By Softgenix Engineering, AI Workflows Team. 6 min read,
The short answer
Give AI the steps that involve reading and interpreting messy input. Keep rules, calculations, money and anything that must give the same answer twice in ordinary code. Put a person wherever a mistake would be expensive, and measure the AI on your own examples before trusting it.
Split the workflow into steps first
Asking whether AI can automate a process is the wrong question, because a process is several different kinds of work joined together. Take a supplier invoice. Someone opens the email, reads the PDF, finds the supplier, the lines, the VAT and the total, checks them against the purchase order, decides whether anything looks wrong, enters a bill in the accounts and files the PDF.
Written out like that, the steps are clearly not alike. Reading the PDF is interpretation. Checking the total against the purchase order is arithmetic. Deciding to pay an unfamiliar supplier is judgement. Each needs a different tool.
What language models are good at
Reading unstructured input and turning it into structure: pulling fields out of documents, classifying an email by what the sender wants, matching a product description to the right item in your catalogue, summarising a long thread.
Drafting text that a person will check: replies, quotes, job reports, meeting notes.
These are the steps that used to make automation impossible, because no fixed rule could cope with every supplier's invoice layout or every customer's way of phrasing a question.
What they are bad at
Giving the same answer every time. Ask a model the same question twice and the wording, and sometimes the substance, can change. That is fine for a draft and unacceptable for a VAT calculation.
Arithmetic, dates and anything with a single right answer that a few lines of code can compute exactly.
Knowing when they are wrong. A model will state an incorrect total as confidently as a correct one, so its output has to be checked by something other than the model.
A useful test: if two people doing the step by hand would always get exactly the same answer, it should probably be code, not AI.
Put rules and people around the AI
In a well-built workflow the AI reads, code checks what it read, and a person decides anything that matters. For the invoice, the model extracts the fields; ordinary code confirms the lines add up and match the purchase order; only mismatches go to a person, shown side by side with the original PDF.
This is the same split we used in YesScope, our own planning product: AI reads the project brief and records only facts, and a deterministic engine works out the schedule, so the same brief always produces the same plan.
Measure it on your own examples
Before a workflow goes live, run it over a sample of real past documents or emails where you already know the right answer, and count how often each field comes out correct. That number tells you where a person still needs to look, and gives you something to compare against once it is running.
Keep measuring after launch. Suppliers change their invoice layouts and customers find new ways to ask for things, so accuracy drifts unless someone is watching it.