One of our clients ran its shop floor on paper travelers, receiving logs, inspection sheets, and cert requests, some on paper and some in spreadsheets. The first thing we built for them had no AI in it. We turned each of those forms into a digital form with fixed fields, so a heat number went into a box marked Heat Number and nowhere else.

That order surprises people who call us about AI. It follows from a simple change in cost. The cheapest part of a software project today is the software.

1. The code stopped being the expensive part

For most manufacturers we meet, custom software used to be out of reach. You bought an ERP built for thousands of companies and bent your work to fit its screens. The median ERP project in Panorama Consulting's 2025 report cost $450,000 and took nine months.

AI now writes software. A coding agent is an AI that writes and tests code on its own, with a person reviewing the result. In 2026 these agents work in sessions that run for hours, and Anthropic bills its hosted agents by the session-hour. Software development is not finished as a discipline, and a person still reviews what the agent builds. But a digital receiving form, a review screen, or a tool that fits how your quality lead works is now a small build. Our view is that it no longer needs a project the size of an ERP rollout.

2. The hard part is what lives in people's heads

When the code is cheap, the cost moves to the question the code has to answer: what exactly is the job?

In most manufacturers that have been around for decades, the answer is not written down. People do things a certain way because that is how they learned it, and they get it done under pressure. Managers often believe they know the process. Then they sit with the person who does it and find steps nobody wrote down. The spreadsheet has forty columns, and nobody can say why half of them exist. Some exist because somebody left and nobody cleaned up after them.

We call the work of fixing this the clarity bill: the time and attention it takes to get a job out of heads and off paper into steps a new person, or a piece of software, could follow. Software got cheap. The clarity bill did not, and nobody can pay it for you.

3. What clarity looks like: a decision map

Clarity sounds vague, so here is what it looks like on paper. For one job, it is a decision map: a written, step-by-step map that marks where the software may act and where a person must verify.

Take a customer email asking for a certificate of conformance. A decision map for it answers three kinds of questions:

  1. What are the steps? Read the request, find the order, pull the part, revision, quantity, and heat from the job record, fill in the template, and put the draft in the review queue.
  2. Where does a person verify? The quality lead checks every heat number against the mill cert and signs before anything goes to the customer.
  3. What happens when the answer is unclear? If the heat on the mill cert does not match the receiving log, or the email names a PO the system cannot find, the work stops and the quality lead gets a flag with the reason.

If a team cannot answer those three for a job, the job is not ready for software or AI. Our rule is plain: if we cannot reach clarity with a client, we do not deploy.

4. Plain software solves the first half

Once the map exists, a surprising amount of the fix is ordinary software. Many problems on the shop floor and in the office are solved first by digital forms and one record for each job, before any AI enters.

At the client with the paper travelers, digital forms gave the company something it had never had: it could see where every job was, who had touched it, and what was waiting. It also got traceability from the mill heat to the shipment. Clear fields did that, with no AI involved.

5. Then AI sits on top and uses the systems as tools

The order is clarity first, then software, then AI. When the AI arrives, it works the way a new employee would: it uses the forms, the record, and the review queue that already exist. It looks up the heat in the record instead of guessing it from a scan. It drafts into the review queue instead of sending. When the map says stop, it stops and flags the person the map names.

That is what keeps the AI inside the lines. Its limits are the same steps, checks, and owners your people agreed on, written down before anyone wrote code.

What to do this week

  1. Pick one request your office handles every day, such as a cert request or an order confirmation, and write its steps on one page.
  2. Mark every point where a person checks something, and write that person's name next to it.
  3. For each step, write what should happen when the answer is unclear. The blank lines are your clarity bill.
  4. Look at the result and ask whether a digital form or one shared record would fix most of it before any AI is involved.

For the full argument, read Manufacturing's ChatGPT Moment Hasn't Happened Yet: Why AI Fits Your Paperwork and How to Prepare. For how an AI employee uses a decision map on the job, read AI Is a New Hire: What an AI Employee Needs on Day One and Why Clarity Comes First

Questions people ask

If AI can write software, why does an AI project still take work?

The code is the cheap part now. The work is getting clear on the job: its steps, where a person verifies, and what happens when the answer is unclear. Without that, cheap software automates a process nobody has agreed on.

Do we need AI, or just better software?

Often better software comes first. Digital forms and one record per job fix many problems for manufacturers on their own, and they give AI something reliable to work from later.

Terms in this piece

Clarity bill
the time and attention it takes to get a job out of heads and off paper into steps a new person or software could follow.
Coding agent
an AI that writes and tests software on its own for long stretches, with a person reviewing the result.
Decision map
a written, step-by-step map of a job that marks where the AI may act and where a person must verify.
Structured data
information stored in labeled fields, the same way every time, like a heat number in a box marked Heat Number.
Flag
a mark the software puts on a value it is unsure about or that does not match another document, so a reviewer checks it first.