Sell outcomes, not AI: what operations leaders are actually buying

02

Four outcomes that get budget

In manufacturing, food processing, energy and distribution businesses, four outcomes tend to attract budget. Each is measurable, owned by a named executive and visible in numbers the board already reads.

• Lower energy cost per unit of output. Energy is one of the largest controllable costs in energy-intensive industries, and many sites now run a mix of grid power, captive generation, solar and batteries that is scheduled largely by hand.
• Less leakage. Material, product and revenue go missing at weighbridges, gates, depots and billing systems. Some of it is theft. Much of it is process noise that nobody has time to chase.
• Compliance data you can trust. Tax stamps, e-invoicing, product serialisation and export traceability all demand records that agree with each other. Reconciling them by spreadsheet at month end is slow and fragile.
• Utilisation of new capacity. A new line or plant is expensive while it runs below capacity. Better planning, scheduling and quality control shorten the ramp-up.

03

Why so many AI projects stall

Analysts have repeatedly warned that many AI projects will be abandoned. Gartner has predicted that through 2026, organisations will abandon 60% of AI projects that are not supported by AI-ready data. It had earlier forecast that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, pointing to weak data, rising costs and unclear business value among the causes.

The causes are rarely the algorithm. They are usually more ordinary:

• The data does not exist, or it exists only on paper, in a controller nobody can reach or in a spreadsheet maintained by one person.
• Nobody owns the decision the model is meant to improve, so its output has nowhere to go.
• The pilot is judged on model accuracy rather than a business measure, so success cannot be turned into a budget.
• The tool sits beside the ERP and plant systems instead of inside the workflow, so people stop using it after the demonstration.

04

Foundations come first

Being ready for AI is less glamorous than it sounds. It means meters and sensors that are installed, calibrated and logged at a useful interval. It means master data for products, customers, vehicles and assets that is clean enough to join. It means events with reliable timestamps, so a weighbridge ticket can be matched to a gate pass and an invoice.

It also means a defined process. Someone must act on a dispatch recommendation, an exception alert or a forecast, and there must be an agreed way to record what they did. Without that loop, even an accurate model produces reports nobody reads.

This groundwork is often the largest part of a first project, and it pays back on its own. Sub-metering, clean master data and reconciled records are useful whether or not a model is ever added. A good partner will tell you when that is all you need.

05

How a paid pilot with agreed measures works

A pilot should answer one commercial question: if we scale this, will it pay back? To answer it, the scope must be narrow, the measures agreed in advance and the work done on live operations rather than a curated sample.

A typical pilot runs for 8 to 16 weeks at one site, line, gate or business unit, in four stages:

• Agree the outcome and the baseline. Define the measure, how the baseline is calculated, which data sources count and who signs off the result.
• Audit and connect the data. Confirm access, close gaps such as missing sub-meters or unsynchronised clocks, and build the minimum integration needed.
• Build and run inside the workflow. Deliver recommendations, alerts or automation where people already work, starting in advisory mode wherever safety or revenue is at stake.
• Measure and decide. Compare results with the baseline over the agreed period, then scale, adjust or stop.

06

Why paid, and what good measures look like

Free proofs of concept feel low-risk but often fail for a simple reason: neither side is committed. A modest fee secures access to data, time from operators and an executive sponsor who will make the scale-up decision. It also keeps the supplier honest about what can be delivered.

Good measures are specific and hard to argue with. Examples include energy cost per unit of output against a baseline adjusted for production and tariffs, dispatch exceptions detected and confirmed, forecast error against the current planner's numbers, and hours of manual reconciliation removed each month. Agree the decision rule as well: which result justifies scaling, and which result means you stop.

07

Questions to ask any AI supplier

Whoever you work with, a few questions separate a serious proposal from a demonstration:

• Which line in our accounts does this change, and by roughly how much?
• What data do you need, and what happens if it is incomplete?
• Who in our organisation acts on the output, and how is that recorded?
• How does it connect to our ERP and plant systems, and who maintains that integration?
• How will we measure the result, and what happens if the pilot misses?

08

Where we start

At TashIT we start with the outcome and choose the simplest technology that will move it. Sometimes that is a machine-learning model. Often it is sub-metering, a reconciliation engine or a workflow that removes manual steps. Our energy monitoring and dispatch optimisation, gate and weighbridge vision and utility loss analytics solutions are each designed around a pilot with measures agreed before work starts.

If you have a problem with a number attached to it, that is the right place to begin.