A language model is good at language and poor at arithmetic.
The useful question is never whether the OpenAI API can be added to your software. It is whether it removes work somebody is doing by hand today. If the answer is a demonstration rather than a job, the project usually goes quiet after a month.
The OpenAI API is one way to put a language model inside a system you already run. It is a paid service belonging to somebody else, and that shapes what you can promise about it.
The OpenAI API is a service your software sends text to and gets text back from, charged by the amount of text. It can read a supplier email and pull out the order lines, draft a reply in your wording, or sort tickets by subject. It knows nothing about your business unless you send your own records along with the question.
Where it earns its place.
The jobs it is genuinely good at
- Turning a message into fields your system can store
- Drafting a reply somebody reads before it is sent
- Sorting enquiries and tickets by what they are about
- Summarising a long thread or a site report
- Pulling line items out of a supplier invoice
- Answering in the language the customer wrote in
What OpenAI decides, not us
This is a supplier relationship, and the terms belong to the supplier.
- Which models exist, and when an older one stops being served
- What each call costs, and when that price moves
- What the model will refuse to produce
- Rate limits, and what happens when you cross one
- Where the request is processed, which matters if your data may not leave India
So we build so that the model can be swapped. Anything that depends on one model behaving exactly as it does this month is a system with a clock running on it.
Where it sits in a workflow
Inside a screen somebody already opens, rather than in a separate window they have to remember.
- Support and sales assistants
- Internal assistants over your own documents
- Invoice and document data extraction
- Enquiry classification and routing
- A summary at the top of a long customer record
- Quotations drafted from a salesman’s notes
The work that makes it reliable
- Name the job
- Pick a model
- Ground it
- Fix the output
- Review
- Measure
- Asking for a fixed structure, so the answer can be stored and not merely read
- Feeding it your own records, so it quotes your policy instead of inventing one
- A person approving anything that leaves the building
- A log of what was asked and what came back, for the day somebody disputes it
- Something sensible happening when the API is slow or down
What it should never be given
- Totals, tax and anything that has to add up
- A decision with a legal or medical consequence, with no person in the loop
- Data you have promised a customer stays on your own servers
- Work a plain rule would do, such as flagging an overdue invoice
A language model is convincing when it is wrong, which is the whole difficulty. Everything it produces needs either a human check or a small consequence.
When this is the right choice.
- Somebody reads the same kind of message all day and retypes it into a system
- Your material already exists in writing, in documents, tickets or a manual
- A draft a person edits is more useful to your team than a blank box
- Customers ask the same twenty questions in three languages
- You are willing to measure whether it actually saved the time
When it is not.
- Anything that must be exactly reproducible, such as a bill or a payroll figure
- Data that by policy or contract cannot leave your own servers, unless you run a model yourself
- A rule you could write down in one sentence, where a line of code is cheaper and never wrong
- A business with no appetite to review the output, because unchecked output is eventually wrong in public
Questions we are asked about it.
Does our data get used to train the model?
Under the API terms as they stand, business data sent to the API is not used for training by default. That is OpenAI’s policy rather than our promise, and policies change. If the data is sensitive enough that the answer matters, read the current terms with your lawyer before you send anything.
What does it cost to run?
It is charged on the amount of text going in and coming back, billed by OpenAI directly to you. We can estimate from the volume you expect, but the rate is theirs to change. The account stays in your name so you see the bill.
Can we use a different provider later?
Yes, and we build so that you can. The same work runs on other providers, and for some jobs a smaller model running on your own server is enough.
Will it make things up?
It can, and it will sound certain while doing it. Giving it your own documents to answer from cuts that down a great deal. A review step on anything that reaches a customer handles the rest.
Can it read our PDFs and scanned bills?
Text PDFs, yes. Scanned images have to be read into text first, which is a separate step, and it is usually the step that decides how accurate the result turns out.
Should we build this now?
Only if you can name the person whose work it removes. If the honest answer is that AI is on the board’s list this year, the free technology audit is a better use of a fortnight than a pilot.
Services that use it.
What it sits with.
Not sure OpenAI API is the right choice?
Tell us what the software has to do and who opens it. If something else fits better, we will say so, and say why.
- No obligation
- We reply the same working day
- Your details stay private