Build or Buy an AI Billing Assistant: How a Law Firm Should Decide

At a Glance

• Who this is for: Finance directors and CTOs who are being pitched AI billing tools, or being asked by a partner why the firm can't just build one.
• The problem: Bill review is slow and repetitive, so it's the first place firms reach for AI. The build-or-buy call usually gets made on licence price, which is the smallest cost in either option.
• What decides it: Who owns the error when the AI gets a client bill wrong, and whether you can prove what it did.
• The takeaway: Price the guideline rules and the people who'll maintain them before you price the software.

What is an AI billing assistant? An AI billing assistant is software that reads time entries and draft bills, checks them against client billing guidelines and e-billing formats, and suggests or makes edits before an invoice goes out. A person in the billing team or the responsible partner approves what it changes.

Why Billing is Where Firms Start

Bill review has everything that makes a task look ready for automation.

High volume, and written rules that define what counts as wrong.

Client billing guidelines spell out what a client will and won't pay for: block billing, vague narratives, travel billed at full rate, 4 timekeepers on a call that needed 1. Somebody in the billing team reads every entry against those rules, for every client, every month.

E-billing adds a format layer on top. The LEDES Oversight Committee has set data exchange standards for legal e-billing since 1995, and a bill that fails a client's format checks bounces back before anyone reads a word of the narrative.

So an AI billing assistant looks like an easy win. The demo usually is one. The decision gets harder once you ask what happens in month 7.

What Building One Involves

A working in-house billing assistant needs more parts than the prototype suggests:

• Access to a hosted model under enterprise terms your clients will accept
• A rules layer that holds each client's guidelines in a form software can check
• A connection to your time and billing system that can read entries and write edits back
• A review screen where a person accepts or rejects each change
• A log of every suggestion, every edit and who approved it
• A team that keeps all of the above working

The rules layer is where builds stall. Guidelines arrive as PDFs, email chains and a paragraph buried in an engagement letter.

Someone has to turn those into checks and keep them current every time a client revises them. That work never finishes, and it lands on whoever built the thing.

Then there's the model itself. Providers update and retire model versions on their own schedule. A prompt that behaved well on one version can drift on the next, so every change needs re-testing against real bills before it goes near a client.

The SRA's Risk Outlook report on AI in the legal market put it bluntly in November 2023: "Only the largest firms will hold enough data to train their own systems." Most in-house builds wrap a hosted model, so they skip training entirely.

The point still holds. The group of firms with the data and the engineering depth to build this well is small.

What Buying One Involves

Buying moves the engineering off your desk. The ownership stays put.

The SRA Code of Conduct for Firms says you "remain accountable for compliance with the SRA's regulatory arrangements where your work is carried out through others." A vendor tool that edits a client bill is doing work on your behalf. When it gets something wrong, it's still the firm's bill and the firm's client relationship.

Before you sign, I'd ask any vendor these 4 questions:

1. Where is our billing data processed and stored, and is any of it used to train models?
2. What does the audit log show for a single edited time entry? Ask them to show you one.
3. Can we control which clients, and which kinds of edits, the tool can touch without approval?
4. If we leave, what happens to the guideline rules we've configured?

The last one matters more than it looks. The guideline rules are the expensive part of this whole exercise. If they only exist in a vendor's format, you'll end up renting your own billing knowledge back.

Where the Cost Actually Sits

Look at the error row. Both columns say the same thing, and it's the row that should drive the decision.

5 Questions to Make the Call

1. How sensitive is the data? Check which client terms restrict where billing data can go. Some will rule out a vendor on their own.
2. How big is the problem today? Count bills per month, rejections and write-downs. If you can't produce those numbers, measure them first. You'll need them to prove anything worked.
3. Who will maintain it in 2 years? A build needs engineers who are still around after the launch.
4. Where is the delay? AI speeds up rule checking. If bills sit for a week waiting on partner review, that's a people and process issue a tool won't fix.
5. How soon do you need it? A build takes longer to reach real clients, even when it goes well.

High volume, in-house engineering and restrictive client data terms point towards building. Modest volume, a thin engineering bench and a deadline this financial year point towards buying. I'd expect most firms to land somewhere in the middle.

Governance Comes with Either Option

NIST's AI Risk Management Framework, released in January 2023, organises AI risk work into 4 functions: govern, map, measure and manage. It's voluntary and US-published, and it's still a useful checklist for a UK firm deciding what to demand from a build team or a vendor.

The Law Society's guidance, Generative AI: the essentials, makes the same point from the professional side: the solicitor is responsible for checking what AI produces.

For billing, these are the controls I'd insist on whichever way you go:

• Log every suggested edit, every accepted edit and the person who accepted it
• Name a reviewer for every bill the tool has touched
• Keep a written list of clients and matter types where AI edits aren't permitted
• Re-test against a sample of approved historic bills before any model or rule change goes live

Our piece on AI governance for law firms covers why the log is the control firms skip and regret.

A Third Route

Some firms don't want to staff a build and don't want their guideline rules sitting with a vendor. There's a middle option: a provider builds and runs the assistant, and the firm owns the rules and the finished system.

That's how Cognitive Outsourcing™ works when billing is the function being transferred. The AI assets hand back to the firm at the end of the term.

It only makes sense if you'd consider handing the whole billing function over. If you wouldn't, skip it and use the 5 questions above.

Common Pitfalls

• Comparing a licence price against a build estimate that leaves out maintenance.
• Letting the tool edit narratives for clients whose guidelines haven't been turned into rules yet.
• Buying before measuring current rejection and write-down rates, so there's no before-and-after.
• Accepting a vendor's claim of a "full audit trail" without looking at one.
• Dropping partner approval once the tool is right most of the time. Most of the time is where the risk hides.

How 3Rive Approaches This

We start with the billing data and the client rules, and choose the software after. Where a firm hasn't decided whether to own or buy, that's a Tech Advisory conversation first.

Where it's building, our AI service line delivers the assistant with the log and the review step built in from day 1. It's the same billing data work behind our guide to getting more from Elite 3E pricing.