Finance professionals already know how this goes. If it is not documented, it did not happen. That rule does not change when the work is done by an AI agent. It gets stricter.
Most advisors and accountants I talk to want the help and do not trust the tools. They are right to hesitate. The reason is laid out in Why SAIN is building the trust infrastructure for AI. This piece covers what it means for a small advisory or accounting practice.
The work an agent should take
Leave the advice to the advisor. Give the agent the work around it.
- Collecting documents from clients, and reminding the ones who forget
- Scheduling reviews and preparing the notes for them
- Drafting client updates for your approval
- Keeping month-end and tax-season checklists moving
- Following up after every meeting so nothing slips
A two-person practice in March knows what that list is worth.
The question you will be asked
One day a client, an auditor, or an examiner asks what your AI did on an account. There are three possible answers.
The first is "I do not know." That ends the conversation badly.
The second is a chat history and a confident explanation from the agent. That is the agent describing itself. It can be wrong about what it did, and anyone with access could edit it.
The third is a tamper-evident audit trail: every action, every approval, every draft, in order, where any later edit shows. That answer ends the conversation well.
Try all three on one account, then make a quiet edit and see which answer catches it:
Privacy is the other half
You cannot hand over client data every time someone wants proof, and you should not have to. A good record holds the facts about the work (who drafted, who approved, when it went out) and a fingerprint of each document instead of the document. The reviewer checks the facts. If a specific letter is in question, the fingerprint shows whether the copy you hold is the one that was approved. Every other client file stays with you.
Questions to ask a vendor
- What can the agent send to a client without my approval? The answer should be nothing.
- Show me the full record for one task.
- Who can edit that record, and what would I see if they did?
- When the agent does something out of the norm, who finds out, and does it wait for a person?
- Where do client files and prompts go, by name?
- Does it replace my planning or accounting software? It should work alongside it.
How to start
Pick one workflow with low risk and high annoyance. Document collection is the usual winner. Before you turn anything on, write down:
- The documents you need from each client, and the deadline
- Who approves outgoing messages (one name is best)
- What the agent never does: give advice, move money, or change an account
Run it with approvals on everything for two weeks. Then read the record the way an examiner would. If you cannot explain an entry, fix the instructions before you add a second workflow.
Aura does the work in iMessage, where you already are, and a new agent sets up in under 10 seconds. Anchor'd is our tamper-evident record. More detail is on our financial services page and our accounting page. For files that should never leave the office, see Privacy, transparency, intelligence.
If you want to think through which workflow to start with, we are glad to listen first.