Connecting AI to your CRM without handing over the keys

The window that knows nothing about you

The question people want to ask an AI assistant is a good one. Which of my open deals have gone quiet. What did we agree with this account last quarter. Summarise everything we know about this company before the call at eleven. All three are answerable, and none of them is answerable by a model that has never seen your account.

So the workaround appears on its own. Somebody exports the open deals, pastes the rows into a chat window, and gets an answer that reads beautifully. It usually is not wrong. It is just no longer connected to anything. The export was true at the moment it was taken, the account moved on that afternoon, and there is no way to click from the answer back to the three deals it was built on.

There is a second cost, and it belongs to whoever administers the account rather than to the person asking. Customer names, contract values and notes about people left the systems you control, in a paste nobody logged and nobody approved. That decision was made by a sales rep in a hurry, which is not where it should be made.

What you can do yourself

Some of this you can arrange without help, and it is worth knowing where that runs out rather than being sold a connection you do not need.

For a one off question, an export is fine. Take the rows, strip what does not need to travel, ask the question, and treat the answer as a starting point rather than as a record. The failure mode of an export is staleness, and a one off question does not care.

For anything repeating, the next step people reach for is an API token. This is where the honesty is needed. A token identifies a user and carries that user's access with it. It is an excellent way to let a system act, and a poor way to say what a system may do, because the sentence you want to write, read deals and organisations, never touch anything else, never see salary fields, is not a sentence a token can express. What a token is and how it behaves is on the API tokens and webhooks page.

You can narrow it a little by creating a dedicated user with restricted visibility and issuing the token from there, which is a real technique and better than the alternative. It is also blunt, it is easy to widen by accident when somebody adjusts a role, and it depends entirely on your permission model being deliberate in the first place.

Where the do it yourself route stalls

Three limits, in the order teams hit them.

Scope cannot be stated. Access ends up described in terms of a person, while the thing you want to control is an action. Nobody can answer the audit question, which is not who has the token but what the token is allowed to do.

Nothing leaves a trace. When an assistant reads a set of deals, or updates one, that has to be recoverable afterwards. Without it, the first time a record looks odd the conversation turns into a search for whether the model did it, and that search has no end.

Writing is all or nothing. The moment you want an assistant to do anything beyond reading, log a note, set a follow up, the crude arrangement gives it far more than that. Most teams respond by keeping everything read only, which is safe and also gives up the half of the value that saves anybody time.

What we built for this

At Sales Surge we implement Pipedrive and build our own software around it, because the same gaps kept appearing at the end of implementations. This one arrived recently and arrived fast: teams with a well built account, an obvious appetite to ask it questions in plain language, and no safe way to let them.

Surge Connect is our own product. It is a connection between an AI assistant and your own Pipedrive, built on the Model Context Protocol, with its own authorisation step rather than a shared token. The scope is decided before anything is switched on: which objects are reachable, which fields, and whether writing is permitted at all. Each customer environment is separated from every other, and what the connection does is recoverable afterwards. It is described on the Surge Connect product page, with the rest of what we make at the Sales Surge product overview.

One thing we will not do is pretend the connection settles the whole question. Where your data travels depends on which AI provider you point it at, and that belongs in a decision you record before you go live rather than in a sentence on a website. We build the part we control, we tell you what it does, and we would rather have the awkward conversation about the rest at the start. We are a Pipedrive Platinum Partner and have been in the global top 8 since July 2026, which is a reason to take the description seriously and not a reason to skip the question.

When you do not need this

If what you want is a general assistant for drafting and thinking, with no view of your pipeline, you do not need a CRM connection and adding one only widens your exposure for nothing. The connection earns its place when the questions are about your own records.

If your data is not in a state you would defend, fix that first. A connected model answers confidently from duplicated organisations and half filled fields, which turns a quiet reporting problem into a loud one. Start with data hygiene, and with duplicate records if you know they are there. Note that finding duplicates and merging them are separate steps, and the merging is reviewed by a person before anything is joined.

And if the real need is a recurring number rather than a conversation, this is the expensive route to it. A scheduled report or a dashboard answers the same question every Monday without anybody typing, which is the territory of reporting Pipedrive cannot do on its own. The full list of what we build, including the parts we tell people to skip, is on the product page.

Questions