— OPENAI FOR BUSINESS
OpenAI for business. Where models actually work, and where they do not
— W SKRÓCIE
OpenAI models earn their place in four areas of a company: work with text and knowledge, repetitive incoming requests, analytical support, and internal assistants. ChatGPT is enough for individual work; API-based solutions make sense where the output has to land in a company system. Where a process is deterministic and written as rules, plain automation is cheaper and more reliable.
Updated: · author: Arleta Marczyńska
Four areas where this works
"Will OpenAI suit us" is too broad a question to answer. A narrower one can be answered: does the company have processes shaped the way models like. The shape is fairly consistent and comes down to four areas.
Work with text and knowledge. Documents, procedures, contracts, correspondence. The model summarises, compares versions, pulls a specific fact out of material nobody has time to read. One condition applies: somebody has to know where that knowledge sits and what state it is in.
Repetitive incoming requests. Classifying cases as they arrive, drafting a first response, preparing a case for the human who will decide. The payoff comes from the number of similar cases, not the difficulty of any one.
Analytical support. Describing what the data shows, preparing a summary, a first interpretation. The model does not replace an analyst; it shortens the road to a first draft.
Internal assistants. An entry point to company knowledge for employees. This works well in organisations where knowledge is scattered across drives and inboxes, and badly where it is inconsistent, because the model will repeat the inconsistency with full confidence.
In every one of these, quality depends on whether the model receives the right context. The mechanism that supplies it is covered in our entry on RAG, and the data condition it depends on in AI-ready data.
ChatGPT or an API solution
This is the most common fork in the road, and the one most often taken too quickly.
| ChatGPT and ready-made tools | API-based solution | |
|---|---|---|
| Who uses it | individual employees | a system, a process, many people at once |
| Where the output lands | on the user's screen | in a company system, a database, a document |
| Time to launch | days | weeks or months |
| Control over model behaviour | limited to product settings | full, on the company's side |
| Cost | per-user licences | usage plus implementation and maintenance |
| When to choose it | lifting team productivity | a repetitive process with output in a system |
A common mistake is buying licences for the whole company and treating that as the implementation. Licences lift the productivity of people who already knew how to use the tool. A business process changes only when model output reaches where people work, meaning the system, not a separate browser tab.
A separate risk of licences handed out without agreement is tools used outside the IT department's knowledge. The phenomenon is called shadow AI and it is more common than boardroom declarations suggest, as we wrote in a separate piece.
When AI is not needed
This section is here deliberately, because it comes up least often in implementation conversations.
A model is not needed when the process is deterministic. If the rule reads "an invoice above this threshold goes to the director for approval", that is a condition in a system, not a task for a model. A model will get it right in most cases, and from a system we expect all of them.
It is not needed when the data does not exist, or exists in a state nobody trusts. Deploying a model onto bad data does not fix the data; it distributes the consequences faster.
And it is not needed when the problem is the absence of a process. Automating something nobody has written down entrenches the improvisation, only cheaper and faster.
An order that works
- Pick one process that genuinely hurts and repeats often.
- Check what data that process leaves behind and whether it is usable.
- Settle what has to change and how you will know. The measure comes before the start.
- Only now choose the tool. Sometimes the answer turns out to be an integration.
- Launch on a narrow scope, with the condition for going further agreed in advance.
- Plan who maintains the solution after handover.
Step three is where projects most often come apart, because without a measure any outcome can be told as a success. We expand on this order in where to start with an AI implementation.
What next
If you want to check whether your organisation is ready for any of these scenarios, we start with an audit of processes and data, not with a tool. The scope of that work is on our AI implementation page, and ready scenarios for specific areas of a company sit under solutions.
If you are looking for a company to run it and want to know what to ask them, we put the questions together on our page about choosing an OpenAI partner.
— OUR STATUS IN THE PROGRAM
YES FOR sp. z o.o., the company behind yesfor.ai, is an OpenAI Select Partner. What this status means, and what it does not.
Frequently asked questions
- Is ChatGPT enough, or do we need an API solution?
- ChatGPT works for individual use: writing, summarising, analysing a document on one person's desk. An API solution is needed when the output has to reach a company system, repeat hundreds of times a day, or be logged and controlled. The line runs at whether the result ends on an employee's screen.
- Where do OpenAI models pay off most?
- Where people process text in repetitive ways: handling enquiries, preparing documents, searching scattered company knowledge, first-pass classification of incoming cases. The common factor is a high volume of similar cases and an output a human checks anyway.
- When does a company not need AI?
- When the process can be written as rules, the data is structured, and the output must always be identical. A script, a form or an integration between systems is then cheaper, faster and more reliable. A model introduces variance, and there is no reason to pay for variance you do not need.
- Is company data sent to OpenAI used to train models?
- The terms depend on the product and plan, so the answer belongs in the terms of the specific service rather than in a general claim. The practical rule during an implementation is different: before you settle what the vendor does with data, settle which data is allowed to leave the company at all.
- What does using OpenAI models cost in a company?
- The cost is licences for tools, model usage billed by text processed, implementation work and maintenance. Usage grows with the number of users and the length of the documents processed, so a bill from a ten-person pilot does not scale linearly.
- Do we need a partner to deploy OpenAI?
- No. A company with a team that understands its data and integrations can do it alone. A partner helps where there is no time, no experience with a comparable implementation, or nobody to say that this particular process does not need a model.
- Where should we start?
- With one process that hurts, and with checking whether the data that process leaves behind is usable. Choosing a model is the last decision, not the first.
Sources and further reading
- OpenAI for businessProduct page for the business offering.
- OpenAI Partner Network, program pageThe partner program and partner roles in implementations.
- OpenAI API pricingCurrent rates for model usage.
Who wrote this
Arleta Marczyńska, founder of YES FOR sp. z o.o., CEO of the international freight forwarder EXPORTSY, a customs agent with fifteen years in the transport industry, an expert of the European Institute of Road Transport. OpenAI Select Partner.
Read next
Let us check whether your process fits a model
An intro call takes thirty minutes. Sometimes it ends with the conclusion that an integration is enough, and that is an answer too.