AI agent access control is the set of rules that decide which person, through which AI app or key, can read or change which client's books, and how you see and revoke that access. For an accounting firm it comes down to five controls: who signs in, which companies they reach, whether they can write, which keys exist, and what was recorded. A prompt is not one of them.
This guide is for firm owners, operations leads and anyone writing the firm's AI policy. It explains each control, shows how it works in DeepLedger for QuickBooks Online and Xero clients, and ends with a one-page policy you can adapt. For the write controls inside one company, read the companion checklist on AI write access to QuickBooks.
Why AI access is a firm problem
One bookkeeper with one client can watch what an AI does. A firm cannot. It has many staff, many clients, and staff who use different AI apps: Claude for one person, ChatGPT for another, a script with an API key for a third. Without controls, every AI app a person connects can reach everything that person can reach, and nobody can say afterwards who did what.
The five controls
| Control | The question it answers | How it works in DeepLedger |
|---|---|---|
| 1. Identity | Who is acting? | Each person signs in to DeepLedger as themselves, from their own AI app; two-step sign-in is available |
| 2. Company scope | Which clients can they reach? | Owners and admins assign companies under Settings → Team; members reach only those, and owners and admins see every company |
| 3. Read or write | Can they change the books? | An owner or admin can set a member to Read-only for a company |
| 4. Credentials | Which keys exist, and what can each reach? | API keys act as their creator; All my companies or One company |
| 5. Record and revoke | What happened, and how do we stop it? | Settings → Audit history; revoke under Settings → Connected apps or API keys |
1. Identity: one person, one sign-in
Shared logins make every other control useless. In DeepLedger, each person connects their own AI app to https://mcp.deepledger.ai/mcp and signs in as themselves, so the AI acts with that person's access and the record shows their name. Nobody needs a QuickBooks or Xero login of their own to work through DeepLedger: an admin connects each client's ledger once.
2. Company scope: assign clients, not everything
A staff member who works on eight clients should reach those eight. Owners and admins assign companies under Settings → Team, and members reach only the companies assigned to them (owners and admins see every company), in every AI app they use.
The AI also works in one active company at a time. Every write must name its company, and if the company does not match, nothing is written. That stops a request meant for one client landing in another.
3. Read or write: decide per person and per client
Many people only need to ask questions: a partner reviewing numbers, a client's own manager, a new hire in their first weeks. Make them Read-only for that company, and the AI they use cannot write there at all.
For people who can write, DeepLedger adds checks of its own: deletes and voids need confirmation, a QuickBooks delete is refused for a reconciled transaction or one dated on or before the closing date, and in Xero new invoices, bills, credit notes and manual journals are drafts by default.
4. Credentials: scope every API key
Scripts and custom agents use API keys instead of a sign-in. Any team member can create one under Settings → API keys → Create key, covering All my companies (the default) or One company. A one-company key works only in that company and cannot switch, and every key acts as the person who created it, so it never reaches more than that person can. Keep keys in a secret store, name them after their job, and revoke the ones nobody owns. The custom agents guide covers keys in code.
5. Record and revoke
Settings → Audit history shows what the AI read and recorded in each company, with the result of each call. Owners and admins also see the trail across people, AI apps and API keys. The record holds metadata (who, when, which company, which app or key, which tool, record ids and the outcome), not amounts or report contents. See AI accounting with an audit trail for what reviewers check.
To stop access, any one of these works: remove the connector in the AI app, revoke its grant under Settings → Connected apps, revoke the key under Settings → API keys, or disconnect the ledger on the company's Connections tab.
What a prompt cannot do
"Do not change anything" in a prompt guides the AI. It does not stop a write. Access control has to sit where the AI cannot talk its way past it: in the member's access level, the key's scope and the server's own checks. The same goes for the ledger's user roles: a QuickBooks or Xero role limits a person inside the ledger, not an app connection, as the read-only access guide explains.
A one-page AI access policy for a firm
Adapt this to your firm:
- Approved apps. Staff connect only the AI apps on this list, each through DeepLedger with their own sign-in.
- Client assignment. Staff are assigned only to the clients they work on. Assignments are reviewed every quarter.
- Default access. New staff start as Read-only. Write access is granted per client by an owner or admin.
- Keys. API keys are one-company keys unless a partner approves wider scope. Each key has a named owner and lives in the firm's secret store.
- Writes. The AI prepares; a person approves adjusting entries, write-offs and anything that cannot be undone. Signed-off periods are not changed.
- Client rules. Each client's rules are written in its AI Brain, and the AI reads them before it categorises.
- Review. A reviewer samples Audit history each month for every client with AI writes.
- Leavers. On the day someone leaves: remove them from the team, revoke their keys and check Connected apps.
Frequently asked questions
What is AI agent access control?
It is the set of rules that decide which person, through which AI app or key, can read or change which client's books, and how you see and revoke that access. For a firm it covers five things: identity, company scope, read or write, credentials such as API keys, and the record of what the AI did.
Can I stop an AI agent from writing to a client's books?
Yes, if the control sits outside the prompt. In DeepLedger, an owner or admin can set a member to Read-only for a company, and the AI that member uses cannot write there, whatever the prompt says. A prompt that says do not change anything is guidance, not a permission.
Can an API key be limited to one client?
In DeepLedger, yes. Any team member can create a key under Settings, API keys, and choose All my companies or One company. A one-company key works only in that company and cannot switch, and every key acts as the person who created it, with that person's access.
How do I see what an AI agent did in a client's books?
In DeepLedger, Settings, Audit history records what the AI read and recorded in each company, with the result of each call. Owners and admins see the trail across people, AI apps and API keys. The ledger's own audit log is a second record.
How do I revoke an AI app's access?
Remove the connector in the AI app, revoke its grant under Settings, Connected apps in DeepLedger, revoke the API key under Settings, API keys, or disconnect the ledger on the company's Connections tab. Any one of these stops that access.
Related guides
- AI write access to QuickBooks: controls to check
- AI accounting with an audit trail
- How to give read-only access in QuickBooks Online
- One sign-in, every company
See how DeepLedger protects client data on the security page.