AI and Automation

Why your AI pilot never leaves the pilot

In this article

Nori.mx7 min read

Out of every 100 companiesEvaluated60Reached a pilot20In production5MIT NANDA, July 2025Why they stall1Picked for show2Lives on another screen3Does not learn4Nothing measured first5No owner in operations

In August 2025 a headline went viral: 95% of companies get no return at all from generative artificial intelligence. It came from an MIT study, and it was quoted both to say AI is a bubble and to sell more AI. Both readings skip the most useful part of the study, which is not the number but the explanation of why pilots stay pilots.

If your plant already had an AI trial that looked good in the meeting and that nobody uses today, this article is for you. And if it has not happened yet, even better: the first project is the easiest one to pick wrong.

What the study says, and what it does not

The report is called The GenAI Divide: State of AI in Business 2025, and MIT’s NANDA project published it in July 2025. It is based on more than 300 AI initiatives that companies made public, interviews with 52 organizations and surveys of 153 senior leaders, between January and June of that year.

What it found, in its own words: 95% of organizations see no measurable effect on their results. With tools built for a specific task the drop is steep: out of every hundred companies, sixty evaluated them, twenty reached a pilot and five put them into production. The study counts as “in production” what users or executives said left a clear, sustained effect on productivity or results.

What it does not say is that AI does not work. More than 80% of organizations have tried general tools like ChatGPT or Copilot, and nearly 40% have them deployed. More telling still: only 40% of companies bought an official subscription, but at more than 90% people use AI on their own for work. AI did get into companies, just through the side door and without changing the processes.

It is worth reading with its own caveats. The authors present it as preliminary findings, say their figures point in a direction because they come from interviews rather than official reporting, and acknowledge that each company defines success differently. It is a good map, not an exact measurement.

One of the quotes in the report comes from the COO of a mid-market manufacturer, and you have probably heard it in your own plant: “The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted. We’re processing some contracts faster, but that’s all that has changed.”

Why they stall

The study says the problem is almost never the model, regulation or a lack of talent. It is that the tool does not get into the daily work and does not learn from it. Seen from a plant, that turns into five mistakes that keep repeating.

1. People pick what looks good, not what saves. According to the study, AI budgets go mostly to sales and marketing, because results there are easy to show off, while the best return is usually in back-office operations, where nobody sees it. At a maquiladora, the website assistant looks better in the meeting than the person who keys into the system, by hand, the purchase orders that arrive as PDFs. The second one is what saves hours every day.

2. The pilot lives on a separate screen. The demo works in its own window, but the work still goes through the same ERP, email and spreadsheet as always. If using the tool means copying and pasting between two systems, people drop it as soon as the pilot ends. The report describes it as brittle workflows and tools that do not fit day-to-day operations.

3. It does not learn from corrections. This is the core of the study: most of these systems do not keep what they are corrected on or improve with use. If the supervisor fixes the same mistake every Monday, by the third week they stop fixing it and do the job themselves. Corrections have to feed back into the system, and that is the reason for having the agent prepare and the person sign during the first weeks: every signature with changes is a lesson.

4. Nobody defined what “it worked” meant. A pilot without a prior measurement ends in an opinion. If you do not know how many minutes each document takes today, or how many errors come out each week, at the end of the pilot nobody can show it got better, and what cannot be shown does not get a budget.

5. It has no owner in operations. According to the report, companies that make it do better when the project is pushed by the manager whose area’s work changes, not by a central innovation team. A pilot that belongs to IT stays in IT.

There is good news for mid-sized companies, like many of the suppliers in the region: the study found they move faster. The best of them went from pilot to full implementation in about 90 days, while large enterprises took nine months or more.

How to pick the first one

For the first project to reach production, it should meet all of this:

  1. Work that already exists and is done by hand every day. Keying into the ERP the purchase orders or packing lists that arrive as PDFs, matching invoices against orders, sorting supplier emails, preparing the shift handover report. Without volume there are no savings to measure.
  2. Something you can measure before you start. Two weeks of writing down how long it takes and how many errors come out is enough. That is the line it will be compared against.
  3. Mistakes that are cheap and visible. At first, a person reviews everything before it goes into the system. A misread field on a purchase order gets caught in review; a misread field on a customs entry can end in a fine.
  4. It lives where people already work. The result has to show up in the ERP or in email, not on another screen someone has to open.
  5. An owner in operations and an exit date. Someone in the area decides whether it worked, with a rule agreed from the start: if in so many weeks it does not go from so many minutes to so many, it gets switched off and nothing is lost.

And some things should not be the first project: the website assistant “because everyone has one,” an agent that takes over the whole purchasing process, or anything where a mistake costs a shipment without a person checking it first.

A caveat about us

The study also found that projects done with an outside partner reached production about twice as often as those the company built on its own: 67% against 33%. We are an outside partner, so that number suits us, which is why we give it with the caveat the authors themselves give: these are outcomes reported by the people interviewed, and other factors may play a part. What we do stand behind is what sits underneath it: an AI project works when it fits how your plant works, not when your plant bends to fit the tool.

Where we come in

That is why we start small and measurable. On a twenty-minute call we pick the process with you, and then there is a free prototype you test with a sample of your own documents. If you need one, we sign an NDA before you share them. If the prototype does not convince you, that is where it ends. If you want to know what it would cost to run afterwards, we already covered that in what an AI agent actually costs. And if you do not know yet which process goes first, Chira asks you the questions in five minutes and leaves you a document with what to fix before that.

References

  • The GenAI Divide: State of AI in Business 2025, MIT NANDA — Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, July 2025. Preliminary findings from research conducted January to June 2025. It is the source of the 95%, the drop from 60 to 20 to 5%, personal AI use at more than 90% of companies, the 90 days for mid-market firms and the 67% against 33%, along with the authors’ caveats on method.

Related service

AI and Automation

Automate the repetitive work and see what is coming.

See the service

Articles on this service

WhatsApp · 9 p.m.Do you machine6061 aluminum?Could you sendthe drawing?Complete requestDrawing and rev.Material, finishVolume and lotTimingQualityShip-to, freightWho is askingNever saysPricesDatesCapacityA person confirmsLaura, by 5

AI and Automation

Your sales chatbot should not quote

What a sales assistant does well at an industrial company: it answers what is public, completes the quote request and hands it to a person. Never prices or dates.

9 min read

DRAFTQuoteagent preparedPriceA person signsWhen to let goWhat can be undonefirstWhat gets seenlaterPrice, date, authorityneverAcceptable correctionsInternal1 in 10Customer1 in 50

AI and Automation

The agent prepares, the person signs

How to test an AI agent before trusting it with work: what to measure, when to loosen the leash, what it never signs alone and what happens when it slips.

10 min read

MODEL COST · USDInput · 15,000 tokens0.030Output · 1,500 tokens0.015Per quote0.045440 a month × 3 calls≈ $48 a monthBy hand, 25 min each183hours a monththe same 440 quotes

AI and Automation

What an AI agent actually costs

It is not billed per user per month, it is billed per unit of work. Where the money goes, the three discounts almost nobody uses, and the part no vendor invoices.

10 min read