Test whether a client can adopt AI, not just whether they want to

Every first AI founder meets the enthusiastic buyer. The CEO has read the headlines, the board has signed off a budget, and the meeting ends with handshakes. George's warning is that enthusiasm is the wrong signal to qualify on.

"Most companies assess their willingness to adopt AI and skip their ability to adopt it. They check leadership appetite and budget. They rarely check whether the business can actually connect AI to its own data."

The gap, he says, usually hides in the plumbing. "The blind spot is usually the systems. Client data sits in legacy platforms that were never built to be reached from outside." And when that happens, no amount of goodwill closes it: "If AI can't get to that data through a secure, well-defined interface, the chain has a broken link, and executive enthusiasm doesn't fix it."

The research backs him up. A Gartner survey of 1,203 data management leaders found 63% of organisations either lack, or are unsure they have, the right data management practices for AI. That is the market your first AI startup is selling into.

George's fix is two questions, asked before any budget is committed:

1. Can AI reach the data it needs, API-first, with the right security around it?
2. Is that data in a form AI can use?

The second question is the one founders misread. George is clear that AI does not need a pristine data warehouse: "AI doesn't need perfectly structured data the way reporting does. Unstructured data works too. It just can't be locked away."

What this means for your first AI company: build access into your discovery process. Before you sign a pilot, map which systems hold the data your product needs and whether you can reach them through an API. A client who cannot connect is not a client yet, however excited they are.

Lesson 2: Build the collection layer on day one

Ask George what he would lay down first if he started again from zero, and he does not say the model, the stack or the database. He says collection.

"The collection layer. I'd map out how data gets collected and ingested on day one, and treat it as a use case in its own right rather than plumbing to sort out later."

His reasoning flips the usual founder worry on its head. "Turning unstructured data into structured data is now the easy part. The hard part is having the data at all." And you rarely know in advance what you will need: "The honest difficulty is that you often only discover what data you need at the moment you need it. That's exactly why collection belongs in the foundation. Processing is cheap now, so capture broadly first and structure it later."

We practise this inside Novosapien. George describes our vault as the place where "every meeting, conversation, spec and idea is captured and organised in one place," with the goal of "an x-ray view of the whole business." Every file sits under version control, and every agent trace is kept: "The lead that responded well and the lead that responded badly both become material for future agents to stand on the shoulders of."

Capture broadly, but feed agents carefully

Capturing everything does not mean dumping everything into a prompt. George is blunt about this: "Extract structured data from sources like meeting transcripts rather than handing over raw text, which bloats the context and causes context rot."

Context rot is a documented effect. Anthropic's engineering team, in Effective context engineering for AI agents, explains that as the number of tokens in a model's context grows, its ability to recall information accurately falls, a finding first measured in Chroma's context rot research.

George also pushes founders past the default retrieval setup. "Classic RAG, which chops documents into chunks and runs semantic search over them, looks good on the surface but limits how well a capable model can search." Instead: "What works better is giving the model what looks like a local file system that it searches with bash commands. Current models are very good at that. Link the files into a knowledge graph, attach metadata to each page, and the model finds what it needs quickly." Anthropic describes the same shift in the same article, noting that agents increasingly hold lightweight references such as file paths and load data at runtime, rather than pre-loading everything.

What this means for your first AI company: treat data capture as a product feature, not an afterthought. Decide on day one where every conversation, document and agent action will land, then invest in processing it into clean, structured context your agents can search.

Lesson 3: Adoption is part of the product, so trace everything

The story George tells most often is not about a technical failure. It is about a launch that looked like a success.

"We’ve seen systems built to fix how businesses collected data and ran their client-facing work. Delivered over several months, staff trained, adoption looked fine at launch."

Then the months after launch arrived. "The people closely involved used it and liked it. Others drifted back to old habits, often the quieter people who hadn't raised gaps in their own workflows during development."

With half the team in the system and half outside it, the records stopped matching, and people blamed the software rather than the habits behind it. "It nearly cost the whole project."

The fix was not a code change. The team worked with senior leadership to make raising issues an expectation from the top down, and usage spread.

His verdict: "It was a people problem that showed up in the data."

George breaks adoption into three layers: leadership's willingness to adopt, the business's ability to adopt, and the willingness of the people who use the system every day.

"Companies underestimate the third." He has seen what happens when nobody owns it: "We delivered a solution once where the junior staff meant to use it had no mandate from leadership. Using it was optional, so it didn't get used."

This is not unique to AI. Prosci's Best Practices in Change Management research, drawn from more than 2,600 practitioners, found that projects with excellent change management were roughly seven times more likely to meet their objectives than those with poor change management.

Then make every action visible

If adoption is how you win, tracing is how you keep winning.

Asked for the one control that must exist before an agent touches enterprise data, George does not hesitate: "Tracing. Every action an agent takes has to be recorded and stored. At some point performance will dip somewhere in the business. Without a trace of every action, you can't diagnose what went wrong, let alone fix it. With traces, the dip becomes a data set you can learn from."

Traces feed evals, and evals keep your product honest. "Test your models rather than guess. Run evals so the choice is data-driven. Even a so-called upgrade can regress in places, and in agentic work, tool calling is one of the most important things to test."

And as autonomy grows, George designs for checks, not solo heroes: "Full autonomy doesn't mean one agent running alone. It works like a good human team: one agent acts, another critiques."

Anthropic calls this the evaluator-optimizer pattern, where one model generates and another evaluates in a loop. At Novosapien, humans stay in that loop. Their feedback in the early stages is what builds the evals, and hard stops sit at every level. As George puts it, those stops "aren't there to slow the work down. They're there to stop catastrophic errors."

What this means for your first AI company: budget for adoption the way you budget for engineering. Agree who mandates usage, give quieter users a route to raise gaps, and ship tracing and evals in version one, not version three.

The founder's checklist

George's three lessons come down to one idea: the model is the easy part. What decides whether your first AI company survives is everything around it. Before you launch, check that you can answer yes to each of these:

1. Have you confirmed, system by system, that your product can reach the client's data through a secure API?
2. Do you know where every conversation, document and agent action will be captured, and how it gets processed into usable context?
3. Are you feeding agents structured, searchable context rather than raw text dumps?
4. Has someone senior mandated that the people who use your product every day actually use it?
5. Is every agent action traced, and are your evals built from real human feedback?
6. Do your agents work in pairs of actor and critic, with hard stops and humans in the loop?

At Novosapien, this is how we build AI Workforces: agents and people working side by side, with autonomy earned over time rather than assumed on day one. If you are starting your first AI company, or trying to get a stalled AI programme moving again, talk to us at novosapien.ai.

Which of George's three lessons is your business closest to getting wrong right now? Tell us in the comments.

Sources

• Novosapien, Stop Fixing Your Data. Start Fixing Your Requirements. co-founder interview, 6 October 2026 (all George Westbrook quotes)
• MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, via Computing
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025
• Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, 26 February 2025
• Anthropic, Effective context engineering for AI agents, 29 September 2025
• Chroma, Context Rot
• Anthropic, Building effective agents, 19 December 2024
• Prosci, The Correlation Between Change Management and Project Success

Publishing notes

• Primary keyword:
• Secondary keywords:
• SEO title (58 characters): How to Start Your First AI Company: 3 Lessons From a Founder
• Meta description:
• Slug: /blog/how-to-start-your-first-ai-company