Generative AI Billing Policies for Law Firms: Ethics, Models, and Engagement Letter Language

A generative AI billing policy for a law firm no longer starts with whether the firm permits lawyers to use AI. It starts with the engagement letter: what tool may touch the work, what work it may touch, what client information may be entered, who verifies the output, and how the client will be charged when the tool changes the time required.

That sounds narrower than most AI governance conversations, but it is where the disputes will surface. A general clause saying the firm may use “technology” or “automated tools” does not tell the client whether a generative AI system will draft a research memo, summarize discovery, translate a contract, or receive confidential facts. It also does not tell the billing department whether a saved three hours became a lower invoice, a flat-fee margin, an overhead cost, or an unapproved client expense.

As of Q3 2026, the safest consensus is simple on hourly billing and less simple everywhere else. ABA Formal Opinion 512, issued in July 2024, is now joined by formal or published guidance from at least seven state bars: Texas 705, Florida 24-1, Oregon 2025-205, Virginia LEO 1901, DC 388, North Carolina 2024-1, and Washington 202505.[1][2] The opinions do not all answer the same questions in the same way. For billing policy purposes, that difference matters more than any broad claim that AI will end the billable hour.

The Hourly Rule Is the Least Ambiguous Part

For hourly work, the operative rule is not difficult to state: bill actual time. If a lawyer would once have spent four hours on a research task and, using generative AI, completes the task in forty minutes, the lawyer bills forty minutes. The client is not billed for the hours the technology eliminated.

That conclusion is not a new AI exception. ABA Formal Opinion 512 relies on the older billing principle that a lawyer may not charge a client for time not actually expended, and current commentary describes the same baseline across jurisdictions addressing AI-assisted work.[3] The tool changes the workflow; it does not create a reserve of billable time that the firm may recapture because the task used to take longer.

The operational consequence is plain enough to write into a billing policy without embellishment: timekeepers record the time they actually spend using, reviewing, correcting, and finalizing AI-assisted work. They do not enter the historical or estimated time the task would have taken without AI. If the invoice narrative mentions AI, it should not imply that saved time is chargeable as lawyer time.

The more difficult internal question is compensation, not ethics. A practice group may worry that actual-time billing penalizes efficient lawyers. That is a management issue; it is not a reason to put four hours on an invoice for forty minutes of work. If the firm wants to price for value rather than time, it should move the matter into a fee structure that has been disclosed, agreed to, and reviewed under the governing jurisdiction’s reasonableness standard.

Flat Fees Are Where the Authorities Stop Agreeing

Flat fees create the real interpretive problem. Virginia LEO 1901, adopted by the Supreme Court of Virginia, says it is “not per se unreasonable” to charge the same non-hourly fee for AI-assisted work as the lawyer would have charged without AI. ABA Formal Opinion 512 takes a more cautious view, warning that a fee may be unreasonable if generative AI drastically reduces the lawyer’s effort.[4]

Those positions are not identical, and a law firm should not paper over them with one national template. Virginia’s formulation gives more room for the familiar flat-fee premise: the client buys the result or defined service, not a particular number of lawyer hours. The ABA formulation puts more pressure on the proportionality between price and effort when the tool materially changes the economics of the work.

The practical difference is easy to miss until the invoice is challenged. Suppose, hypothetically, a firm offers a fixed price for a standard contract review. If generative AI reduces the first-pass issue spotting and clause comparison work substantially, Virginia’s “not per se unreasonable” language does not automatically condemn the original fixed price. Under the ABA framing, the same facts may require a closer reasonableness analysis if the lawyer’s effort was drastically reduced. The facts still matter: scope, expertise, risk, responsibility, review, negotiation, and outcome do not disappear merely because a first draft moved faster.

A defensible flat-fee policy therefore needs more than a sentence saying “AI may be used.” It should identify which matters or tasks are eligible for flat fees, require jurisdiction-specific review where the client or matter is governed by stricter authority, and preserve records showing what professional work remained after AI assistance. The point is not to convert every flat fee back into an hourly bill. The point is to be able to explain why the agreed fee remained reasonable after the firm used a tool that changed the labor profile.

Subscription Costs Usually Belong in Overhead

The temptation to treat AI subscriptions as a new reimbursable technology expense is understandable. It is also where firms can create a billing problem without any hallucinated case citation. Oregon Opinion 2025-205 states that lawyers who cannot determine the actual per-client cost of flat-rate AI subscriptions may not prorate those charges to clients and must treat them as overhead, applying the same logic associated with Westlaw billing in Waggoner v. Chadbourne.[5]

That does not mean every technology cost is unrecoverable in every circumstance. It does mean the firm needs a cost-allocation answer before the charge appears on a bill. A monthly enterprise license, used across matters, practice groups, training, internal administration, and client work, is not automatically a client disbursement because a lawyer happened to use it during a matter.

There is also a commercial signal here. In a 2025 Harvard Center on the Legal Profession study, none of the ten AmLaw 100 firms interviewed planned to pursue direct AI cost recovery from clients; the prevailing top-firm strategy was to absorb AI costs as an investment.[6] That is not an ethics rule, and the sample should not be overstated. It does show that treating AI infrastructure as overhead is not merely a defensive posture for smaller or more cautious firms.

What the Engagement Letter Has to Say

The engagement letter is where the billing policy becomes client-facing. ABA 512 is commonly read to require more than boilerplate technology language for informed consent under Rule 1.6 when confidential information may be disclosed to or processed by a generative AI tool. Current engagement-letter commentary frames the needed disclosure in concrete terms: name the AI tools, identify the categories of work for which they may be used, and explain the data-handling architecture.[7]

A clause that says the firm may use “artificial intelligence and other technologies to improve efficiency” leaves too much undecided. It does not tell the client whether the firm will use a public chatbot, a legal research platform with AI features, an enterprise model configured not to train on client data, a document-review system, or an internal drafting assistant. Those distinctions matter because confidentiality, privilege, data retention, vendor access, and verification duties do not look the same across tools.

A useful engagement-letter provision should cover at least four categories, with enough specificity that a billing reviewer and the client can later understand what was authorized:

• Tool identity: the specific platform or category of platform the firm may use, with enough naming precision to distinguish a public tool from an enterprise or legal-specific system.
• Permitted work categories: research assistance, document summarization, drafting support, contract analysis, discovery review, translation, administrative work, or other defined uses.
• Data handling: whether client confidential information may be entered, whether prompts or outputs may be retained, whether the vendor may use data for training, and what security or contractual protections apply.
• Human review and billing treatment: who remains responsible for verifying outputs, how AI-assisted time will be recorded, and whether subscription costs are included in overhead or separately charged if allowed.

This is not the place to promise that AI output will be perfect, or to bury the issue in a policy attachment no one discussed. It is the place to secure consent for the actual workflow the matter team intends to use. If the firm later changes tools or expands AI use into a more sensitive category of work, the consent analysis should be revisited rather than treated as covered by a generic opening clause.

The engagement letter also should line up with the firm’s internal AI governance policy. A broader policy can address approval workflows, vendor diligence, training, supervision, incident response, and retention. The engagement letter should not try to carry all of that architecture. For that wider structure, see Building a Law Firm AI Policy: 8 Essential Components Based on Bar Guidance and Policy Frameworks. The billing document’s narrower job is to connect consent, fees, confidentiality, and verification in a form that can survive a later billing or ethics review.

Sample Consent Topics, Not Plug-and-Play Language

Firms often ask for model language. The safer answer is model topics. Local authority, client sophistication, matter sensitivity, and the tool’s architecture can change the wording. Still, the provision should be concrete enough to answer these questions before work begins:

• Will the firm use generative AI on this matter, or only with later client approval?
• Which named tools or approved tool classes may be used?
• Which tasks are in scope, and which tasks are excluded unless separately approved?
• May confidential client information, privileged material, personal information, or trade secrets be entered into the tool?
• Does the vendor retain prompts or outputs, train on client data, permit human review, or store data outside an agreed environment?
• How will the firm bill time saved by AI, AI review time, and any tool-specific charges?

The billing question should not be left to invoice time. If the matter is hourly, the letter or billing policy can say that lawyers will bill only time actually spent, including time spent prompting, reviewing, verifying, revising, and applying professional judgment. If the matter is flat-fee or value-priced, the letter should make clear that the fee is not a disguised hourly estimate and that AI use has been considered in the pricing and reasonableness analysis.

Documentation Is the Control That Makes the Policy Auditable

An AI billing policy that cannot be audited is mostly a statement of preference. The records do not need to preserve every prompt in every matter, especially where confidentiality or retention rules counsel against it. But the firm should be able to show that it approved the tool, obtained required client consent, recorded actual time, treated subscription charges consistently, and verified the work product before filing, sending, or billing it.

For hourly matters, the most useful controls are time-entry rules and billing-review flags. Entries should reflect the work performed, not advertise novelty. “AI research” is usually less helpful than a narrative explaining that the lawyer researched a stated issue, reviewed authorities, and revised the analysis. If the time entry is only forty minutes because AI accelerated the first pass, the invoice should not be padded by administrative convention before it leaves the firm.

For flat-fee matters, the better record is not a shadow timesheet kept for every task. It is a pricing file or matter note showing the basis for the fee: scope assumptions, expected deliverables, risk allocation, staffing, level of attorney responsibility, anticipated AI use, and jurisdictional reasonableness review. If a regulator or client later asks why the price remained fair after AI reduced the first-draft time, the firm should not be reconstructing that answer from memory.

For subscription costs, the record should show whether the firm treated the cost as overhead or, if it charged the client, how it determined the actual client-specific cost and why that charge was permitted under the applicable authority. A flat allocation across clients because the software invoice arrived once a month is the kind of convenience that looks poor under scrutiny.

Sanctions Make Verification a Billing Issue Too

The sanctions cases are often treated as a separate hallucination story. For billing governance, they are also evidence that verification is not an optional quality-control step. If a firm charges for AI-assisted work product that has not been checked, the problem is not only that the filing may be wrong. The invoice may also reflect professional work the lawyer did not competently perform.

The trajectory is no longer theoretical. In Mata v. Avianca in the Southern District of New York, the court imposed $5,000 in sanctions in 2023. In Lacey v. State Farm in the Central District of California, sanctions reached $31,100 in 2025. In Couvrette v. Wisnovsky in the District of Oregon, sanctions exceeded $110,000 in 2025 and involved 15 nonexistent cases and eight fabricated quotations.[5]

Those cases should not be used to suggest that every AI use is a citation disaster waiting to happen. They do support a narrower and more useful rule: the billing record should align with a verification record. When AI is used for research, the lawyer should verify the authorities. When it is used for drafting, the lawyer should review the legal and factual content. When it is used for document analysis, the firm should be able to explain the review protocol and the human responsibility that remained.

A Policy Choice, Not a Template Exercise

The current materials support a restrained but firm conclusion. A law firm should not rely on boilerplate technology clauses if it intends to use generative AI in client work. It should choose a billing model with jurisdiction-specific awareness, especially for flat fees. It should bill actual time on hourly matters, treat unknowable subscription allocations as overhead, and maintain records showing consent, tool approval, billing treatment, and verification.

There may be commercially attractive ways to explain AI efficiency as value, particularly in fixed-fee and hybrid arrangements. But the ethics file has to be strong enough before the client-relations conversation begins. If the engagement letter does not identify what the client consented to, and the billing record does not show what the firm actually did, the firm has left too much to be argued after the invoice is disputed.

This article is informational coverage, not legal or ethics advice. The Virginia and ABA flat-fee tension makes local authority decisive, and firms should verify the rules, opinions, court orders, and client requirements that govern their own matters.

References

1. ABA Formal Opinion 512: The Paradigm for Generative AI in Legal Practice, UNC Law Library
2. Beyond the Ban: Why Your Law Firm Needs a Realistic AI Policy in 2026, North Carolina Bar Association, January 13, 2026
3. Be Reasonable: AI and Legal Billing Practices, Attorney at Work
4. How Generative AI Is Disrupting Law Firm Billing Practices, LexisNexis
5. AI Legal Ethics, GC AI
6. The Impact of Artificial Intelligence on Law & Law Firms' Business Models, Harvard Center on the Legal Profession
7. Engagement Letter Language AI, Zusman Partners