How we work
Connecting an AI agent to your existing systems
Integration decides the timeline, not the model. A modern CRM with a documented API is a day of work; a twenty-year-old system with a nightly CSV export is a different project entirely. You can check which one you have in about ten minutes, before committing to anything.
By Gorden WübbeAutomation, agents and search visibilityUpdated 10 August 2026The model is not the variable
Buyers spend their evaluation time on which model a supplier uses. It is close to irrelevant to the outcome. Models are commodities that all handle a booking conversation competently, and they change every few months anyway.
What decides your timeline and your price is what your systems allow. Two companies asking for the identical agent can be four weeks apart, and the difference is entirely in the software they already run.
The ten-minute check
For each system the process touches, answer three questions. You can usually get all three from whoever administers it, or from the vendor’s documentation page:
- Is there a documented API? Not “integrations” with named partners — an API you can call.
- Does it allow writing, not just reading? Read-only access produces an agent that can suggest but not act, which reintroduces the manual step you were removing.
- Can you get credentials without a negotiation? Some vendors gate API access behind a higher tier or a sales conversation. That is a cost and a delay, and it is better discovered now.
Three yeses across the board is a straightforward build. A single no does not stop the project, but it changes the estimate materially — and anyone quoting before asking these questions is guessing.
What tends to be easy and what does not
| System type | Typical effort | What to watch for |
|---|---|---|
| Cloud CRM, cloud calendar | Around a day | Rate limits on higher volumes |
| Modern VoIP phone system | A few days | Whether transfers can carry context with them |
| Industry software with an API | Days to a week | Documentation quality, and whether the API is actually maintained |
| On-premise system, export only | A different project | Whether the process can use yesterday’s data at all |
The CSV trap
When a system has no API, the workaround on offer is usually a scheduled export. It is technically possible and often the wrong answer.
An agent reading a nightly export sees yesterday. It will confirm an appointment for a slot that was booked this morning, and it will do so confidently. For anything where currency matters, a stale source does not make the agent slower — it makes it wrong, in a way that is hard to notice until a customer arrives to a double booking.
Where the data genuinely does not change during the day, an export is fine. That judgment is worth making explicitly rather than by default.
Failing visibly
Every connected system will be unavailable at some point, usually while a customer is mid-conversation. What the agent does then is a design decision, and there is only one acceptable version: say so, record what could not be completed, and notify a person.
The failure mode to design out is the silent one, where the conversation ends pleasantly and nothing was written. The customer believes they have an appointment. Nobody finds out until they turn up.
Migration is not a prerequisite
Some suppliers will suggest replacing the awkward system first. Occasionally that is genuinely right — and it is still a separate decision with its own budget and its own timeline, not a preamble to a 30-day project.
If a system cannot be connected, the better move is usually to pick a different process for the first agent. There is almost always one that touches systems you can reach today, and finishing that one buys the credibility to tackle the harder one later.
What people ask about integration
- Do we have to replace our software?
- No, and a project that starts by replacing something has already lost its timeline. The agent connects to what you run today. If a system genuinely cannot be connected, the honest answer is to pick a different process rather than to migrate first.
- What makes a system easy or hard to connect?
- Whether it has a documented API with write access. A modern cloud CRM is typically a day. An older on-premise system with only a nightly export is not an integration, it is a separate project — and the cost difference is an order of magnitude, not a percentage.
- Can you work with a CSV export?
- Technically yes, and it is usually a bad idea. An agent working from a nightly export sees yesterday, which means it will confirm a slot that was taken this morning. Where the data must be current, a stale source produces confident wrong answers rather than slow right ones.
- How do we check our systems before committing?
- For each system the process touches, answer three questions: is there an API, does it allow writing and not just reading, and can we get credentials without a vendor negotiation. Three yeses is a straightforward build. Any no changes the estimate materially.
- What happens when a connected system goes down?
- The agent has to fail visibly rather than silently. It tells the person it is talking to, records what it could not complete, and notifies someone. Silent failure is the worst outcome, because the work looks done and is not.
Read next
Not sure what your systems allow?
Bring the list of what the process touches. Thirty minutes is usually enough to say whether it is a day or a different conversation.
Book a scoping call