How to manage AI agent access across your company in 2026: 5 steps

Our sales agent prepares a brief before each customer meeting and keeps our CRM up to date as conversations move forward. With two of us, access was simple. We both knew what the agent could touch, and we saw every change it made.

That stops being true when a whole team shares an agent. This is where AI agent governance starts: someone has to decide what it can read, what it can change, and who checks its work. Here is how we manage AI agent access in five steps, using one example: a sales team with an agent that prepares customer briefs from HubSpot and updates records after each call.

• Operation policies. The agent reads freely from the tools you connect. Its changes wait for a person's approval, and risky actions stay blocked.
• Approvals. A person checks each suggestion before the agent writes it to HubSpot.
• Owners. People own the agent's instructions and decide on its approvals.
• Data rules. Keep every request on models that never train on your data.
• Access reviews. A new task, account or audience is the moment to review access again.

What you need

• A d5s workspace where you are the owner or an admin. Only owners and admins can add connectors and manage the agent's permissions for each connected system.
• A HubSpot account for the agent, or the CRM your team already uses.
• An agent with a clear goal, such as "prepare a brief before each customer call and keep HubSpot up to date after every conversation". If you don't have one yet, create an AI agent in a few minutes.

Step 1: Set AI agent permissions for each tool

Set each connector operation to Allow, Ask first or Block. The agent then reads what it needs without making risky changes.

Start with what the job actually needs. A brief needs customer records, so the agent has to read them. It never needs to archive a contact, so there's no reason to allow it.

Open HubSpot on the Connectors page and go to Tool permissions. New connections start on Allow reading: the agent reads freely and asks before it changes anything. That's a sensible default, and you can adjust each operation from there:

• Allow runs without a prompt.
• Ask first pauses until a person approves it.
• Block never runs.

Figure 01

The agent reads freely, asks before it writes and cannot archive

d5sConnectors · HubSpot

HubSpot toolssales@northwind.example

Always askAllow readingAllow all

Read automatically; ask before making changes.

Search CRM recordsAllow

Get CRM recordAllow

Update CRM recordAsk first

Archive CRM recordBlock

HubSpot still applies the account's own user permissions. Telling the agent to "focus on this customer" helps it stay on task, but it isn't a hard limit. The real limits are the account's roles and these operation policies.

Before you go further, try a simple read. Ask for one brief, then look at which account the agent used, which records it read and the tool activity in the conversation. The connector docs list every setting.

Step 2: Add human-in-the-loop approvals before the agent changes records

With Ask first on contact updates, every change waits on a card where a person checks the record and the new value.

This is where most mistakes get caught. Say a salesperson asks the agent to update a contact after a call, and it picks someone with a similar name at another company. The agent will explain its choice confidently either way. The approval card is what shows you the wrong record.

Choose Ask for approval in the conversation bar. From then on, every operation set to Ask first pauses on a card with the account, the record and the exact change. You have three options:

• Allow once runs this change. Open the record in HubSpot to confirm the result.
• Decline stops it. Ask the agent to fix the target, then review the new card.
• Always allow skips future prompts for this operation. Only workspace admins can choose it.

Figure 02

Every contact update waits for a person

d5sSales · Northwind brief

Maya

Otto, mark Dana Meyer at Northwind as an opportunity.

Otto

I found the contact. The update needs your approval.

Allow Otto to access HubSpot?

update_crm_record on sales@northwind.example · Set Dana Meyer (Northwind) lifecycle stage to Opportunity

DeclineAlways allowAllow once

An approval tells you a person agreed with the change. It doesn't tell you the change worked. In the first few weeks, open the record after each approved update and check it. You're testing the whole workflow, including the person who clicked Allow.

Once the team trusts the workflow, you can switch to Allow all, which skips the cards for Ask first operations. Blocked operations and the account's roles still apply, and some actions, like sending email, always wait for a person.

Step 3: Share the project and name owners for AI agent governance

Project sharing controls who can open the work. Named owners control how the agent behaves.

In d5s, the conversations and files behind the work live in a project. Share it with Everyone in workspace, or choose Restricted and add specific people and groups. Colleagues can then open the same conversation to check a brief or pick up where someone left off. Sharing doesn't hand over the connection's credentials, and it doesn't change any of its permissions.

Access alone isn't enough, though. Good AI agent governance means every decision has a person behind it. Before the team starts using the agent, agree who owns which decision:

• The sales lead owns the instructions and makes sure the briefs are useful.
• The CRM owner manages the HubSpot connection and its permissions.
• The reviewer checks the record and the new value on each card.
• The IT or AI lead reviews access whenever the job, the account or the audience changes.

The sharing docs explain project access and roles in more detail.

Step 4: Set AI data rules for training, retention and data residency

Keep your data on models that never train on it, and check retention and location before rollout.

Customer records don't stay in HubSpot. They also pass through an AI model, so that belongs in your review too.

Organization owners and admins can turn on Require no-training routes for everyone. Every request then runs only on models that never train on your data. On Team and Enterprise plans, the primary models d5s provides also run with zero data retention at the model provider. Your workspace still keeps its conversations and files, so the team can come back to them.

Then check where everything runs. d5s runs in our EU cloud, as a single-tenant deployment, in your own VPC or on your own Kubernetes. Whatever you pick, look at model processing and each connected service, not only where files are stored.

The security page and the security docs cover the details for each deployment.

Step 5: Review AI agent access when the job changes

A new task, a new account or a wider audience is the moment to review access again.

Agents rarely keep the job they started with. The more we work with ours, the more we hand over, and each new task usually brings a new tool. When that happens, go through each new operation again. Should it run on its own, wait for a person, or stay blocked? Try it on test records before it touches real data.

It also helps to know how to take access back:

• Stop one type of change. Set that operation to Block. The agent can still read.
• Stop all HubSpot use in a project. Remove the connection under the project's Context → Connectors. It stays removed even when the account is reconnected.
• Check what happened. Open the tool activity in the conversation, then check the record in HubSpot. Changes that already ran stay in HubSpot until you fix them there.

Organization owners and admins can also open the Audit log, which records security events such as a change of approval mode. The connection guide shows where each connection is available.

What to do next

Start small. Pick one workflow, one account and one reviewer, and get that right before you bring in the rest of the team. Write down which operations the agent may run and who reviews each change, and keep the results of your first read, approval and decline tests with it.

• Connect your tools. See the connectors that d5s supports.
• Start from a ready-made agent. Browse the agents and adapt one to your team.
• Plan a rollout with us. Book a demo and we'll set up the first workflow with your team.

Questions and answers