AI-Generated Billing Narratives Aren't Unethical. Lazy Ones Are.

The debate over AI billing narratives has been framed backwards. Every panel I attend treats the question as "is it ethical to let a model write your time entries?" That's the wrong question. Attorneys have been dictating narratives to paralegals for forty years, and nobody argued that was per se unethical. The real question is whether the narrative accurately describes work actually performed at the time claimed. AI doesn't change that standard. It just makes violations faster to commit and easier to detect.

Firms that treat AI narrative generation as a productivity hack without redesigning their review process are walking into a fee-dispute buzzsaw. Firms that treat it as an accuracy tool - one that forces better contemporaneous records - end up billing more defensibly than they did before. The technology is neutral. The workflow around it isn't.

The Real Ethics Problem Isn't What You Think

Model Rule 1.5 requires reasonable fees. Rule 7.1 prohibits misleading communications. Neither says a human must type each character of a billing entry. The ABA's Formal Opinion 512 (2024) explicitly permits generative AI use in legal work provided the lawyer maintains competent supervision and doesn't misrepresent the nature of the work.

So where's the ethical failure? It's in the drift between what the timer captured and what the narrative claims. A recent LexisNexis study found that 61% of attorneys admit their time entries are reconstructed at end of day or later, and 23% reconstruct entries more than 48 hours after the work. When a human reconstructs from memory, the narrative gets vague ("reviewed correspondence, analyzed issues"). When AI reconstructs from calendar entries, emails, and document activity, the narrative gets specific - sometimes more specific than the underlying work justifies.

That's the trap. An AI system that pulls "drafted motion to compel discovery re: defendant's insurance policy limits" from a document title creates an entry that reads like 2.3 hours of substantive drafting. If the actual work was 20 minutes of edits to an associate's draft, you've just billed a client for work that didn't happen the way the entry describes. The AI didn't lie. Your workflow did.

Why PI Firms Are More Exposed Than Big Law

Personal injury firms operating on contingency assume this doesn't apply to them. It does, and arguably more acutely. Fee petitions in fee-shifting cases (bad faith claims, EAJA matters, sanctions motions) get scrutinized by opposing counsel and judges with time on their hands. Common fund cases require lodestar cross-checks. And clients requesting case files at termination increasingly ask for time records.

A 2023 analysis of federal fee applications found that courts reduced requested fees by an average of 34% when narratives were deemed vague or block-billed. Firms using AI to generate uniformly detailed narratives across every entry are creating a new problem: entries that look consistent and specific but that opposing counsel can attack as templated. If your motion practice narratives all include the phrase "analyzed evidentiary standards under Fed. R. Evid. 702," a savvy opponent will ask for the underlying research notes. If those don't exist, you have a bigger problem than a fee reduction.

The firms getting this right treat AI narrative generation as a draft, not a deliverable. They require the billing attorney to verify each entry against contemporaneous notes, calendar blocks, and document metadata before submission. That takes about 45 seconds per entry in the systems we've deployed. It cuts write-offs by roughly 18% because entries survive client review.

The Disclosure Question Nobody Wants to Answer

Do you have to tell your client that AI drafted their bills? Formal Opinion 512 says generally no, unless the client asks or unless disclosure is material to the representation. I think that's technically correct and practically insufficient.

Sophisticated clients - insurers, corporate defendants funding coverage counsel, sophisticated plaintiffs in mass tort MDLs - are already asking. A 2024 Wolters Kluwer survey found that 39% of corporate legal departments have added AI disclosure requirements to their outside counsel guidelines. That number was 4% eighteen months earlier. If you're waiting for the client to ask, you're behind.

The better practice: include a one-sentence engagement letter provision stating that AI tools may assist in administrative tasks including time entry drafting, subject to attorney review. This solves the disclosure question, sets expectations, and gives you cover if a fee dispute arises. It costs nothing.

A Practical Framework You Can Deploy Monday

Before your firm bills another AI-drafted narrative, run every entry through four checks:

1. Source verification. Does the entry trace to a specific artifact - a document, calendar event, email thread, or call log? If the AI can't cite the source, the entry doesn't go out.

2. Time-to-narrative proportionality. Does the described work plausibly require the time claimed? A 0.2 entry shouldn't describe "comprehensive analysis." A 3.5 entry shouldn't say "reviewed file."

3. Specificity floor and ceiling. Entries should identify the matter task and legal issue without inventing detail the underlying work didn't produce. If your AI adds legal theories not reflected in the file, retrain it or replace it.

4. Attorney sign-off with attestation. The billing attorney clicks a confirmation that they reviewed each entry against their recollection of the work. This is your ethical firewall and your evidentiary record.

AI-generated billing isn't the ethics problem. Unreviewed AI-generated billing is. The firms that understand the difference will bill more, collect more, and defend more successfully. The firms that