Trust Accounting AI: TrustBooks vs Clio vs Custom Agent

By Jude Lee · September 8, 2026 · Comparison

Lawyers and a bookkeeper reviewing trust account reconciliation reports on a laptop in a small law firm office

The rules set the boundary before any tool does

Start with what you’re actually obligated to do. Under ABA Model Rule 1.15 (Safekeeping Property), a lawyer must hold client and third-party funds separate from the lawyer’s own property, keep complete records of those funds, and retain records for a period after the representation ends. The ABA’s Model Rules on Client Trust Account Records go further, setting out the specific records to maintain and the expectation of periodic reconciliation of the trust account against bank statements and individual client ledgers.

Critically, these are model rules. Your obligations come from your state’s version — and states differ meaningfully on reconciliation frequency, whether interest goes to an IOLTA program, and what triggers a duty to report. Confirm the specifics with your state bar’s rules and any published ethics opinions before you design an automation around them, and involve your firm’s ethics counsel or a qualified professional on anything with disciplinary stakes.

Why generative AI is the wrong tool for the ledger itself

Three-way reconciliation — bank balance, book balance, and the sum of individual client ledgers, all agreeing — is arithmetic with a fixed rule set. It has one right answer. Software that computes it deterministically will be right every time; a language model asked to “reconcile these” will be right most of the time, which is worse than useless in a compliance context because you now have to verify all of it anyway.

This is the clearest illustration of the distinction we’ve written about between generative AI and agentic AI in law firms. Generating text about your trust balances: bad idea. Taking a multi-step action around a deterministic system of record — read the ledger, compare it to a rule, flag the exception, draft the notice, route it to a human — that’s the agentic pattern, and it fits here well.

An agent should never move a dollar. Its job is to make it impossible for a dollar to go unnoticed.

The three realistic options for a small firm

Built-in trust features in your practice management system

Most platforms in this category — Clio, MyCase, Smokeball and similar — advertise client trust ledgers, trust-specific reports, and guardrails against common billing errors such as paying an invoice from an unfunded retainer. Treat every one of those as a claim to verify in the vendor’s current documentation and a trial account, not as a fact. Advantages: no new vendor, trust data sits next to matters and invoices, familiar to staff. Watch the sync path to whatever general ledger your accountant uses.

Dedicated trust accounting software

Products like TrustBooks and LeanLaw exist because some firms outgrew the trust reporting in generalist software or needed tighter alignment with QuickBooks and their bookkeeper’s process. They tend to be more opinionated about the reconciliation workflow itself. The tradeoff is another subscription, another integration, and a second place where matter and client data must stay in sync.

At the product level, the differences worth checking before you demo anything (all of this changes; confirm against each vendor’s own documentation):

Clio’s trust module (inside Clio Manage): reconciliation happens where your matters and invoices already live, so there’s no second data set. Reports are generated from the same ledgers your billing runs on. Best fit for firms whose trust activity is mostly retainers tied to matters already in the system, and whose bookkeeper is comfortable exporting to the accounting package.
TrustBooks: built specifically around trust compliance rather than billing, with a guided reconciliation process and reporting aimed at bar examination requirements. Check which state’s rule set its reports are modeled on and whether that matches yours. Best fit for firms whose main pain is producing defensible, audit-ready trust reports.
LeanLaw: oriented around QuickBooks Online as the system of record, with time, billing, and trust flowing into it. Best fit for firms whose accountant already lives in QuickBooks and wants one ledger, not two.

The third option is a custom compliance agent that sits on top of whichever of the above you use — the one most firms haven’t considered, because they assume “AI for trust accounting” means asking a chatbot to do the math.

What a custom trust-compliance agent actually does

The technical building block is MCP — the Model Context Protocol, an open standard for giving an AI assistant secure, governed, permissioned access to specific data and tools. Instead of pasting reports into a chat window, you connect the assistant through a defined interface that logs every call and enforces scope. We’ve covered the mechanics of connecting Claude to Clio via MCP; the trust use case is a narrower, stricter version of the same pattern.

Expose read-only trust data through a custom MCP server

Build a small server exposing exactly four or five operations: list client trust ledgers, get balance by matter, list transactions in a date range, fetch the latest reconciliation report. No write methods exist in the code at all — not disabled, absent. That’s the guarantee you’ll want to explain to a bar investigator or a malpractice carrier.
Write a skill that encodes your state's rules

A skill is a reusable, packaged instruction set that teaches the assistant to do one job the same way every time. Here it encodes your jurisdiction’s requirements: reconciliation cadence, retention period, what counts as an exception, who gets notified, what the notice says. Written once, reviewed by whoever owns compliance, applied identically every run.
Run a daily exception sweep

The agent reads every client ledger and flags: any negative balance, any matter below the engagement letter’s replenishment floor, funds sitting past the point where they should have been billed against or refunded, and transactions lacking a matter assignment.
Draft, don't send

For each exception the agent drafts the output — a client replenishment request, an internal note to the responsible attorney, an earned-fee transfer request for a human to approve and execute — and queues it. Nothing leaves the firm or hits a bank without a person clicking approve.
Assemble the monthly reconciliation packet

The deterministic software computes the reconciliation. The agent gathers outputs, writes the plain-English exception narrative for the file, and flags anything unexplained. Human signs.

A quieter benefit: staff can ask “how much is left in trust on the Alvarez matter and when was it last topped up?” and get a sourced answer instead of interrupting the bookkeeper. Whether that matters is measurable — count the interruptions for a week before you decide.

What the agent option costs you

The first two options cost a subscription and some configuration. This one costs more, and the depth of the build guide above shouldn’t disguise that. Expect a narrow read-only MCP server plus one skill to take weeks, not days, including the rule research. Someone must own it: when your practice management vendor changes an API, the sweep can fail silently, so you need a heartbeat check that alerts if a day’s run returns zero ledgers or doesn’t run at all. Budget a recurring skill review — quarterly is a reasonable default, plus any time your state bar amends the rule — and record the review date in the skill itself. Add the assistant subscription and hosting. If nobody at the firm will own all of that, the honest recommendation is the boring option.

Modeling the value without making up numbers

We have no benchmark data on trust accounting hours, and you should distrust anyone who quotes you one. Build your own estimate:

(hours/month × loaded rate)

Recovered bookkeeping and follow-up time — plug in your own numbers

Illustrative formula, not a benchmark

(unreplenished retainers × recovery rate)

Revenue captured by catching depleted trust balances earlier

Illustrative formula, not a benchmark

Acceptable number of AI-initiated trust transfers

Editorial position

Track a baseline for a month before buying anything: hours spent on reconciliation and trust follow-up, matters that ran to zero in trust before anyone noticed, and work-in-progress sitting unbilled because the retainer was dry. In our view the third is usually the largest and least visible. Same modeling logic as our billing and AR agents piece — recovered hours plus captured revenue, from your figures, not a vendor’s.

The honest disadvantages

The real risks here are mundane, and each needs a detection method, not just a warning.

Stale or partially-synced data produces confident, wrong exception reports. Detection: make every MCP response return a “data as of” timestamp, print it at the top of each report, and have the agent refuse to run if the last successful sync is older than 24 hours. Spot-check one known balance against the system of record before acting on any report.

A skill encoding last year’s rule quietly enforces the wrong standard. Detection: version and date the skill, store the “last verified against state bar rules” date inside it, and have each run print that date. If it’s older than your review interval, the run flags itself.

Alert fatigue is the failure that actually kills these systems. Detection: log the disposition of every flag. Our suggested threshold — a judgment call, not a measured finding — is that if fewer than roughly half of a month’s flags lead to real action, stop adding rules and tune the existing ones.

Confidentiality: any AI touching client financial data implicates Model Rule 1.6 and the technology-competence expectation in Comment 8 to Model Rule 1.1. Read the vendor’s retention and sub-processor terms before, not after.

One named human owns the queue, exceptions get a same-week disposition, and the skill gets reviewed on a calendar. If you don’t have that person, don’t build the agent. Our breakdown of AI agent oversight models covers structuring that ownership at small-firm scale.

Which option is right for your firm

The durable principle: get the deterministic layer right first, then add intelligence only where the cost of a missed exception exceeds the cost of maintaining the agent.

Solo or under ~5 timekeepers, low trust volume: use your practice management platform’s built-in trust ledger and reconcile monthly with your bookkeeper. Adding AI here optimizes a workflow that takes an hour. Don’t.
Meaningful trust volume, contingency or family law, messy retainer replenishment: move to dedicated trust accounting software first.
Multiple offices, multiple bank accounts, or a compliance history you’re managing carefully: the custom read-only agent starts to justify itself, because a missed negative balance isn’t measured in hours.

The best AI program for a law firm’s trust accounting is whichever one you don’t have to babysit — and for a lot of firms that’s a boring deterministic ledger plus a disciplined monthly review. Add the agent when your exposure, not your curiosity, justifies it.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit