Skip to content
Odysı
See if we fit
← Guides
Guide AI agents

An AI agent and the software you already have

For two months the same question has come up in our calls and emails, from very different businesses: can an AI agent work with the software we already run? A management system whose API nobody has written down. A platform with an old web service. A tool built in-house. A vendor that charges extra for every change. Too often we answered that it depends on your developer. Partly true. Here is the full answer, with the EU rules on your data checked on 12 November 2026.

T
Updated 19 November 2026
Topic AI agents
Read 7 min
No. 01

First: what the agent needs from your system

Before asking how to connect, ask what for. An agent needs one of three things from your software:

  • To read. Stock, free slots, the status of a case, a price, a customer’s last order.
  • To write. Create a booking, a lead, a note on a file.
  • Both. Read the free slots, then book one.

Then ask how fresh the data must be. Opening hours and a price list can be a day old. This afternoon’s free slots and the units left in stock cannot.

Start with read-only. An agent that can only read cannot change anything in your system, and after a few weeks you will know which writes are worth adding. When it does write, give it the smallest permission that does the job: create bookings, not edit every booking. What the AI provider may then do with the data the agent reads depends on the plan and the contract: see what happens to your data when you use AI.

No. 02

Five ways in, from best to worst

  1. An official API. The door the vendor builds so other software can read and write. Newer systems usually offer a REST API; older ones often have a SOAP web service. Both work: age is not the problem. An undocumented API is, because it can change without notice and the agent stops working. Ask the vendor, in writing, whether it is supported and what it allows.
  2. The vendor’s own integrations. Many tools send a notice (a webhook) when something changes, or publish a ready-made connector for AI agents. Quick to set up; the risk is that you only get what the vendor chose to expose.
  3. An export or a copy. With no API, a scheduled export (a CSV or Excel file, a report emailed every night) or a read-only copy of the database. Fine for what changes once a day: catalogue, prices, past orders. Not for live stock or today’s slots.
  4. Through your developer, or the vendor’s. When one person controls the database, ask them for three things: a small endpoint that returns only what the agent needs, credentials with the least privilege that works (read-only where possible), and a test environment, so nothing is tried on live data. The risk is depending on one person, so agree in writing who maintains it.
  5. Driving the screens. Software that clicks through the interface, or reads it, as a person would. Last resort: any update to the screens can break it, and nobody warns you.

At Fixlab, the agent identifies the customer against Fixlab’s platform and opens a certification request directly through the platform’s API (Fixlab case study): the first way in, reading and writing.

The second way is becoming more common. On 10 November 2026 Anthropic announced that, where a tool’s connector supports it, you can open that tool beside the chat in Claude and ask Claude to make changes in it, approving its actions; the tools named include Asana and monday.com. These interactive connectors run on MCP Apps, an open extension to the Model Context Protocol (MCP), an open standard for connecting AI agents to other software, and Anthropic says an MCP App is not limited to Claude. Anthropic donated MCP on 9 December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI. So add one question for your vendor: do you publish an MCP connector?

No. 03

What to ask your vendor

  1. Is there an API, and is it in our plan or charged separately?
  2. Is there documentation we can read, and who answers questions about it?
  3. Is there a test environment?
  4. Can we create our own keys, per user or per use, including read-only ones?
  5. What limits apply, such as requests per minute or per day?
  6. Do you send webhooks, or publish an MCP connector?
  7. Can we export all our data, in what format and how often?

Ask for the answers by email: a yes on a call is not a yes in a contract. If your vendor charges for every change or report, the export and API answers decide whether that cost comes back every month.

No. 04

A tool you built yourselves: keep, improve or rebuild

Many businesses run on a tool someone in the team built: in a no-code tool, a spreadsheet with scripts, or with help from someone who is not a programmer. It works, and people like it. The question is not whether it looks professional. It is whether it is safe to put an agent on top of it.

  • Keep it and connect to it when someone still maintains it, there are backups and you have restored one, each person sees only what they should, and each fact lives in one place.
  • Improve it when the screens work but the data is fragile: the same customer in three places, no permissions, no backups, one person who knows how it works. Move the data to a proper database, keep the screens your team likes on top, and connect the agent to the database, not the screens.
  • Rebuild it when nobody can maintain it, every change breaks something, or it cannot do what the business now needs. Rebuild the part that hurts first, not everything at once.

We compare both paths in no-code vs custom code for internal tools.

No. 05

Having a system is not a reason against, or for

Some businesses tell us they already have a service, or are happy with their booking system, and take that as a reason not to look at an agent. Others assume an agent means replacing what they have. Neither follows. The agent sits next to your system: the booking system, the ERP or the case-file tool stays the record, where the truth lives, and the agent reads from it and, once you allow it, writes back to it.

If your software already comes with AI, test that first: see when built-in AI is enough.

No. 06

Your data is yours, and EU law says so

If your software is a cloud service, the EU Data Act, Regulation (EU) 2023/2854, has applied since 12 September 2025 (Article 50). Its chapter on switching covers data processing services, which include software as a service (recital 81). Providers must remove the obstacles that stop customers moving their exportable data to another provider or to their own infrastructure (Article 23). The contract must set out the switching rights in writing, including a notice period of no more than two months, a transition of no more than 30 calendar days (up to seven months if the provider justifies that 30 is technically unfeasible), and at least 30 calendar days to retrieve the data (Article 25). Until 12 January 2027 a provider may only charge reduced switching charges, no higher than its own direct costs; from that date, none at all (Article 29). Switching charges do not include the normal fees until the contract ends or a proportionate penalty for ending a fixed-term contract early (Article 2).

Two limits. Services mostly custom-built for one customer, and not sold at broad commercial scale from the provider’s catalogue, are partly exempt, including from the rule on charges (Article 31). And software installed on your own server is not a cloud service, so the export clause belongs in your contract. Whatever applies, ask for a full export once a year and check you can open it. More on what you own and what you keep paying for in after an AI agent goes live.

No. 07

How we do it

When we build an AI agent or a single customer record, we connect to each system by the best route it allows, start read-only, and tell you before we start if something cannot be connected. See custom CRM development on the tools you already use.

If you keep one thing: decide what the agent must read and write, take the best way in your system allows, and start read-only. Your system stays the record; the agent works beside it.

Frequently asked questions

Questions about AI agents and existing software

Can an AI agent work with old software?
Usually, yes. An old SOAP web service works as well as a modern API. With no API at all, a scheduled export or a read-only copy of the database covers data that changes once a day; live data needs an API or a small endpoint built by your developer.
Our system’s API is not documented. Is that a problem?
It is a risk: it can change without notice and break the agent. Ask the vendor in writing whether it is supported, or ask your developer for a small, documented endpoint with only what the agent needs.
Should the agent write to our system from day one?
No. Start read-only, check which answers it gets right, then add writes one at a time, each with the smallest permission it needs.
Do we have to replace the tool we built ourselves?
Not if someone maintains it, it has backups and permissions, and each fact lives in one place. If the data is fragile, move it to a proper database and keep the screens your team likes.
Can a cloud software provider charge us for leaving?
In the EU, under the Data Act, switching charges are capped at the provider’s direct costs until 12 January 2027 and banned from that date. The normal fees until the contract ends, and a proportionate early-termination penalty in a fixed-term contract, are not switching charges. Services mostly custom-built for one customer, and not sold at broad commercial scale, are exempt from that rule.

Sources, checked on 12 November 2026: EUR-Lex, Regulation (EU) 2023/2854 (Data Act), recital 81 and Articles 2, 23, 25, 29, 31 and 50; Anthropic, Your apps can now open in Claude, and Claude can work in them (10 November 2026); Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation (9 December 2025); Fixlab case study. This guide is not legal advice.

Keep reading
Your systems · your data

Not sure your software can be connected?

Tell us what you run today. We will tell you plainly which way in it allows, what to ask your vendor, and whether to keep, improve or rebuild.