Why AI in a vet practice should live inside the PMS, not in a separate chat tool

MCP servers may be useful in veterinary software, but they should not be the main answer to how a clinic uses AI day to day.

Model Context Protocol gives tools such as ChatGPT and Claude a standard way to talk to another system. That can help developers and integration partners. It is less convincing as the routine workflow for a practice team that needs to ask questions, draft tasks, or summarise work without leaving the PMS.

Practice managers and clinical leads already deal with enough separate dashboards, logins, and browser tabs. If the AI plan starts by asking staff to leave the PMS and open another product, it is worth asking whether the workflow has been designed around the clinic or around the easiest technical shortcut.

What an MCP server does

An MCP server is a standard interface that lets an external model ask your software for data or request an action. The model sits outside the PMS and connects to it when needed.

That can work technically, but it often means a second window, a second tool, and a second place where staff have to decide what they can safely ask. It can also mean the AI has limited awareness of the screen the user is currently on unless context is pasted or passed across in some other way.

Why vendors keep talking about MCP

For some older systems, adding a clean in product agent is difficult. Permissions may depend on a desktop client, the database may be awkward to query safely, and the product may not have been designed with audited AI actions in mind. In that situation, an MCP server or partner integration can look like a quick route to an AI story.

That does not make it the best clinic workflow. It may simply reflect how the product is built today. Practice managers should recognise the difference between a useful integration channel and a complete answer for daily use in reception, nursing, and consult rooms.

Why the PMS is the better place for clinic AI

The PMS already knows which user is signed in, what permissions they have, which patient or client record is open, and what tools are available on that screen. That is the right place to ask for AI help. Staff should be able to stay on the consult, diary, task list, or invoice and ask for assistance in context.

If the AI drafts a message, suggests a follow up task, or summarises a history, it should do so using the same permissions the user already has. It should also leave a clear audit trail that the practice manager or clinical lead can review later. That matters for governance even when the AI task is optional.

For most practices, the practical test is simple. Can a vet or nurse use the tool without opening ChatGPT or Claude in a separate tab? If not, the workflow is probably being added beside the PMS rather than properly inside it.

The model should be behind the workflow, not in front of it

Different jobs may suit different models. A routine administrative draft may not need the same model as a more careful client communication. Equally, some practices may want AI features available only for certain jobs, and others may want them switched off completely.

When AI lives in the PMS, the model choice can remain an implementation detail managed by the vendor and the practice settings. Staff use the workflow they need, and managers retain control over what is enabled. When AI lives in a separate chat product, the choice is driven by whichever external tool the team has been told to use.

Joined up integrations matter more than a side channel

External AI tools are often proposed because the practice data is scattered. Accounts sit in Xero, payments sit elsewhere, communications sit in another system, and the PMS only holds part of the picture. In that situation, an outside agent may be used to gather answers across several sources.

A better long term arrangement is for the PMS to hold the connected operational picture through proper integrations, so the practice can ask questions in one place and get answers in context. Then nobody needs to export a file or paste information into a chat just to understand unpaid invoices or revenue differences.

The kind of questions staff should be able to ask in the PMS

The useful questions are usually practical ones that help the team run the day, complete follow up work, or understand a discrepancy.

• Why does this month's revenue differ from Xero?
• Which diabetic patients have not had a weight recorded in six months?
• Which invoices went out last week and are still unpaid?
• Who is on the diary this afternoon with an overdue vaccine?

The important point is where those answers appear. They should come back inside the system the team already uses, with the right permissions and a record of any actions taken from the result.

Where MCP still has a sensible role

MCP can still be useful as a partner or integration channel. It may help a university, hospital group, or software partner connect tools to a PMS in a standard way. It may also be helpful for one off workflows or external developer work where a published protocol avoids custom interfaces.

That is a different claim from saying MCP is the future of clinic AI. For most practices, it is more accurate to treat it as one technical option in the wider integration layer while the day to day AI experience remains inside the PMS.

Questions to ask a veterinary software supplier

• Can staff ask for AI help from the screen they are already on, without opening ChatGPT or Claude separately?
• Does the AI use the same login and permissions as the rest of the PMS?
• Is there an audit trail for actions, drafts, and task creation?
• If we use Xero, payments, or communications through your integrations, can the AI work with that context inside the PMS?
• Can we switch AI jobs on or off without affecting the rest of the system?

If the answer is mainly about connecting an external chat tool, that may still be useful in some cases, but it is not the strongest clinic workflow. For a veterinary practice, the safer and more usable approach is usually an AI layer that lives inside the PMS, uses existing permissions, and keeps every job optional.