Adding AI to a product you already run, without a rewrite
Where language models fit into an existing codebase, and how to measure whether they are helping.
You do not need to rebuild your product to get value from AI. In most of the systems we work on, the best first step is small: pick one slow, repetitive workflow, put a model in the middle of it, and keep a person in charge of the result.
This is how we approach it, using our own garment ERP, TB-ERP, as the running example.
Start with a workflow, not a model
The question is never "where can we use GPT?" It is "which task costs us the most time for the least judgement?" Good candidates share three traits:
- High volume. The same task happens dozens or hundreds of times a week.
- Unstructured input. People read PDFs, emails, scans or photos and retype what they find.
- Checkable output. A person can quickly tell whether the result is right.
In garment manufacturing, the clearest example is the tech pack. Merchandisers receive a buyer's tech pack and purchase order, often a long PDF plus scans or photos, and retype styles, materials, colours and size quantities into spreadsheets before costing can begin.
Add AI as a side service
We rarely put model calls inside the core application. Instead, we add a small, separate AI service with a narrow contract: documents in, structured data out. The main system calls it, stores the result, and carries on as before.
In TB-ERP, the core is a .NET application and the AI service is a separate Python service. Neither needs to know how the other works. That keeps the existing product stable, lets the AI side change quickly, and means a model provider can be swapped without touching the core.
Ground the model in the document
The most important rules in our extraction prompts are about what the model must not do:
- Extract only what is explicitly present. Never estimate, calculate or invent a value.
- Leave a field empty when the value is missing, rather than guessing.
- Skip anything that is cancelled, crossed out or marked as dropped.
- Keep codes, names and sizes exactly as written.
Those rules sound obvious, but they are what turn an impressive demo into something a costing team can trust.
Cross-check before anyone relies on it
Extraction is only half the job. The other half is comparing sources. TB-ERP checks the purchase order against the tech pack, style by style and size by size, and flags any mismatch before an order is costed. A person reviews the flags and makes the call.
This is where most of the value comes from. Retyping errors that once surfaced weeks later, at sampling or shipment, now show up on the first day.
Measure whether it is helping
Before switching anything on, we agree how success will be measured, and we compare against the old way of working:
| Measure | What it tells you |
|---|---|
| Time per task | Whether the workflow is actually faster, end to end, including review. |
| Field-level accuracy | How often a person has to correct the extracted values. |
| Mismatches caught | Errors stopped before they reach costing, sampling or shipment. |
| Adoption | Whether people choose the new flow when the old one is still available. |
For TB-ERP, this approach has meant around 80% less time spent on costing, bulk orders, tracking and management, with AI cross-checking reaching 99% accuracy.
Key takeaways
- Pick one high-volume, checkable workflow with unstructured input.
- Add AI as a separate service with a strict output schema.
- Tell the model what not to do: no guessing, no invented values.
- Cross-check sources and keep a person in charge of the decision.
- Agree the success measures before launch, then compare honestly.
Have a project in mind?
Tell us what you want to build, improve or automate. We reply within one working day.