Read-only by default · Per-agent permissions · Every query logged
Your advisers are already asking Claude, ChatGPT, and Copilot to prep meetings and answer client questions. If those agents ever touch your CRM as it stands today — duplicates, unlinked households, half-empty fields — they will answer confidently from bad data, with no record of what they saw. agentdatasync cleans and normalizes your client data once, then exposes it to the agents you choose through MCP: read-only, permissioned per agent, and with every single query written to an audit log your CCO can hand to an examiner.
The Scorecard runs entirely on your device. Your file is never uploaded. How it works.
Agents read. Humans decide.
agentdatasync never moves money, never executes trades, and never gives advice. It prepares data and controls access to it. Every decision stays with your firm’s people — exactly where your CCO, and your examiners, expect it to be.
For the Chief Compliance Officer
At most firms we work with, the CCO can veto this purchase. So we wrote this page for the veto-holder first. These are the questions a CCO should ask any vendor putting AI agents on client data — and our answers, in writing.
No. Access is read-only by design. The MCP tools agentdatasync exposes — list_households, list_clients, get_client, search_notes, list_accounts — have no write path. There is no update, delete, or export-bulk tool for an agent to call, and no tool that touches a custodian, a portfolio system, or a trading system.
Yes. Every tool call, successful or failed, is appended to an audit log the moment it happens: which agent, which user it acted for, what it asked, which records it touched, and how many results came back. The log is yours — exportable at any time, in a plain JSON Lines format your examiner can read without our software. See the Audit tier for the full record format.
Only what you permit, per agent. Permissions are granted per agent and per data domain: the meeting-prep agent can be allowed client contact details and notes, and denied account details entirely. An agent asking outside its permission set gets a refusal — and the refusal is logged too. Access is limited to authorized personnel and the agents you name, nothing broader.
You hear from us within 72 hours. agentdatasync contractually commits to notify your firm within 72 hours of becoming aware of a security incident involving your data — the term amended Regulation S-P expects you to require of your service providers (sourced: Reg S-P service-provider oversight, 17 CFR §248.30(a)(6)). That clock is built to leave room for yours: the rule requires notification to affected individuals not later than 30 days after you become aware sensitive customer information was, or reasonably likely was, accessed without authorization (sourced: 17 CFR §248.30(a)(5)). Our controls are designed to support your firm’s Reg S-P program; they do not replace it, and nothing here is a guarantee of compliance.
No. Your CRM stays. agentdatasync works from an export of the system you already run — Wealthbox, Redtail, Salesforce, or a spreadsheet — cleans a copy of the data, and serves agents from that cleaned layer. Your team keeps working in the CRM they know; the original records remain your system of record.
None that we will claim. We do not hold SOC 2 or any regulatory endorsement, and we will not imply otherwise. What we offer instead is inspectability: a published security & audit whitepaper that maps our controls to Regulation S-P and Rule 204-2 in plain language, a reference server you can read and run yourself, and an audit log format you can verify. Judge the controls, not a badge.
How it works
We take an export from your existing CRM and do the unglamorous work agents can’t be trusted to do for themselves: duplicate client records resolved, households linked, phone numbers and dates standardized, missing fields flagged in a report your ops team can act on. You approve the cleaned dataset before anything connects to it.
Start by scoring your own export with the free Scorecard — the same checks, run in your browser.
Your agents — Claude, ChatGPT, Copilot, or in-house tools — connect to the cleaned data through the Model Context Protocol, the open standard for giving AI tools structured access to data. Each agent gets its own credentials and its own permission set. MCP’s specification is explicit that the protocol itself cannot enforce consent, privacy, or tool safety — that is the implementer’s job (sourced: MCP specification, Security and Trust & Safety). The controls are our job, and they are the product.
Each agent query is written to an append-only audit log: agent identity, timestamp, tool, a hash of the arguments, the records touched, and the result count. Your CCO reviews it on a monthly rhythm, anomalies are flagged, and the full log exports cleanly when an examiner asks what your agents have been doing with client data. Retention support is aligned to Rule 204-2 — five years, the first two easily accessible (sourced: 17 CFR §275.204-2).
Details on the Audit tier page.
The audit log
This is the artifact your CCO actually cares about. One line per query, written as it happens — including the queries that were refused.
Illustrative excerpt. Firm, people, and records are fictional; the field format is the real one — see the log format specification. Note line four: a denied, out-of-scope request is a log entry, not a silent failure.
Export a CSV from your CRM and run the free Data Readiness Scorecard. It checks completeness, duplicates, formatting, household linkage, and notes quality, and gives you a 0–100 score with the specific findings behind it. It runs entirely in your browser: your file never leaves this device, and nothing is uploaded.
No account, no email required, no upload. Prefer to look first? Read the docs or the security whitepaper.