Why AI Projects Fail (It's Not Your Data): Lessons from Novosapien's Co-Founders

The short answer

Most AI projects don't fail because of bad data. According to Novosapien co-founders Brett St Clair and George Westbrook, they fail on four things around the data:

• Requirements: nobody captures what the business actually wants in a form an AI agent can act on.
• Access: data sits in legacy systems that AI can't reach through a secure, API-first interface.
• Adoption: staff aren't required to use the new system, so they drift back to old habits.
• Trust: governance either blocks agents entirely or lets them act without tracing what they do.

The fix is to put the time saved on engineering into requirements, build a data collection layer from day one, make usage a leadership mandate, and give agents autonomy in stages, with every action traced.

Who we asked

Brett St Clair and George Westbrook are the co-founders of Novosapien, which builds AI Workforces for enterprise and SME clients. Brett is a former senior executive at Google and Barclays and leads Novosapien's sales. George is an entrepreneur who shapes what Novosapien builds and how it is delivered to clients, from first idea to live deployment.

We asked both the same questions about AI readiness, data strategy and AI governance, and they answered separately. Their answers overlapped far more than either expected. Below, each section starts with the takeaway, followed by what each co-founder said.

How do you know if your company is really ready for AI?

A company is ready for AI when its leaders have used the technology themselves and its systems let AI reach the data it needs. Budget and appetite alone don't count as readiness.

Brett St Clair sees three common mistakes. The first is believing the entire business must be mapped before an agent can touch anything, which leads to nine-month consultancy exercises that document a company that has already moved on. The second is cutting headcount too early. Building an agentic business takes around two years on a good run, and early cuts remove the people who would have taught the agents how the work gets done. The third is that most executives approve budgets without ever seeing agents working alongside their people. His advice: "Before any critical decision, get the C-suite spending a few weeks with their own hands on the technology."

George Westbrook puts it as willingness versus ability. "Most companies assess their willingness to adopt AI and skip their ability to adopt it." The usual blind spot is legacy systems that were never built to be reached from outside. Before committing budget, he asks two questions:

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

On the second point, he notes that AI doesn't need perfectly structured data the way reporting does. Unstructured data works too. It just can't be locked away.

Is data quality the biggest reason AI projects fail?

No. Both co-founders say data quality is now one of the easier problems to solve. The bigger blockers are confidence, requirements and adoption.

Brett St Clair: "Data quality is the easy part now. I know that is not what the industry wants to hear." The hardest part, he says, is getting what a human wants out of their head and into a shape an agent can act on. Across Novosapien's clients, the real blocker is confidence: "It is rarely skills or tooling. They cannot bring themselves to start."

George Westbrook points to people. He describes three layers of adoption: leadership's willingness, the business's ability, and the willingness of the people who use the system every day. Companies underestimate the third. "We delivered a solution where the junior staff meant to use it had no mandate from leadership. Using it was optional, so it didn't get used." The ownership gap, in his view, is ownership of adoption, not ownership of data.

Why do AI projects stall after launch?

AI projects often stall after go-live because part of the workforce drifts back to old habits, and the mismatch then shows up as a data problem.

George Westbrook shared an anonymised example. Novosapien built a system for a financial services client to improve how it collected data and ran client-facing work. It was delivered over several months, staff were trained, and adoption looked fine at launch. In the months after, people who had been closely involved kept using it. Others, often quieter staff who hadn't raised gaps in their own workflows, drifted back. With half the business in the system and half outside, records stopped matching and people blamed the system. "The cost was nearly the whole project." Novosapien worked with senior leadership to make raising issues an expectation from the top down, and usage is now much more widespread. "It was a people problem that showed up in the data."

Brett St Clair traces stalled projects to two things: turning what a human wanted into a usable output, then checking the output was right. Underneath both sits poor requirements. He cites PMI research attributing 47 percent of unmet project goals to poor requirements management. Because AI has cut software engineering time so sharply, he argues teams should spend most of a project on the requirement, prototype quickly, then refine. Done this way, he sees success rates above 80 percent, with delivery tracked to the exact line count.

Why is RAG underdelivering, and what works better?

When GenAI or RAG projects underdeliver, the model is rarely the cause. Retrieval design, data preparation and missing checks are.

George Westbrook says classic RAG, which chunks documents and runs semantic search over them, limits how well a capable model can search. A better approach is to give the model what looks like a local file system that it searches with bash commands, link the files into a knowledge graph, and attach metadata to each page. Data should be processed, not dumped: extract structured data from sources like meeting transcripts rather than passing raw text, which bloats the context and causes context rot. He also recommends running evals to choose models, since even a so-called upgrade can regress in some areas, and tool calling is one of the most important things to test in agentic workflows.

Brett St Clair agrees that Novosapien is finding less and less use for RAG. "What works better for us is structured Markdown, organised deliberately for agents to consume." The real causes, he says, are prompt structure, guardrails, and whether you built loops that check the agent did what you expected. "Everyone audits the retriever. Hardly anyone audits the loop."

Why are older data strategies failing to support AI agents?

Most data strategies written between 2020 and 2024 were designed for humans and BI dashboards, not for AI agents to consume.

Brett St Clair: "Because they were written for humans to read." A modern data strategy, he says, has to be written for agents first and humans second, so both understand the same data together, with outputs vetted continuously by agents rather than a quarterly steering committee. On spotting a strategy built for BI rather than AI, he is blunt: "There is no tell, because I have not yet seen the exception."

On tenders and statements of work, he notes most already include some data readiness line. What matters is accessibility: can an agent reach the data, with the right permissions, today? Because AI can now solve readiness quickly, the buyer-versus-vendor argument over who owns it matters less than it used to.

What does "data as a product" look like for AI?

Treating data as a product for AI means capturing everything the business knows, keeping its history, and structuring it so both people and agents can use it.

George Westbrook describes three practices inside Novosapien:

• The vault: every meeting, conversation, spec and idea is captured in one place, giving AI an "x-ray view of the whole business" across finance, engineering, sales and marketing.
• Version control on every file: a full history of what changed and when, which serves people and agents equally.
• Agent traces as data sets: good and bad outcomes, such as a lead that responded well or badly, become material for future agents to improve on.

Brett St Clair adds structured Markdown, which he calls "badly underrated", and capturing every human-to-agent and agent-to-agent interaction into an agentic brain. Client onboarding is his favourite example. Novosapien runs two weeks of requirements workshops, everything lands in the vault, and the requirements come out written for agents and humans at once.

What should you build first in an AI initiative?

Build requirements and data collection first. Architecture and processing are now the easy parts.

Brett St Clair: "Requirements. Requirements. Requirements." They should be granular enough to become a real backlog, because "that is the foundational data set now."

George Westbrook would start with the collection layer, mapping how data is collected and ingested on day one and treating it as a use case in its own right. Turning unstructured data into structured data is easy now. The hard part is having the data at all, and you often only discover what you need when you need it. His rule: "Capture broadly first and structure it later."

How should AI governance work: gates or guardrails?

Effective AI governance uses both, in sequence. Agents start behind human gates and earn their way to automated guardrails as trust builds. Brett St Clair and George Westbrook described the same three-stage model independently.

[embed: node/3a49b3be-de04]

Human checks in the early stages act as gates. Once success rates are high enough, they give way to guardrails.

Brett St Clair: "All three phases exist to do one thing: build trust between human and AI. That is what the governance is for."

George Westbrook adds that full autonomy doesn't mean one agent running alone. "It works like a good human team: one agent acts, another critiques." Novosapien splits work across many specialised actors and critics, organised into AI Workforces, and measures each layer against evals built from human feedback in the earlier stages. Hard stops, meaning strict code-level detection of things like malicious commands, apply at every stage. "They don't exist to slow the work down. They exist to stop catastrophic errors."

What governance must exist before AI agents act on enterprise data?

Before an autonomous agent touches enterprise data, every action it takes must be traced, and it must follow the same access and security rules as your people.

George Westbrook: "Tracing. By far, that's the most important mechanism." When performance dips somewhere in the business, a full trace of every agent action lets you diagnose the cause, fix it and improve the next run. Without one, you can't.

Brett St Clair: "Do not change your access, security and governance model." Agents follow the same rules as humans, with no carve-outs because it's a bot. What changes is resolution: agents handle small tasks, so governance breaks into a micro layer, with guardrails written at prompt level.

Frequently asked questions

What is the main reason AI projects fail?

According to Novosapien's co-founders, the main reasons are poor requirements, low adoption by staff, data locked in legacy systems, and a lack of executive confidence. Data quality is now one of the easier problems to fix.

What is AI readiness?

AI readiness is a company's ability, not just its willingness, to adopt AI. It means leaders have used the technology first-hand, and AI can reach the data it needs through secure, API-first interfaces.

Does AI need perfectly structured data?

No. George Westbrook says AI doesn't need data structured the way reporting does. Unstructured data works, provided AI can access it. Processing unstructured data into structured data is now one of the easier steps.

Is RAG still the best way to give AI access to company knowledge?

Novosapien is moving away from classic RAG. It favours structured Markdown, file systems the model searches directly, knowledge graphs and page metadata, combined with evals and checking loops.

What are the three stages of AI agent autonomy?

Stage one: a human starts the work and a human checks it. Stage two: an agent starts the work and a human checks it. Stage three: an agent starts the work and another agent checks it. Each stage is reached by earning trust.

What is the most important AI governance control?

Tracing. Every action an autonomous agent takes should be recorded and stored, so any dip in performance can be diagnosed and fixed. Agents should also follow the same access and security rules as human staff.

How long does it take to build an agentic business?

Brett St Clair estimates around two years on a good run, which is why he warns against cutting headcount before the agentic business is standing up.

Want to see AI Workforces built this way? Talk to Novosapien.