How do you integrate AI into your existing software systems?

You do not need to replace your ERP, CRM or core systems to use AI. In almost every case the right move is to connect a model to what you already run, through an API layer, an event stream or retrieval over your own data. Uliasti, a Zürich engineering firm working with banks and SMEs, builds these integrations for a living. This is the approach we use.

Start with one workflow, not with a model

The most common false start is picking a technology and then looking for a place to use it. It works better in reverse. Pick one workflow that is frequent, rule-bound and measurable: answering routine support email, extracting data from invoices or contracts, drafting the first version of a compliance summary, triaging incoming requests.

Then measure how that workflow performs today. Minutes per case, error rate, cost per month. If you cannot state the baseline, you will never know whether the AI actually helped, and the project will be judged on impressions instead of numbers.

The three integration patterns that work

Nearly every successful integration we ship uses one of three patterns, often in combination:

• An API layer. The AI runs as a separate service and talks to your existing systems through their existing interfaces: REST APIs, database views, file exports. Nothing inside the ERP or CRM changes. This is the most reversible pattern, and reversibility is worth a lot in a first project.
• Event-driven processing. A new invoice arrives, an event fires, the AI processes it and writes the result back for a person to review. This suits asynchronous work with clear inputs and outputs, and it keeps the AI out of your critical path.
• Retrieval-augmented generation (RAG). The model answers questions using your documents and data, fetched at the moment of the question. Your knowledge stays current without training a model on your data, and access controls can be enforced at retrieval time.

What these patterns share: the existing system remains the source of record, and the AI is a layer you can adjust, swap or remove. Deep couplings and forked systems are how AI projects become legacy systems of their own.

What your data needs before integration

Not perfection. Three things are enough to start:

• Access. There must be a sane way to read the data: an API, a replica, a scheduled export. If your data is locked inside a vendor system with no interface, that is the first problem to solve, before any AI work.
• Clear permissions. Whoever could not see a document before must not be able to see it through the AI either. Access rules have to carry over into the retrieval layer, not get flattened by it.
• Enough quality for the one workflow. Fix the data the workflow touches. Do not launch a two-year data program as a precondition; that is how AI initiatives die of old age.

How regulated companies handle the risks

For banks, insurers and other regulated firms, the model choice matters less than the controls around it:

• Every input and output is logged, so there is an audit trail.
• A person reviews anything that reaches a customer or a booking system.
• Data residency is explicit: Swiss or EU hosting, or on-premise inference where policy requires it.
• There is a written rule for which categories of data may reach which model.

We build software for SIX and Aargauische Kantonalbank, and the integration pattern that passes review is boring on purpose: deterministic pipelines around a model, with defined inputs, checked outputs and a human in the loop. Not a model improvising in production. More on this on our page for regulated industries.

A realistic timeline

A first integrated workflow, including evaluation against the baseline, takes four to eight weeks. Not a year. From there you expand workflow by workflow, reusing the integration layer you built for the first one. The expensive part is the first mile: access, permissions, logging, evaluation. Every workflow after that gets cheaper.

If you want a structured starting point, our AI readiness check covers exactly these questions, and our AI, software and data engineering team builds the integration itself.

Frequently asked questions

Do we have to replace our ERP or CRM to use AI?

No. The systems you run stay the source of record. AI connects to them through an API layer, an event stream or retrieval over your data. Replacing a working core system to get AI is almost always the wrong order of operations.

Which AI model should we use?

Start with a hosted frontier model behind your own abstraction layer, so you can switch models without rewriting the integration. The abstraction matters more than the initial choice; models change every few months, your integration should not.

Can our data stay in Switzerland?

Yes. Swiss or EU hosting is available for the major model providers, and on-premise or private-cloud inference is an option where policy requires it. Data residency should be decided per data category, in writing, before the first integration ships.

How much does an AI integration cost?

A first integrated workflow typically lands between CHF 30,000 and 80,000 depending on how many systems it touches and the level of controls required. See our breakdown of what secure custom software costs in Switzerland for the full picture.