We once took cycle time out of hundreds of parts across a plant and braced for chaos. Almost nothing happened. The floor barely noticed. It worked because of everything we had put in place first: better cooling that gave the process room, control of the setup sheets, and process techs trained to optimize. Then we applied steady pressure. AI is the same. The plants that get real results from it build the system first. The ones that skip that step get a pilot that never pays off.
The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?
What the AI field has found
In 2025, a preliminary report from MIT’s NANDA initiative, “The GenAI Divide,” found that about 95% of the organizations it studied were getting no measurable profit-and-loss return from their generative AI pilots. Its authors pointed less at the AI models and more at what they called a learning gap: tools that don’t fit how the work is actually done and don’t retain feedback, dropped into organizations that hadn’t adapted around them. (The sample was small and the report preliminary, but the pattern matches what many people have seen.)
A widely used industry guide to building agents gives the same advice from the builder’s side: “Start small, validate with real users, and grow capabilities over time.”
From the shop floor: the leap
At one plant, the cycle-time gains came at the end of a lot of groundwork.
The infrastructure. We upgraded the plant’s cooling system. Better, more even cooling improves the whole process, and it gave the process techs room to work. That was probably the single biggest reason the rest was possible.
Control. We got control of the setup sheets: six to eight hundred different parts, each with its own sheet, pulled into one system, so we knew where every process stood and could control it.
Method. We trained every process tech in how to optimize a process: where the time actually is in a cycle, and how to take it out without hurting the part.
Then we started optimizing molds one at a time, highest volume first. After a few dozen parts, I looked across what we’d done. On average, we were taking about 10% out of the cycle on every part we touched. We’d been too conservative almost everywhere.
So instead of going part by part forever, we crept the targets down across the plant in a series of small steps over about three months. Every time a mold was set, the tech was expected to optimize the press. That was the pressure. They could meet it because the cooling gave them room and the training gave them the method.
I braced for the avalanche of complaints. It never came. Quality tests kept passing, good parts kept coming off the presses, and the floor barely noticed. One product family couldn’t hit the faster cycle, and we dealt with those. When my boss ran the numbers later, we had more than 5% cycle improvement, plant-wide.
It wasn’t one clever move. It was putting all the right things in place, then putting the right pressure on everything.
How it carries over
AI on top of a plant with weak infrastructure, scattered data and untrained people is the leap without the foundation. It looks impressive in a demo and stalls in production. The foundation comes first:
- Infrastructure that gives room to work. Like the cooling: equipment, data access and tools good enough that improvement is possible at all.
- Records. The data the AI agent needs, kept somewhere it can read, in a form someone trusts.
- Procedures. Written down clearly enough that a new person, or an AI agent, could follow them.
- Trained people. Who know the process well enough to land a change, and to catch the AI agent when it’s wrong.
- A way to flag problems. So when something goes wrong, you hear about it early.
Then start small: one process, one tool, one real job. Prove it there. Then apply what you learned more widely, in small steps, with steady pressure and people trained to land each one.
How you know it’s working
Keep the measure simple. If what you’re doing with AI means you’re spending less money, spending less time, or less stressed about the work, you’ve made an improvement. There’s more you can measure, but if you skip those three, the fancy metrics are theater.
And some things that look like progress usually aren’t, on their own: more AI conversations, more tokens used, the biggest model in the news, or a slide that says “AI strategy.”
Try this
Before your next AI project, write down three things: the records it will read, and whether they’re complete; the procedure it will follow, and whether it’s written down; and the person who’ll check its work. If any of the three is missing, build that first.
Sources
- MIT NANDA, “The GenAI Divide: State of AI in Business 2025” (preliminary report), as reported by Fortune, 18 Aug 2025: fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- “A practical guide to building agents” (2025), page 32: source
These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.