Can a firm bill a client for time that AI saved?
Not as though a human had worked it. ABA Formal Opinion 512, issued 29 July 2024, reiterates that fees must be reasonable and addresses this directly: a lawyer cannot bill a client for time AI saved as if the work had been performed by a person.
Conduct rules are adopted state by state and diverge, so what a particular jurisdiction requires is a question for that jurisdiction. But the framework in Opinion 512 has been broadly influential, and firms planning around a contrary assumption are planning around an unusual position.
The practical effect is to close off the path of least resistance. A firm cannot simply keep billing the old hours and absorb the efficiency as margin.
Are firms actually changing how they price?
Yes, and faster than the public conversation suggests. Among firms Clio’s 2025 Legal Trends Report classified as wide AI adopters, 45% had already adjusted pricing. Alternative fee arrangements are no longer a minority practice: 72% of US firms offer some form of AFA, rising to roughly 90% at firms of 50 or more lawyers.
The readiness picture is less flattering. BigHand’s 2025 legal pricing and budgeting report found that 100% of surveyed firms said AI and technology were reshaping pricing while only about a third felt ready for it. Firms that use matter budgets are doing measurably better: 70% of those firms reported realisation gains of 9% or more.
Where does the efficiency gain go if the firm cannot bill it?
To one of three places, and choosing deliberately is the whole exercise: to the client as a lower bill, to the firm as margin under a fixed or capped fee, or to capacity as more matters handled by the same lawyers. Only the second and third preserve firm revenue, and both require moving off pure hourly billing to capture.
That is the uncomfortable structural point. Under pure hourly billing, an efficiency gain is a revenue loss by construction. Every other pricing model converts it into something the firm keeps.
Do AI time-capture tools help or make it worse?
They recover time that was genuinely worked and never recorded, which is a real problem, and they sit directly on top of the Opinion 512 constraint. Vendors report 5–20% more billable time captured, and those figures are vendor-reported and should be treated as such. At a $400 blended rate against 1,600 annual hours, even a 5% recapture is roughly $32,000 per timekeeper per year, which is arithmetic any firm can run against its own numbers.
Two cautions travel with it. The monitoring is surveillance-adjacent and raises employee-privacy and morale questions that are separate from the ethics of the billing. And the captured data includes privileged content, which makes the vendor’s security posture a confidentiality question rather than a procurement detail.
The ethics floor does not move. Time that AI found still has to be time honestly billed.
Are clients using AI on the bills as well?
Yes, and firms should assume the invoice is being read by a machine before a human sees it. Corporate clients run AI bill review against firms through established vendors in that market, checking narratives against outside counsel guidelines for block billing, vague descriptions, staffing violations and non-billable task codes.
That mirror market reframes what prebill AI is for. A firm adopting AI prebill scrubbing is substantially playing defence, matching a capability the client already has, rather than opening a new source of revenue.
What is the actual strategic question for firm leadership?
Not whether to adopt AI but what the firm sells, and whether that is still hours. A firm that sells hours has an efficiency problem it cannot solve within its own pricing model. A firm that sells outcomes, or access, or a defined scope, converts the same efficiency into margin.
That decision sits above the billing system and is difficult to reverse once clients have been trained on a rate structure. The Institute’s Billing, Pricing & Fees area works through the models, and New Revenue & Service Lines covers what firms are selling instead.