One factory first: an AI pilot that proves ROI in 90 days

Most organisations now use AI, but few can point to a financial result. McKinsey's 2026 survey of 1,719 respondents found nearly nine in ten organisations use AI regularly, yet only 37 per cent attribute any EBIT impact to it, unchanged from 2025. [1] In Gartner's survey of 2,002 CIOs and technology executives, only 13 per cent say their enterprise has realised significant or transformational value from AI, and Gartner puts the likelihood of use case success at one in five. [2] Among infrastructure and operations leaders, only 28 per cent of AI use cases fully met ROI expectations. [3]

The cause is rarely the model. Pilots start without a baseline or an owner, so when a result arrives nobody can say whether AI caused it.

The manufacturer in our session was not short of ideas or data. Its managing director described data "right from production shop floor, machine breakdowns to monthly dispatches, utilisation of raw material", sitting alongside an established ERP. His overarching goal was a material bottom-line gain, and he listed where it might come from:

• Lower inventory and better working capital management.
• Better procurement of the main raw material, where buying well drives a large share of margin.
• Production planning and better use of CNC machine time.
• Predictive maintenance based on recurring breakdowns.
• Sales forecasting, and help deciding whether to accept a low-margin order to fill spare capacity.
• Overtime management, using years of HR data on when overtime was and was not given.
• Fewer quality problems and less scrap.

Every item is a reasonable AI use case. Together they are a programme, not a pilot. As he put it, "the more data you have, the more time you need to analyse". The goal was not to become an AI company. It was to become a more profitable manufacturer.

From this session and others like it, a strong first pilot has five traits:

1. One site and one metric. A single factory and a single number, such as scrap rate or material yield, that finance already trusts.
2. Data that already exists. Production plans, coil or batch records, cut lists and scrap logs. If the data needs a year of cleaning, pick another pilot.
3. A clear owner. Here, a factory coordinator reporting directly to the managing director, with authority to get answers quickly.
4. Attribution built in. The managing director asked the right question: "Would we be able to identify that this intervention led to this?" Gains can also come from more business or other changes, so the baseline and comparison method must be agreed before the build.
5. A path to scale. The pilot should be repeatable at other sites or sister companies. In this case, a result at one company was meant to give "the confidence to apply to other companies" across the portfolio.

The managing director was specific about the output he wanted: not a one-off analysis, but "an engine which will keep on analysing the data as it comes and predict what will be the scrap". For a given production plan, it should estimate scrap, then show how a different plan would reduce it.

His example came from coil processing. Large coils are slit and cut, and leftover slit coils are scrapped or held as stock. Because the order book is known months ahead, a smarter plan can reserve leftovers for a job due in two months. That is a planning problem AI handles well: many constraints, lots of history, one clear objective.

The business also had an in-house slitting optimiser, built years earlier. The sensible first step is to improve that tool with the data now available, not to replace it. Good pilots extend what the plant already trusts.

In our experience these engagements work best when the team learns the process, not just the data. Handing over spreadsheets for analysis rarely produces a change on the floor. Sitting with the planners, seeing how a plan is made and where judgement comes in, is what turns a model into a tool people use.

Ninety days is long enough to build and measure a narrow pilot, and short enough to hold attention:

• Weeks 1 to 2: baseline and scope. Interview process owners, pull 12 months or more of scrap and production data, agree the metric, the baseline and how gains will be attributed. Our AI Compass roadmap covers this ground, including an ROI feasibility matrix, in 14 days.
• Weeks 3 to 8: build and shadow. Build a working prototype rather than slides. Run its recommendations alongside the planners' own plans so the two can be compared without risk.
• Weeks 9 to 12: run and measure. Let planners use the recommendations on live jobs. Track scrap against the baseline weekly and record every case where they overrode the tool, and why.

The board then sees one number moved, how it was attributed and what it cost. That is a decision, not a demo.

At the end of the session the managing director gave us blunt advice on our own pitch. Lead with impact that can improve profitability immediately. "Then you got my attention." He did not want training as the opening offer, and he did not want "a fancy project for the name of using AI in my company". His summary was three words, repeated: "Bottom line, bottom line."

That applies to any internal sponsor. Gartner found 44 per cent of CEOs struggle to demonstrate AI-driven financial returns. [4] "We will become AI-enabled" invites scepticism. One point of scrap, with a baseline and a date, invites a yes.

Quantify first, then expand. Once scrap has moved, the same data foundation and working method can be pointed at the rest of the list: predictive maintenance, CNC scheduling, overtime, inventory and finally forecasting, where the business will by then have the confidence to judge a model's advice. Upskilling still matters, and helping leaders use the tools they already have often pays back early, but it lands better after a visible win.

If you want help choosing and scoping your first pilot, get your AI roadmap. For the full method, read our enterprise AI roadmap guide.