The bridge to applied AI

Boards now ask every chief executive the same question: what is our AI plan? Investors ask it of every company they back. The honest answer, in most companies, is a handful of trials, a chat assistant, and a slide. The money has been spent on the idea of AI. Very little has reached the work.

The gap between the demo and the P&L

The record is sobering. In 2025, 42 per cent of companies abandoned most of their AI initiatives.¹ Large, transformational deployments cost between $5 million and $20 million.² For a company of a few hundred people, that is not a pilot. It is a bet on the whole year.

The demos were not the problem. The demos were good. What failed was the crossing: from a model that can write a paragraph to a business that closes more deals, answers customers faster and knows its cash position every morning.

Why pilots stall

They start with the technology, not the job. A model is a capability. A business needs an outcome: a follow-up sent, a ticket resolved, an invoice chased. Without a named job, a pilot has no finish line.

They sit beside the work, not inside it. A separate AI tool is one more tab. People try it, then go back to the system where the work actually lives.

Nobody owns the decision. When an AI suggestion goes wrong, who answered for it? If that is unclear, cautious managers switch it off, and they are right to.

The build is too big. Internal projects try to assemble the whole stack: data, models, connections, controls. MIT's research found that buying from specialist vendors succeeds about 67 per cent of the time; internal builds about 33 per cent.³

A pilot without a named job has no finish line. A pilot without a named owner has no future.

What carries a business across

The companies that cross do three things differently.

They buy applied, not general. Applied AI is software built around one job, such as running a pipeline or answering customers, with the AI already inside. The work happens where it always did, and the agents do most of it.

They keep judgement human, on purpose. Agents prepare, draft and chase. Anything that commits the business, a price, a promise, a payment, waits for a named person. That is not a brake on adoption. It is what makes managers willing to switch it on.

They measure from day one. A baseline before the start, the same measure at day 30, and a decision agreed in advance: expand, extend or stop. Evidence, not enthusiasm, decides the next step.

The case against buying

Some leaders fear that bought software makes them the same as everyone else, and that the real advantage lies in building their own. There is truth in it. Your data, your processes and what your people know are the advantage. But that advantage is expressed through the software, not in writing it. The companies that win will own their data and what is learned from it, and let specialists carry the engineering. That is why we insist the learning stays with the customer, as we argue in Sovereign intelligence.

What this means for boards and investors

Ask for jobs, not experiments. Which named jobs will agents do by the end of the quarter, and who approves their work?

Ask for the baseline. What does the job cost today, in money and hours? Without it, no return can be shown.

Ask what it costs to run, all in. Licences are the smaller part. The hours your people spend feeding systems are usually the larger one. When Kintoro moved onto piMonk, our software run rate fell 83 per cent in three months once that time was counted.⁴

Ask how fast it can stop. A good programme can be stopped at day 30 with the data returned. If it cannot, the risk is larger than the plan admits.

The bridge to applied AI is not a leap of faith. It is a series of short, measured crossings, one team and one job at a time. Where to cross first is the subject of The applied-AI grid.