Where the hours actually go
The pain here is not that financial services work is complicated. It's that most of it is simple, repetitive, and stuck in the gap between systems. A paraplanner opens four tabs to build one suitability file. An onboarding coordinator sends the same "we're still missing your proof of address" email for the third time this week, then loses track of which of a dozen open files actually replied. An account manager at a brokerage spends a Friday afternoon manually checking which applications are missing a payslip instead of chasing the two that are actually ready to go to underwriting.
None of that work needs a qualified adviser's, accountant's, or underwriter's judgment. All of it currently uses their time anyway, because somebody has to open the file, read the document, check it against the checklist, and write the follow-up. That is the specific gap AI agents for financial services close: the reading-and-preparing layer that sits in front of every decision a regulated professional actually has to make. This page covers that agent layer specifically — for the fuller picture across reporting, automation, and voice, see AI for financial services firms.
What the system does, day to day
Set up against your practice management platform, document store, portfolio or loan origination system, and inbox, the agent runs a defined set of jobs continuously rather than only when someone remembers to ask. On a normal day it might:
- Open a new KYC file, check which documents have arrived against your checklist, and chase whatever is missing.
- Pull the account and portfolio data a reviewer needs before a meeting and lay it out in the format your team already uses.
- Draft the narrative sections of a reporting pack from current data, ready for a reviewer to check the numbers.
- Read the shared mailbox, sort what has landed, and either draft a safe reply or route it to the right person with context attached.
- Run the daily reconciliation match and surface the handful of exceptions that need a human decision.
- Pull together the evidence trail for a file under review — what was sent, when, and who approved it.
Every one of those is preparation, not decision. The system never tells a client what to do with their money, never approves a loan, and never sends a recommendation that a qualified person hasn't seen first. That boundary is fixed in how the system is built, not left to good judgment on a busy day.
Three workflows, walked through end to end
Three of the jobs above are worth walking through in full, because the shape of each one — and exactly where the human sign-off sits — is the part that actually matters.
Chasing KYC documents for a new client
A new client is added in your CRM and the checklist for their file type — individual, joint, trust, corporate — attaches automatically. The agent reads what has already been uploaded, checks each document against the checklist (a passport that's expired, a utility bill with the wrong address, a proof-of-source-of-funds that never arrived), and works out exactly what's outstanding.
For anything missing, it sends the chase — by email, or by phone for firms that onboard that way — in your firm's tone, referencing the specific document rather than a generic reminder. It logs every attempt and every reply against the file, so nobody has to ask "did we already chase this" in a client meeting.
When the file is complete, it doesn't wave it through. It flags the file as ready for review and hands it to the onboarding team member responsible for that client type, with everything gathered in one place instead of scattered across an inbox. A person opens the file, checks it, and makes the call on whether the client is onboarded. The agent shortens the distance between "client signed up" and "someone qualified can actually look at this," and stops there.
Getting an adviser ready for a review meeting
Ahead of a scheduled review, the agent checks the calendar, identifies which client the meeting is with, and starts assembling what the adviser will need: current holdings and valuations from the portfolio system, any outstanding actions from the last meeting, correspondence since then, and anything flagged as changed — a new address, a change in circumstances noted in the CRM, a support ticket that's still open.
It drafts a short pre-meeting brief in the format your advisers already use, with the source of every figure attached so nothing reads as unexplained. If something looks like it needs the adviser's attention before the meeting — a large withdrawal, an unanswered query, a document that never came back from onboarding — it's called out at the top rather than buried in the detail.
The adviser reads the brief, not the four systems it came from. They still decide what to raise in the meeting and what to advise; the agent's job ends at getting a complete, sourced picture in front of them with enough lead time to actually read it.
Drafting the quarterly reporting pack
On the practice's reporting cycle, the agent pulls current valuations, performance, and transaction data from the portfolio or accounting system for each client due a pack. It builds the standard sections — holdings, performance summary, transactions, fees — directly from that data, and drafts the narrative commentary in the firm's usual tone, flagging anywhere the commentary can't be written confidently from the data available.
The draft pack goes to the reviewing adviser or manager, not to the client. They check the numbers against source, read the commentary, adjust anything that doesn't sound like the firm, and approve. Only after that sign-off does the pack go out. The agent removes the hours spent assembling near-identical packs across many client files from several systems; it does not remove the adviser reading and approving each one before a client sees it.
Inbox triage, reconciliation exceptions, and audit evidence
The same pattern — read, prepare, flag for a person — covers the rest of what tends to eat a financial services team's week.
A shared mailbox that never gets ahead
A firm's shared inbox is rarely one kind of message. It's status checks, document uploads, address changes, complaints, and genuine advice questions, all sitting in the same queue. The agent reads what comes in, classifies it, pulls the relevant account or case context, and either drafts a safe reply — a status update, an acknowledgement, a request for a specific document — or routes it straight to a person with the context attached. Anything touching money, advice, or a complaint goes to a person by default; it is never a judgment call the agent is trusted to make on its own.
Reconciliation exceptions
Matching client money, fees, or transactions across a portfolio system, an accounting ledger, and a custodian's records is exactly the kind of work that is fine in volume and painful in the exceptions. The agent runs the match, and where two records don't agree, it surfaces the mismatch with both source records attached rather than a bare "doesn't match" flag. A person resolves the genuine exceptions; the agent's contribution is making sure that's the only work left for them.
Compliance evidence and audit trails
When a file is pulled for review, the useful question is rarely "what happened" — it's "can we show it." The agent assembles what already exists across your systems into one evidence trail for the case in question: what was sent, when, on what basis, and who signed off at each step. That's assembly of a record, not compliance judgment, and it doesn't replace your firm's own review process — the same approach we take to compliance reporting automation elsewhere: the system organizes what already exists so the person responsible for judging it can do that faster.
Integration notes
None of this works from outside your systems, so every build starts with what your firm already runs — typically a CRM or practice management platform, a document store, a portfolio or loan origination system, and email.
Access stays narrow. The agent gets read access where reading is enough, and write access only for the specific action a workflow needs — sending a chase email, updating a checklist status, drafting a document. It never gets broader access than the job requires, and access follows the same client-confidentiality rules that already govern your team.
Every action is logged. A chased document, a drafted pack, a routed email, a flagged exception — each one is recorded with what triggered it and what it produced, so a review can trace the trail rather than take the output on trust. Explainability and records retention aren't added at the end; they're part of the design brief from the first conversation.
The sign-off boundary is built in, not assumed. This is the point worth reading directly: the agent does not make advice, suitability, credit, or investment decisions, and it never sends a recommendation to a client that a qualified person hasn't reviewed first. Drafting, preparing, and flagging are the agent's job. Deciding stays with your staff. That's the same human-in-the-loop principle behind every AI agent Calfy builds, made concrete for a regulated firm rather than left as a general promise.
Not everything here needs an agent. A reminder sequence, a fixed document assembly step, or a routine status sync is often better handled by plain workflow automation for financial services firms that follows a set path rather than making any judgment calls at all. Part of scoping the work is telling you which of your processes actually needs the heavier tool.