aceArt · AI Agency

AI integration into existing software

Most companies do not need a new AI product. They have an application their people work with every day, and in exactly three places inside it a language model would be useful. Building that is development work, not a consulting exercise, and it is the point where we differ from most AI providers.

Get expert advice

How a model gets into your application

Technically, reaching a language model is a call over an interface, the API. You send text, you get text back. That is the easy part, and there are plenty of guides for it. The effort sits elsewhere.

  • What do you send in the first place? The prompt has to come out of the state of your application, not out of a text field.
  • What do you do with the answer? Free text is rarely what a form or a database can use.
  • What happens when the service is slow or down? A model call takes seconds, not milliseconds, and the interface notices.
  • How do you stop the cost from growing unnoticed? Every request is billed, and a careless call inside a loop only shows up on the invoice.
  • How do you check whether it got better? Without cases to compare against, every change to the prompt is a matter of belief.

Those five points are the difference between a demo and something that stays in operation. They are also the reason we build integrations ourselves instead of only recommending them.

Where we build it in

Web applications and custom software

The most common case: an application we built, or somebody else did, is meant to gain one function. Suggest text, summarise input, turn free text into structured data, sort content into the right place. Custom software development is our daily work, and the AI function is one part inside it.

WordPress and TYPO3

In editorial work the benefit shows up quickly: suggestions for meta descriptions, translations, summaries, alternative text for images, ideas for internal linking. We build websites in both systems ourselves, so we implement functions like these as a proper extension rather than a third-party plugin hooked on the side. See also WordPress agency and TYPO3 agency.

CRM, ERP and line-of-business systems

This is where the real value sits, and it is also where the work is. Usually it comes down to making unstructured input usable: filling fields from call notes, recognising line items in free-text orders, pulling numbers out of service reports. The effort almost always lies in the interfaces of the existing system, not in the model.

Which data leaves the building

With an integration this is not an abstract question. Every call sends exactly the data that is in the prompt, and a prompt is built out of your application. So we settle early which fields are allowed to travel and which get removed or replaced beforehand.

For the choice of provider there are three workable routes. Providers with processing inside the EU and a proper data processing agreement cover most cases, and Azure OpenAI is a frequent compromise between capability and contract situation. Where that is not enough, open models in your own environment come into play, at the price of more operations work. And in some cases the right answer is to use no model at all for that one field.

Keeping the running cost under control

This is the point where AI projects behave differently from anything you have bought before. Classic software costs money once in development and then in operation. A language model is billed per request, more precisely by the amount of text processed. The bill grows with usage, and if the function is well received it grows fast.

  • We work out the expected volumes before we build, from your real usage figures.
  • For recurring requests we cache the answers instead of paying for the same question ten times over.
  • For simple tasks we use smaller models. Not every summary needs the expensive one.
  • Spending limits, and a warning before a limit is reached, are part of the basic setup.

We also build in a way that lets the model be swapped out. The market moves quickly, and anyone who ties themselves to a single provider ends up building the switch a second time later.

Frequently asked questions about AI integration

How does ChatGPT get into our website or app?
Through the provider’s API. Your application sends a prompt and gets an answer back. The work sits in building the prompt out of the state of the application, bringing the answer into a usable form, and catching waiting times and outages so the interface stays workable. The connection itself is the smallest part.
Is using the API GDPR compliant, and which data leaves the building?
Exactly the data that stands in the prompt leaves the building, and you decide that together with us. For the choice of provider there are three routes: processing inside the EU with a proper data processing agreement, open models in your own environment, or deliberately no model at all for particularly sensitive fields. Which one fits depends on what is in the data.
What does running it cost?
Billing goes by the amount of text processed, so the invoice grows with usage. We estimate the cost before we build, using your real volumes, cache recurring answers, pick smaller models for simple tasks and set up spending limits with a warning.
Can we build ChatGPT into TYPO3 or WordPress?
Yes, and we build websites in both systems ourselves. In editorial work the things that pay off are suggestions for meta descriptions, summaries, translations and alternative text for images. We implement it as a proper extension, so the function survives an update of the system.
Which model should we take?
The task decides that, not the brand. For sorting and summarising, smaller and cheaper models are often enough, while difficult text comprehension is worth a large one. We build the connection so the model can be changed later, because the market moves faster than your application does.
What happens when the provider goes down?
That does happen, so a behaviour for the case belongs in the planning. Depending on the function that means a clear message instead of a hanging interface, a switch to a second provider, or a way to get the task done without AI. Which one is right depends on how important the function is.
Can we not solve this with a ready-made plugin?
For standard tasks often yes, and we will say so. As soon as your own business logic or your own data is involved, the reach of ready-made solutions ends quickly. The boundary runs where the plugin does not know your data structure.
How long does an integration take?
A clearly bounded function inside an existing application is often finished within a few weeks. What takes longer is almost always the existing system, not the AI. If an old ERP has no usable interface, that is where the bulk of the work sits.

Related topics

Custom software development

For when the application itself is meant to be built from scratch. To the software development page.

MVP workshop

To try out beforehand whether the function carries. To the MVP workshop.

AI chatbots & RAG

For when the application should answer from your own company knowledge. To the chatbot development page.

Ready for your KI project?

Tell us about your project. We'll get back to you promptly with an honest assessment.

Contact us
HOME › AI-INTEGRATION