An associate at a small estate practice uses an AI drafting tool to produce the first version of a revocable living trust. The tool takes four minutes. Without it, she estimates the draft would take forty-five. The firm is paying a $175 software subscription per month. She runs the math in her head: forty minutes saved, multiplied across dozens of matters, at her billing rate. The tool pays for itself many times over.
Then she reads the draft.
The document is fluent and structurally sound. It uses the right headings and flows in the right order. But one distribution clause uses a formulation she doesn't recognize, and she can't immediately tell whether it's a valid alternative construction or a plausible-sounding error. She looks it up. It's an error. She also notices that the trustee succession language doesn't match the jurisdiction-specific requirements her firm uses as standard, and the pour-over will cross-reference contains a section number that doesn't exist. None of these are catastrophic. All of them require correction. All of them require research she wasn't expecting to do.
The review takes thirty-five minutes. She saved ten.
What the Savings Calculation Actually Includes
The time-savings pitch for AI drafting tools is structurally incomplete in a way that is easy to miss. It compares the time to generate output against the time to produce that same output from scratch. This is a real comparison: generation is faster. But it omits the cost of the step that is non-negotiable for any work product leaving the firm: verification.
Verification is not optional. It was not optional before AI tools existed, and it becomes more demanding, not less, when the source of the draft is a model that generates confident text regardless of whether the underlying content is correct. The attorney who signs a document is responsible for what it says. The tool's confidence is not a defense. Neither is the subscription fee paid for it.
The model's fluency is precisely what makes verification harder. A poorly drafted document signals its own problems. A well-formed error does not.
The model's fluency is precisely what makes verification harder. A document with obvious grammatical errors, awkward construction, or clearly missing sections signals its own problems. A reviewer scanning it will slow down. A document that reads professionally, flows logically, and uses correct terminology throughout does not trigger the same instinct, even if one clause in twelve is wrong in a consequential way. The reviewer's vigilance is calibrated to the apparent quality of the work, and AI drafts consistently appear higher quality than they are.
This is what the book calls the verification tax: the time cost of reviewing AI output at the level of rigor the output actually requires, rather than the level of rigor its surface appearance suggests.
The Billing Problem Is Distinct from the Accuracy Problem
For most businesses, the verification tax is a cost-efficiency question. You spend time checking work that turned out to be mostly right, and you weigh that against the time you saved generating it. If the math favors the tool, you use it. If it doesn't, you don't.
For a billing practice, the calculation is different in kind, because verification time is not uniformly billable. The time an attorney spends reading and correcting an AI-generated draft sits in an ambiguous category. Some of it may be legitimately billable as attorney review. Some of it (particularly the time spent fixing errors the tool introduced) starts to look less like professional service and more like quality control for a software product the firm chose to use.
Bar guidance on this point is still developing, but the direction is clear. Billing a client for the full time required to fix AI-generated errors (errors the client did not cause and had no knowledge of) raises ethics questions that prudent attorneys are already thinking about. ABA Opinion 512 addresses billing directly: fees must be reasonable, and that standard applies to AI-assisted work as it does to any other. If the tool creates rework, someone is absorbing that cost. The ethical question is who.
In early 2024, Klarna announced that its AI assistant was handling the work of 700 customer service employees and reported significant per-case resolution improvements. The announcement was widely cited as evidence of transformative AI efficiency gains.
The part of the story that received less attention: those resolution metrics were self-reported, the quality assessment methodology was not independently verified, and the figures were released as part of a pre-IPO marketing campaign. By late 2024, Klarna had quietly begun rehiring human agents in categories where AI performance had proved inadequate.
The lesson is not that AI tools don't improve efficiency. Some do. The lesson is that vendor-reported efficiency figures are promotional material, not measurement. The only savings number that matters is the one your firm measures on your own matters, accounting for actual review time.
Running the Real Numbers
The math is not complicated, but most firms aren't doing it. They're accepting the vendor's before-and-after comparison (generation time versus from-scratch drafting time) without measuring the third column: post-generation review time on actual work product.
Consider what an honest ledger looks like for a single document type at a small estate practice. The numbers below are illustrative, but the structure of the problem is real.
Two minutes saved per document is not zero. Multiplied across hundreds of matters, it adds up to something. But it adds up to something very different than forty-one minutes, and it changes every downstream decision: whether the tool justifies its subscription cost, whether it reduces or merely relocates attorney workload, and whether the firm's efficiency claims to clients are accurate.
It also changes the staffing logic. The argument for AI drafting tools often includes the implication that firms can handle more matters with the same headcount. That argument depends on the gross savings figure. If the net figure is close to zero, the headcount argument collapses, and the firm that restructured around the assumption has a problem it didn't see coming.
The Error Rate Is Not Uniform
One reason the verification tax is underestimated is that error rates are not stable. An AI tool tested on fifty straightforward revocable trusts will perform better than the same tool applied to a complex special needs trust, a trust with unusual distribution mechanics, or a matter that involves jurisdiction-specific requirements the model hasn't encountered in proportion to their frequency. The firm that pilots a tool on its most routine work and then deploys it across all matters is not generalizing from representative data.
The error rate also shifts over time in both directions. Model updates can silently change output quality, sometimes improving it, occasionally degrading it in specific areas. A reviewing attorney who develops calibrated trust in a tool's output after six months of consistent performance has no mechanism for knowing whether the model she's trusting today is the same model she tested. Most tools do not version-notify their users when underlying models change.
Two attorneys submitted a brief in federal court that cited multiple cases generated by ChatGPT. The cases did not exist. When opposing counsel and the court could not locate the citations, the attorneys represented that the cases were real and produced fabricated copies to support that representation.
The court sanctioned both attorneys and their firm. The opinion was direct: the attorneys had a duty to verify the accuracy of their submissions, and that duty was not satisfied by trusting the output of a tool they had not adequately reviewed. The sanctions were not primarily about using AI. They were about certifying to the court that what they filed was accurate without having checked.
The verification tax, in that case, was not paid. The price of not paying it was public sanctions, reputational damage, and a documented professional failure that will follow both attorneys indefinitely.
What Competent Adoption Looks Like
None of this is an argument against using AI drafting tools. Some are genuinely useful, and for specific, bounded tasks (first-pass research summaries, initial clause suggestions, formatting standardized sections) the net time savings can be real and worth capturing. The argument is against adopting them without measuring what they actually cost, and against letting vendor benchmarks substitute for firm-level data.
ABA Opinion 512's competence requirement speaks to this directly. An attorney who uses AI in the representation must have sufficient understanding of the tool to evaluate its output. That understanding includes knowing what kinds of errors the tool is prone to, which document types are within its reliable range, and how much review time those documents realistically require. A firm that has never measured its own verification burden does not have that understanding, regardless of how long it has been using the tool.
Competent adoption involves three things that most small firms have not formalized.
Baseline measurement before deployment. Before a tool becomes part of the workflow, run it on twenty or thirty real matters of the type it will be used for. Measure generation time, review time, and correction time separately. The net figure from that sample is the only number that belongs in an ROI calculation.
Ongoing error logging. Track the types of errors the tool introduces, not just the frequency. A tool that is wrong about minor formatting 15% of the time is a different problem than a tool that introduces substantive legal errors 3% of the time. The second creates liability exposure the first does not. An error log also lets you detect if the tool's performance has changed after a model update.
Explicit review protocols by document type. Different document types require different review intensity, and that intensity should be written down rather than left to individual attorney judgment. A junior associate reviewing an AI-generated pour-over will for the first time does not have the same calibration as a senior partner who has reviewed two hundred of them. The protocol makes review requirements explicit and auditable.
This is not a large bureaucratic undertaking. A single spreadsheet tracking matters, tool used, generation time, review time, and errors found would capture most of what a small firm needs to know. The issue is that almost no one is doing it.
The Honest Conversation With Clients
There is a client-relations dimension to the verification tax that gets even less attention than the billing ethics question. Law firms that use AI drafting tools are, in most cases, not telling their clients they're doing so. Engagement letters don't mention it. Invoices don't reflect it. And clients who are paying attorney rates for document preparation have a reasonable expectation (whether or not it is legally enforceable) that the work product they're receiving was produced by the person billing for it.
ABA Opinion 512 doesn't require disclosure of AI use in every case. But it requires informed consent when the use implicates confidentiality, and it requires that fees remain reasonable. Both of those constraints bite differently depending on how the tool is being used and how the resulting time is being billed. A firm that is billing full rates for AI-assisted work while absorbing the verification time as a write-off is making an allocation decision that it hasn't made explicit anywhere. A firm that is billing for verification time without disclosing what it's verifying is in slightly different territory.
The firms that will handle this cleanly are the ones that think it through before the bill goes out, not after the client asks. That means having a considered position on disclosure, on billing for AI-assisted work, and on how verification time is categorized, and writing it down somewhere that a supervising attorney can point to.
The verification tax is not a reason to reject AI tools. It is a reason to measure them honestly, adopt them deliberately, and stop letting a vendor's efficiency claim do the work that your own data should be doing.