Model provider · in regular use
OpenAI API agency in Nice
The interface that gives your own applications access to OpenAI's models.
What we do with it
OpenAI API at our clients, concretely
Used on a client portal for automated document processing and reading. We always scope exactly what is sent to the model before going live.
The limits
When we recommend something else
For sensitive data or data under professional secrecy, we run the model on your own infrastructure.
Our work
Projects we delivered with OpenAI API
Paradigm
Platform for a Nice business centre: registered address, digitised mail, messaging and assistant.
OpenAI API: Analysis of incoming mail and conversational assistant.
Read the Paradigm case studyLLM Monitor
SaaS that measures how visible brands are in AI answers.
OpenAI API: Models queried and answers analysed, including in batch mode.
Read the LLM Monitor case study
Method
How we bring OpenAI API into a project
For document processing like the client portal where we use it, scoping means listing the information to extract, its format and what should happen when an item is missing. We describe that output as a strict structure: the API then returns fields the application checks before saving them, and any document that does not conform goes into a review queue.
Before anything is sent, details the processing does not need are removed or replaced with neutral identifiers, then matched back on return. OpenAI offers several models with different balances of cost and quality; we compare them on a sample of your own documents and keep the cheapest one that gets the results right. The exact model version is pinned in the configuration, so its behaviour does not change without you knowing.
The OpenAI account is opened in your name, with separate keys for test and production, stored outside the code. Usage is tracked per feature, so you can see which process drives the bill. On handover, you receive the access details, the prompts and the procedure for revoking or rotating a key.
For your business
What it changes in practice
For a small business that receives documents from its customers, reading them and keying in their contents is often quiet but constant work. Once the API takes that on, your staff only step in for files flagged as incomplete or doubtful. Your customers, for their part, see their upload dealt with without waiting for someone to be free to open it.
You get access to recent models without buying servers or hiring a specialist, and you pay according to the volume actually processed, with no fixed licence to carry. There are two trade-offs: your data passes through a foreign provider, and you depend on its commercial decisions. That is why the decision to use OpenAI is made process by process, according to what the documents contain, and why the integration stays replaceable. Staff have nothing to learn about the tool itself: they find the information already entered and concentrate on the exceptions.
Our reading of it
What we take away from OpenAI API
The OpenAI API is often where an AI project starts, for a practical reason: the models are documented, client libraries exist in every language, and structured JSON output lets you attach a model to existing processing without writing a brittle parser.
We systematically isolate those calls behind a layer of our own. The motive is leverage: a vendor can change prices, deprecate a model or alter its behaviour, and the application must be able to move to the Claude models or to Mistral AI without being rewritten. Where the data is hosted is a question too, and we raise it before the first line is written: what you hand to a model should be a decision taken, not a consequence discovered.
Decision guide
Should you choose OpenAI API?
The OpenAI API or an open-source model hosted on our own infrastructure?
The OpenAI API gives access to capable models with no servers to buy or maintain, and billing that scales with use. It suits cases where the documents processed hold nothing you are unwilling to let leave your perimeter. For health data, client files under professional secrecy or strategic information, we run an open model on your own infrastructure: the upfront investment is heavier, but no data passes through a third party.
The OpenAI API or ChatGPT subscriptions for our staff?
They serve different purposes. A ChatGPT subscription lets each person write or rephrase by hand in a chat window. The API lets your own software call the model automatically: read a document uploaded to a client portal, extract its information and store it in the right place. If you want processing to run without anyone stepping in, on every document received and always in the same way, the API is what you need.
Call OpenAI directly, or add a layer that lets us switch provider?
Calling OpenAI directly is simpler and fine for a clearly bounded task. The risk is lock-in: pricing changes, a model being withdrawn, revised terms. So we isolate the calls in a dedicated part of the code, keeping prompts and output formats separate from everything else. Swapping OpenAI for another provider, or for a European model, then stays a contained piece of work rather than a rewrite of the application.
Client questions
What clients ask us about OpenAI API
What data goes to OpenAI, and what do they do with it?
Only what the application puts in the request is sent: the text of a document, an instruction, sometimes just an extract. Before going live, we scope that content precisely and strip out anything the processing does not need. How OpenAI may use data depends on the plan and contract in place, and those terms change over time, so we review them with you during the project rather than promising a fixed rule.
How much does using the API cost each month?
The bill depends on the volume of text processed and the model chosen: each request is charged according to the length of what is sent and returned. Handling a few documents a day stays modest; a continuous stream of large files much less so. We estimate this cost from your actual volumes before development starts, and set a spending cap so there are no surprises.
What happens if OpenAI goes down?
Like any online service, the API can suffer outages or slowdowns. We design the application so that an outage does not halt your operations: incoming documents are queued and processed once the service is back, and your users are told about the delay. For critical processing, a second provider or a model hosted on your own servers can take over.
Should the OpenAI account be in our name?
Yes, and that is how we set it up. The account, the payment method and the access keys belong to you; we work on it with permissions you can withdraw at any time. You receive the provider's invoices directly, with no mark-up in between, and you keep the usage history. If you change provider, there is nothing to transfer or rebuild.
Does the model always give the same answer for the same document?
Not necessarily word for word. A language model has some built-in variability, and a provider may update a model over time. We reduce that variability with suitable settings, an enforced output structure and a pinned model version. Before any version change, we rerun a set of reference documents and compare the output with results already approved.
Going further
OpenAI API sits within our Artificial intelligence
A technology does not make a project. The method around it decides the outcome: scoping, prototypes, short cycles, and a clean way out.
Neighbouring technologies
A OpenAI API project in mind?
The first conversation is free. We will tell you plainly whether this technology is the right one, and if not, which is.
Updated on