Connecting your tools

Connecting HubSpot to an AI agent

A CRM does not go stale because it is badly designed: it goes stale because filling it in is the work nobody does. An agent connected to HubSpot reads and updates records, activities, files and marketing content — through HubSpot's official interface, and strictly under the permissions of the user who connected it. It is that last point which makes the thing governable.

1. What the agent does in HubSpot

The connector goes through HubSpot's official server, which guarantees the scope follows HubSpot's rules and not ours.

The records. Read and update contacts, companies and deals — the address that changed, the contact who left, the stage the file has actually reached.
The activities. Log what happened: the call, the follow-up, the reply received. That is the part always missing, because it is done after the fact.
Files and marketing content. Find a document, consult or prepare content attached to HubSpot.

2. The agent inherits your permissions, never more

This is the central mechanism, and it is worth understanding before connecting.

The connection is made with the HubSpot user whose permissions must be respected. The agent therefore has no rights of its own: it acts with that person's. If that account cannot see a pipeline, neither can the agent. If that account cannot delete, neither can the agent.

The practical consequence is a pleasant one: you have no new rights model to learn. You decide an agent's scope by choosing which HubSpot user it is connected with — often by creating a dedicated user with deliberately narrow rights. Each person on the team can also connect their own account and grant only their own permissions.

The connection also requires designating the portal and the object type concerned: an agent is never connected “to HubSpot” in general.

3. What stays under approval

HubSpot is classed among the high-stakes connectors, and the default setting reflects it: any action that writes or deletes requires human approval before execution, until you have explicitly relaxed it.

In practice, most teams end up relaxing the update of a record or the addition of an activity — the risk is low and the daily gain real — and keep approval on what commits: changing a deal's stage, altering an amount, deleting. That is a sound trade-off, and it is set action by action rather than wholesale.

As always, each agent's spending is capped before the model call: envelope exhausted, the call does not happen.

4. What it changes over time

On a CRM, the first month proves nothing. The twelfth week is what decides.

Week 1. Recently touched records are completed, missing activities logged. The gap between what the CRM says and what actually happened starts to close.
Month 3. The pipeline reflects reality. It is the first time you can rely on it to decide, instead of “asking whoever is on the file”.
Year 2. The history becomes data: which deals move, which sleep, which follow-ups produce. None of that is readable in a half-filled CRM.

No quantified gain is promised. What is certain is that a CRM that is kept beats a CRM rich in features — that is the subject developed separately.

5. Frequently asked questions

Do I have to create a HubSpot app?

An authentication app is created once for nullbot; each person then chooses their HubSpot account and grants their own permissions when connecting. There is therefore no configuration work to redo per user.

Can the agent change a deal without my knowing?

Not by default: write and delete actions require human approval before execution. And even once that approval is relaxed on harmless gestures, everything stays traced — in nullbot's log and in HubSpot's own history, which records on its side who changed what.

Can an agent be limited to a single pipeline?

Yes, through two overlapping routes: the connection requires designating the portal and the object type, and the HubSpot user the agent is connected with carries their own restrictions. Creating a dedicated HubSpot user with narrow rights is the simplest and clearest method.

What if we use a different CRM?

Pipedrive, Close, Attio and Folk are also in the catalogue, and nullbot has its own CRM if you have none. The principle stays the same: the agent works in the tool that is the reference for you, rather than creating a parallel database that will go stale in its turn.