Almost every small business has one: a program nobody dares touch. Maybe it is an old Access database that prints the day’s route sheets, or a tool a contractor built years ago that adds up the monthly invoices. It works, mostly. The person who built it is long gone, and everyone is afraid that changing one thing will break payroll.

The newest AI tools are getting better at exactly this problem, and that changes what is worth trying. It does not change how carefully you need to try it.

What was announced

According to press reports on October 20, OpenAI announced that its latest coding models got better at older programming languages such as COBOL, Fortran and Visual Basic. These languages have much less public code and documentation for AI to learn from than newer ones, so the tools have struggled with them. OpenAI has also said it built tests for reading these languages and for moving old programs into newer ones without changing what they do.

These are OpenAI’s own numbers, on OpenAI’s own tests, and nobody outside OpenAI has checked them yet. On MultiLangBench, GPT-6 Astra scored 87.4%. On the COBOL migration test, GPT-6.1 Sol scored 86.6% at low reasoning effort and 97.9% at high effort, against 5.4% for GPT-5.6 Sol. OpenAI’s own warning, as Engadget quoted it: applying this to production software “still requires expert review and testing against the system’s business rules and dependencies.”

That warning is the whole lesson. The code is rarely the hard part. The rules hidden inside it are.

Why the rules are the risk

An old program does hundreds of small things nobody wrote down. It rounds every invoice line down to the cent. It skips weekends when it counts days. It applies a discount after a certain order size. It treats one customer’s account in a special way because someone asked in 2012.

A new version can look right and still be slightly wrong. If it rounds differently, every invoice is a few cents off and nobody notices until the month-end totals do not match.

A made-up example

Here is a made-up example. A pest control company runs an old Access program that builds each technician’s route sheet and calculates the monthly invoice for its recurring customers. The owner is tempted to have AI “rewrite it” into a modern tool.

Instead, the office manager does it in a safer order. First, she asks an AI tool to read the old program and list every rule it follows, in plain English. The list includes a rule nobody remembered: customers with two properties get a small discount on the second.

Next, she saves last month’s real route sheets and invoices as the “right answers.” Then she connects the old program’s output to the company’s accounting software, so the retyping stops and the old program keeps running. Months later, when they replace the invoice step, the new version has to produce exactly the same totals for last month’s customers before it goes live.

Four steps, in this order

  1. Have an AI explain the old program. Ask for a plain list of every rule it follows: discounts, rounding, cutoffs and odd special cases. That is the documentation nobody wrote. Have someone who knows the business read it and correct it.
  2. Save real inputs and answers. Keep a month of real invoices, reports or route sheets, exactly as the old program produced them. They become your test: any replacement must match them.
  3. Connect around it first. Get the old program’s data into your other tools so nobody retypes it, and add alerts when something looks off. The old program keeps doing its job while the manual work stops.
  4. Replace one piece at a time. Only switch a piece over when it matches your saved answers exactly. Keep the old one running alongside until you are sure.

Questions to ask before anyone touches it

  1. Who reviews the result? Someone who knows your business rules, not only someone who can read the code.
  2. What are we comparing it against? Ask to see the saved real examples, and the plan for matching them exactly.
  3. Can we go back? The old program should stay available until the new one has run correctly for a full billing cycle.
  4. What does it need access to? Old programs often hold customer and payment records. Limit what any tool can see, and keep a record of what it touched.

Your one step this week: write down the one program in your business that nobody wants to touch, and what it does for you each month. That page is the starting point for any plan, and it is worth having whether or not you ever replace it.

If you want help planning that, here is what we build.

Sources

Kush AI Automation builds in accounts you own, and nothing that moves money happens without a person approving it first. Bring the program you are most afraid to touch and book a free call.