Skip to content

Which systems can an AI agent work with?

An AI agent gets into your programs in three ways: through an interface, via export or through the user interface. Here is what we check before we promise anything.

The most common question in a first conversation is: "Can the agent work with our program?" The answer depends less on the program's name than on how you get to the data. This article explains the three routes, gives examples by category and describes what we check before we make a commitment.

Three routes into a system

1. Interface (API). Many programs offer a documented interface through which other software can read and write data. This is the most stable route. Access can be finely controlled, changes are traceable, and the program stays usable as usual.

2. Export and file. Some programs only hand out data as an export, for example a CSV or Excel file, or accept an import. In that case the agent works with these files: it reads the export, prepares the data and provides an import file. This is less direct, but often entirely sufficient, for example for monthly reports.

3. Operating the user interface. If there is neither an interface nor an export, an agent can in some cases operate the user interface, just as a person would with mouse and keyboard. Whether this works reliably is checked case by case. Two-factor sign-in, frequent layout changes or captchas can make this route costly or impossible.

Examples by category

The overview below names common programs as examples. It is not a list of existing integrations. What is possible in your case is clarified during the check.

Category Examples Typical route
Email and calendar Outlook, Gmail Interface
CRM HubSpot or a CRM in your environment Interface
Accounting DATEV, lexoffice Often export and import, sometimes interface
Online shop Shopify Interface
File storage Dropbox, cloud storage Interface or file storage
Team chat Slack Interface
Spreadsheets Excel File
Industry software Practice, law firm, inventory programs Varies widely, checked individually

Industry software shows the widest range. Some programs have modern interfaces, others only offer an export, and for some the user interface is the only option.

Data migration as a special case

Not every connection is permanent. Often the first job is to move data cleanly from one system to the next, for example contacts into a new CRM. Here care matters more than speed.

Two examples from practice: an online language-course provider moved 2,950 newsletter contacts without a single accidental mail. A gallery migrated 5,992 contacts and kept all 1,594 opt-outs. Opt-outs and consent in particular must not get lost in a migration.

What we check before we promise anything

A commitment without a check would not be serious. We clarify these points first:

  • Access: Is there an interface, an export, or only the user interface?
  • Permissions: Can a separate account with least privilege be set up for the agent?
  • Limits: Does the interface have rate limits or paid tiers?
  • Data format: Is the data consistent enough to process automatically?
  • Testing: Is there a test account or copy where we can try things safely?
  • Approvals: Which steps have external effect and need your approval?
  • Logging: Can every step the agent takes be traced?
  • Terms: Do the vendor's terms of use allow automated access?

Only once these questions are answered do we know which route is possible and what it costs.

Least privilege, clear approvals

Whatever the route, an agent only gets the access it needs for its task. An agent that prepares appointments sees the calendar, but not the invoices. Anything with external effect, such as sending, deleting or publishing, is approved by a person. Every step is logged.

This also applies when operating a user interface. There in particular, the agent should use its own account rather than the account of someone on your team.

Explore practical examples, plan a supported migration, or compare your costs.

FAQ

Our software has no interface. Can it still work?

Often yes, through an export or in some cases by operating the user interface. Which route works reliably is checked case by case before we commit to anything.

Is operating the user interface reliable?

It is the most fragile of the three routes. Layout changes or extra sign-in steps can disrupt it. That is why we prefer an interface or export wherever one exists.

What access does the agent get in my programs?

Only what it needs for its task, through its own account. Sending, deleting and publishing are approved by a person, and every step is logged.

Let’s talk

Get more done withthe same team?

We’re happy to think it through with you.

Rather read first? Go to Knowledge