What it takes to make AI work in your business
AI is the part that reads the order. What makes it safe to trust is everything around it: your data, your rules, approvals and the odd cases. Here is where AI earns its place, and where plain software does the job better.
Every conversation about automation now starts with AI. It is a reasonable place to start: the models are genuinely good, and a year ago some of what we build would not have been practical.
But when we look at the finished systems, AI is rarely the hard part. On a typical order-entry project, the code that calls a model is a few hundred lines. The code that checks the order against your customers, parts, prices and rules, routes it to the right person, enters it into the ERP and handles everything that goes wrong is many times that. That is not a failure of AI. It is what a working system looks like.
Where AI earns its place
AI is good at the parts of a business that look like language: inputs that vary, are written by people, and do not follow a template.
- Reading documents in layouts you have not seen. Purchase orders from forty customers, each with its own PDF. Supplier quotes. Delivery notes photographed in a cab.
- Turning rough notes into clear ones. A technician's three-word job note into a summary a customer can read.
- Sorting and routing free text. Which queue does this email belong in? Is this return reason about size, colour or damage?
- Finding what is missing. This request for a quote has no material specified and no delivery date.
In each case the alternative is a person reading, or a template that breaks the first time a customer changes their layout.
Where plain software is better
If a rule can be written down, write it down. Orders over $10,000 go to the sales manager does not need a model. Neither do price lists, stock levels, tax, delivery dates, discounts or arithmetic.
Plain software does these the same way every time, costs almost nothing to run, can be tested completely, and never surprises you. A model asked to apply a price rule will get it right most of the time, and most of the time is not good enough for an invoice.
If a rule can be written down, write it down.
Everything around the AI
- Your data Customer records, part numbers, price agreements, stock. AI reading a purchase order is only useful if there is a clean list to check the result against.
- Your systems The ERP, the store, the accounting package, the field app. Connecting them reliably, with retries and logs, is usually the largest single piece of work.
- Approvals Who checks what, and when. A person always decides anything with consequences: prices, refunds, what ships.
- Exceptions What happens when something does not fit. This is where systems usually break, and where most of the design time goes.
- Measurement How you will know it works: what was read correctly, what needed a person, what went wrong. Without this, nobody trusts the system for long.
An example from the order desk
A contract manufacturer receives purchase orders by email. Here is how the work splits once it is automated:
- The email arrives and the attachment is saved. Plain software.
- The PDF is read: customer, lines, part numbers, quantities, prices, dates. AI.
- Each line is checked: does this part exist, does this price match the agreement, is this date possible? Plain software.
- Anything doubtful is highlighted for a person, who approves or fixes it. A person, on a screen built for it.
- The approved order is entered into the ERP and a confirmation goes back. Plain software.
One step in five uses a model. It is an important step, because it replaces the typing. But the reason the team trusts the result is everything around it.
AI, rules or a person?
| If the step... | Use | Because |
|---|---|---|
| Reads varied documents, emails or free text | AI | Inputs vary and change without notice |
| Follows a rule you can write down | Plain software | It is cheaper, faster and always the same |
| Needs judgment, or carries real risk | A person | With a screen built for the decision |
How to use this when you are buying
When someone proposes an AI project, ask three questions. What exactly does the model do, in one sentence? What checks its output before anything happens in the real world? And what happens when it is unsure?
A good answer is short and specific: it reads the PDF; every field is checked against your part list and price agreements; anything that does not match goes to the order desk with the doubtful field highlighted. A vague answer usually means nobody has thought about everything around the AI yet.
Questions people ask
- Do we need our own AI model?
Rarely. Most business documents are read well by existing models; the work is in the checks and the screens around them.
- How do we know the AI is right?
Measure it on a few hundred of your real documents before going live, and keep a person approving anything that matters.
We start every project by mapping how the work really runs. The AI is easy to add once the rest is right, and useless if it is not.
Canonical URL: https://ugibo.com/blog/making-ai-work-in-your-business