# An AI Team for Your Agency: Delegate Without Abdicating

> AI assistants on CRM data can save time, but client data needs protecting. What to delegate, where a human approves, and which red flags to know.

- Published: 2026-10-06 · Updated: 2026-10-13
- Tech & automation · 9 min read · The Financi team
- Canonical page: https://www.financi.co.il/en/articles/ai-team-for-your-agency-safely

## The short version

- Delegate drafting, summarising and suggested actions to AI. Not decisions, not numbers and not silent writes to client data.
- A core principle: the AI phrases and the system owns the facts. Numbers are displayed from the source, not rephrased.
- Permissions are a restriction, not a formality: grant only what is needed, and know how to disconnect.
- Client information is subject to privacy and data-protection law. Check with legal counsel and your compliance officer before connecting any AI tool.

The promise of AI in an agency is clear: less time on summaries, searches and typing, more time with the client. The risk is just as clear: a system that phrases things confidently, even when wrong, on top of data that must not be wrong and must not leak. The right path is neither to give up nor to get carried away. It is to delegate within limits, and to build the limits before you start.

## What to delegate, and what not

| Task | Who does it | Why |
| --- | --- | --- |
| A client brief before a meeting | AI drafts, the agent reads | A starting point, not a replacement for checking |
| Proposing values for fields in a process | AI proposes, a person approves | Saves typing without silent writes |
| Answering “how do I” questions about the system | AI answers | Information about the system, not about a client |
| Spotting gaps by rules | The system shows, a person judges | A finding is a question, not a conclusion |
| Numbers (returns, balances, fees) | The system displays from the source | We don't let AI rephrase a figure |
| A pension or insurance recommendation | The agent alone | Professional responsibility and licensing |

> **The principle: the AI phrases, the system owns the facts.** In Financi the AI doesn't write numbers. The portfolio analyst “Keren” names which figure she means, and the system displays it from exactly the same source. The reason: when a language model rephrased real figures, the result drifted. That story is the lesson, and it's worth adopting even if you are building something else.

## Human in the loop: where approval must sit

The difference between a safe assistant and a risky one is usually who presses the button. If the AI writes straight into client data, every mistake is saved. If it proposes and a person approves, the mistake stops at one step.

1. **The AI proposes** — For example: “Status: waiting for documents”, based on notes in the process. The proposal is displayed, not yet saved.
2. **A person sees the proposal** — Field by field. Whoever approves understands what they are approving.
3. **A person approves or rejects** — In Financi the process assistant “Rafi” proposes values and you approve each one, with undo. Nothing is written without approval.
4. **History remains** — What was proposed, what was approved and by whom. That way you can go back and check.

## Permissions: when connecting ChatGPT or Claude

Financi has an MCP connection for assistants such as ChatGPT and Claude. You connect with a regular OAuth sign-in, and the assistant works inside your book with the permissions you grant. The tools cover leads, production, referrals and information-order links. What the assistant can do is limited by the permissions you pick on the approval screen, and you can disconnect at any moment.

**Broad, hasty permission**

- “Give everything so it works”
- No distinction between read and write
- Nobody knows which assistant is connected
- No offboarding process when an employee leaves

**Narrow, documented permission**

- Start read-only, widen as needed
- Choose per connection which permissions to grant
- A list of who connected which assistant and when
- A disconnect procedure: delete the app under Developer API

> **Rule of thumb.** Every assistant gets the minimum permission needed for the role you defined. If access to leads is enough, don't grant access to information orders.

## Privacy and protecting client data

Client details and their pension and insurance position are sensitive personal data. When you connect an AI tool to them, ask three questions: which fields does the tool see, where does the information go, and what happens if something slips.

- **An allowlist of fields.** The AI inside a Financi customer file works on pre-defined fields only, not everything on the card.
- **Testing on real data.** The defence was tested over thousands of real clearing-house files, to confirm identifying details don't pass.
- **An egress scanner.** Before text leaves, a scanner redacts sensitive details rather than failing.
- **A data processing addendum.** When a vendor processes data on your behalf, the relationship should be governed by an agreement. Financi has a data processing addendum.

> **Your procedures remain yours.** Even a tool with good protections doesn't relieve the agency of its duties under privacy and data-protection law, or the regulator's requirements for pensions and insurance. Before connecting an AI tool to client information, check with legal counsel and your compliance officer what is required: client consent, use of the data, processing location and retention.

## Red flags

**A check before connecting an AI assistant**

- [ ] We know which fields the assistant sees, and which it doesn't.
- [ ] The assistant doesn't write to client data without human approval.
- [ ] Numbers are displayed straight from the system, not rephrased by the model.
- [ ] Permission is limited to what is needed, and we can disconnect at any time.
- [ ] We keep a list of who connected which assistant, and someone owns it.
- [ ] A suitable data processing agreement is in place with the vendor.
- [ ] We checked with legal counsel and our compliance officer what is required for client data.
- [ ] Staff know what must not be pasted into an external assistant that isn't officially connected.

- A vendor promising “zero AI errors”.
- An assistant that writes straight into client data with no approval and no undo.
- No clear answer to “where does the information go?”.
- Employees pasting client details into an external chat to “speed things up”.
- One blanket permission for everyone, regardless of role.

## Frequently asked questions

### Can the AI be wrong?

Yes, always. That's why in Financi Keren doesn't write numbers and Rafi doesn't write to a field without your approval. The AI phrases, and the system owns the facts.

### Can we use ChatGPT or Claude directly on client data?

Paste client details into an external chat only if you've checked that it is permitted and protected, with legal counsel. The structured connection in Financi is built on an OAuth sign-in, permissions you choose and one-click disconnect, which differs from pasting information by hand.

### How do we disconnect a connected assistant?

In Financi you delete the app under Developer API and access stops. It's worth setting a standing procedure for this, for example when an employee leaves.

### Does the AI make pension decisions?

No. It phrases, summarises and proposes. A recommendation to a client is the agent's, under the licensing and professional-responsibility rules that apply.

### What if we suspect a data leak?

Follow the agency's incident procedure and contact legal counsel and your compliance officer immediately. Reporting duties depend on the circumstances and the law, so involve them early.

> **Want to see the AI team at work?** We'll show how Asi, Keren and Rafi work inside the CRM, and what human approval on every field looks like. [Book a demo](https://www.financi.co.il/en#contact)

## Sources and further reading

- [gov.il: Israel government services and information](https://www.gov.il)

---

*The information in this article is general only and is not financial, pension, legal or tax advice, nor a substitute for personal advice. Demos and calculators are illustrative models: your numbers will differ.*