OpenAI provides hosted language, vision and speech models over an API. We integrate them for classification, extraction, drafting and transcription — always behind a spend cap, always with an evaluation set, and always with a fallback for when the API is down.
What is OpenAI?
OpenAI's API gives you access to models you do not host, priced per token. You send text or an image, you get a response, and you are billed for both directions. There is no infrastructure to run and no model to maintain.
For most business problems that is the right trade. Hosting a comparable model yourself means GPUs, and GPUs mean a monthly bill that starts higher than most of these projects' entire API spend.
The models also handle Arabic and English competently, which is not true of every provider and matters a great deal here.
When we choose OpenAI
For general text work where breadth matters — classification, summarisation, extraction, drafting. The models are strong across a wide range without tuning.
For transcription, where Whisper is the practical default and handles GCC dialects better than most alternatives.
And when the client wants to start small. Per-token pricing means a pilot costs tens of dollars rather than a hardware commitment.
When we do not use OpenAI
When the data cannot leave the organisation. Some clients here have contractual or regulatory limits on where data is processed, and a hosted API is then not an option regardless of its quality. We say so at discovery rather than after the architecture is drawn.
When the task is simple and deterministic. Extracting a date from a consistent form is a regular expression, not a model call — cheaper, faster and correct every time.
And when cost per request at volume does not work. At millions of calls a month, a smaller self-hosted model can be dramatically cheaper, and we will do that arithmetic before recommending either.
What we build with OpenAI
Classification and routing
Incoming enquiries, tickets and messages sorted by intent and urgency and routed to the right person. Measurable, unglamorous, and usually the AI project with the clearest return.
Structured extraction from messy input
Turning a scanned invoice, a WhatsApp message or a handwritten form into fields your system can store. Validated against your own rules afterwards, because a confident wrong number is worse than no number.
Transcription with Whisper
Voice notes, calls and voice ordering turned into text and then into intent. Handles code-switching between Arabic and English mid-sentence, which is how people here actually speak.
Drafting assistants for staff
First drafts of replies, summaries and reports in your own tone, with a person reviewing before anything is sent. The productivity comes from removing the blank page, not from removing the person.
How we ship OpenAI projects
Behind a hard spend cap with alerting. The failure mode of an AI feature is not a crash, it is an invoice, and a loop calling the API without a ceiling can produce a very large one overnight.
With a cheaper model tried first and escalation only when it fails. Most tasks do not need the largest model, and defaulting to it is the single most common way these projects cost more than they should.
With responses cached where the same question recurs, and a fallback path so the feature degrades rather than the page erroring when the API is unavailable.
What OpenAI costs you
You are dependent on a third party for availability, pricing and model behaviour. Models are deprecated and replaced, and output can change under the same model name — which is exactly why we build the evaluation set.
Data leaves your infrastructure. For many clients that is fine and for some it is disqualifying, and it is a question to answer before building rather than after.
Latency is the one users notice. A model call takes seconds, not milliseconds, so the interface has to acknowledge the request immediately and stream or queue the result. A feature that simply freezes the page for four seconds reads as broken regardless of how good the answer is.
Rate limits also apply per account and bite exactly when a campaign is working. Queueing and retry with backoff are not optional extras here.
OpenAI questions we get asked
It depends on volume and which model, but for a typical internal assistant answering a few hundred questions a day it is usually tens of dollars, not thousands. We estimate it during discovery using your real expected volume and set a hard cap so it cannot surprise you.
Not on the API tier as it currently stands — API data is excluded from training by default, unlike the consumer product. We confirm the current terms as part of discovery and put the answer in writing, because this is the question compliance teams ask first.
The feature degrades and the rest of the site keeps working. We build a fallback path on every integration — a queued retry, a cached answer or a plain message — because a third-party outage should never take a page down with it.
Yes, and competently — Arabic is well supported for classification, summarisation and extraction. Quality is lower than English on nuanced writing, which is why the evaluation set includes Arabic questions graded by an Arabic speaker on your team rather than assumed to match the English score.
The cheapest one that passes your evaluation set, escalating only where it fails. Defaulting to the largest model is the most common reason these projects cost more than they should, and for classification and extraction a smaller model is usually indistinguishable at a fraction of the price.
A hard spend cap, the cheapest model that passes the evaluation set, caching for repeated questions, and a measurement of what the feature actually saves in hours. If that number does not clear the API bill, we say so and stop, which is a conversation worth having in week two rather than month six.