Why most AI pilots fail — and what a small business can do about it

The number

MIT NANDA's State of AI in Business 2025 report found that despite heavy investment, about 95% of enterprise generative AI pilots delivered no measurable return on the profit and loss statement. It remains one of the most cited figures on AI adoption.

It is easy to read that as “AI doesn't work”. The report says something more useful.

The cause is not the model

The failures traced back to what the report calls a learning gap: tools that don't adapt to how a business actually works, weak integration with existing systems, unclear ownership, and organisations that couldn't absorb the change.

Research on agents in production points the same way. Failures come from weak verification, missing escalation paths and security gaps such as prompt injection — engineering and process problems, not intelligence problems.

Partners do better than DIY

The same MIT report found that implementations led by an outside vendor or partner succeeded far more often than internal builds. For a business with no engineering team, that matters: the question is less “which AI tool?” and more “who owns making it work?”

What this means for a small business

Start with the data. If the source is a messy sheet or a paper register, automating it only makes the mess faster. Check readiness first.

Start with one workflow. Industry guidance for small businesses reports that a single focused pilot reaches ROI three to five times faster than a company-wide programme, with typical payback in four to eight months.

Measure before you build. Without a baseline, there is no way to show the automation paid off — and no reason to keep it running.

Keep a person in the loop until the results earn trust, and make sure someone is responsible for the automation after launch.

How we apply it

The Trellient framework is built around these findings: a readiness and data audit before any build, one pilot chosen by ROI, human approval on every output, and monthly measurement against the baseline.