Two versions of the same one-matter invoice excerpt, both modeled hypotheticals:
Version A Partner · revise and finalize motion · 1.5 hrs · $1,800
Version B Partner · revise and finalize motion · 1.5 hrs · $1,800 AI document analysis · $180
Nothing about Version B is unethical. With disclosure in advance, the ethics opinions permit exactly this, and the $180 is a real cost that somebody really incurred. But read it the way the client reads it. The second line does not describe a cost. It asks a question about the first line, and the question is why the first line is still there.
The prebill is an identity document, not a cost report.
1 · The identity document
Everything else a firm publishes about itself is elective reading. The invoice is different in two ways no other document shares: it arrives on a monthly cadence, and it carries a payment consequence. The client cannot file it unread; someone must review it, approve it, and act on it, every month, for the life of the relationship. Which means the invoice is the only recurring document in which the firm states, under its own letterhead, to a reader obligated to respond, what it believes it is selling.
The ethics floor beneath that document is settled and old, and we cite it here as authority, never as discovery. ABA Formal Opinion 93-379 (1993): disbursements pass through at actual cost, in-house services at no more than direct cost plus a reasonable allocation of associated overhead, and absent agreement a lawyer may not "create an additional source of profit for the law firm beyond that which is contained in the provision of professional services themselves." The canonical sentence: "The lawyer's stock in trade is the sale of legal services, not photocopy paper, tuna fish sandwiches, computer time or messenger services." ABA Formal Opinion 512 (July 2024) walked the same principles over to generative AI: bill actual time, not the time the tool saved you; treat infrastructure-grade tools as overhead "absent a contrary disclosure to the client in advance"; pass per-use third-party charges through at actual cost.
On that floor, the trade press has spent two years arguing one question: pass the AI cost through, absorb it as overhead, or break out a technology fee. The cost-recovery frame is the consensus; the answers split. Thomson Reuters' pricing lead says absorb it and never line-item it; a Bloomberg Law columnist, engaged by name in Section 6, says be transparent about it. Both camps treat the bill as a cost-allocation problem. Neither asks what the document is. If the bill is testimony, the question is not where the cost goes. It is what testimony a machine line gives.
2 · The substitution line
Here is the mechanism, and as far as we can find it is unclaimed in print. A bill that lists partner time beside AI time does not disclose an input. It itemizes the machine as a substitute input. The two lines sit in the same column, in the same currency, describing work on the same matter, and the client's procurement function reads parallel entries the only way they can be read: as alternatives. The machine line converts augmentation into competition inside the lawyer's own ledger. It hands the client a unit price for the machine and a standing invitation to ask what fraction of the human line above it could be moved down. The rational client does not thank you for the candor. The rational client prices the human residue.
Call any machine-time entry on a professional invoice a substitution line: a line that itemizes the machine as a substitute for the professional selling the bill, and invites the client to price what remains of the human.
The closest prior art stops one step short of this. William Josten of the Thomson Reuters Institute argued in May 2024 that AI billing "is not a question of cost recovery": "The costs of advanced AI technology should not be lumped into this category," the pricing professionals he canvassed agreed that "generally AI should not be treated as a disbursement," and "GenAI is simply the next evolution of necessary technology" · infrastructure, recovered through rates and value, not line items. We agree with every word, which is precisely why none of it is this memo's claim. "Don't bill the robot as a disbursement" is the consensus floor: Josten said it in 2024, Op. 512 codified its ethics the same summer, and the 2025 survey median (Section 5) shows the market already practicing it. What Josten never asks is why the slot destabilizes. The substitution line is the answer: the machine line fails not because it is gauche but because it testifies against the bill it sits on.
And the profession has already run this experiment once, with receipts.
3 · The Westlaw film, complete print
When Lexis and Westlaw arrived in the mid-1970s, billing clients for computer-assisted research was, per the Wisconsin State Bar's InsideTrack (Laura Olsen, October 2014), "commonplace," surcharges included. The ABA's Law Technology Today retrospective runs the rest of the arc: Lexis and Westlaw fees became "common line items," clients accepted them for a while, then "began demanding change, labeling certain costs as overhead," and by the early 2000s the retreat was general. Today few firms openly bill for Lexis or Westlaw at all.
The dollar print of that film exists, courtesy of Mattern & Associates' cost-recovery work: per the ALM Law Librarian Survey Mattern cites, as recently as 2009 more than half of Am Law 200 firms recovered better than 60 percent of their online research costs. By 2014, Mattern's own cost-recovery survey reported net realization on legal research at 36 percent, the lowest in the survey's history to that date. By 2016 it had sunk to 25 percent. Three decades, one direction, the slope steepening at the end: the machine line commoditized to functional zero on a measurable curve.
Now the controlled comparison, from the same 2016 survey: over 90 percent of billable hard costs were still being invoiced, at 86 percent realization. That contrast is the finding. Clients pay for couriers forever and for machines never. A courier is a true third-party pass-through; paying it confirms nothing about the firm, because the courier is not the firm. The research machine sits inside the work the lawyer signs, so its line reads as a statement about how the lawyer's own product is made · a substitution line · and clients do not pay substitution lines. They reprice them toward zero, then start on the human hours beside them.
Medicine looks like the counter-example: a whole profession that minted a machine billing slot, nationally, in public. It is the strongest supporting evidence this memo has.
4 · The slot medicine minted
CPT 92229 is the first CPT code for an autonomous-AI diagnostic: software that reads retinal images for diabetic retinopathy and issues the finding itself, no overreading physician. Medicare had paid for AI before · the add-on payment for Viz.ai's stroke-triage software in hospitals, beginning in 2020 · but 92229 is the first time autonomous AI got its own slot in the professional fee schedule. And the slot's national price has fallen every year it has existed. By the Medicare Physician Fee Schedule national payment amount, as reported in a 2025 Ophthalmology Science study of health-system adoption: $47.06 in 2022, the first year a national rate existed; $45.74 in 2023; $40.28 in 2024. Down every year of its nationally priced life, roughly 14 percent in two years, under conditions no machine line will see again: a statutory payer, a national price, a public-health mandate, no human competitor on unit cost.
The distinction that converts this from counter-example to evidence: a fee-schedule code exists because a third-party payer cannot reimburse what has no slot. CMS needs the line; the line is how payment happens at all. So in medicine the machine line commoditizes and survives, held in place by the architecture of reimbursement while its price grinds down. Relational professional billing has no payer who needs the slot · only a client reading it, and the client reads it as Section 2 says. So in law the machine line commoditizes and dies. Same disease, no life support. The general law, now with a medical control group: the machine line commoditizes wherever it appears.
Which brings us to the firms minting slots right now.
5 · Act one, again
LawFuel reported in May 2026 that "Several Am Law firms have introduced 'AI technology fees' or built AI costs into engagement letters as a separate line item." LawFuel names no firms and cites no survey, so weight the sentence accordingly; the directional corroboration is independent. An Axiom-commissioned survey of 607 senior in-house legal leaders, fielded April to May 2025 and released that July, found 33 percent of law firms charging more for AI-assisted work and 58 percent not reducing rates at all. Meanwhile the market median sits exactly where the Westlaw film predicts: in Thomson Reuters' 2025 Generative AI in Professional Services report, the plurality law-firm answer on GenAI costs, at 42 percent, is absorb as firm overhead, an expectation the report attributes to the precedent of past technologies. The same report: 71 percent of law-firm clients have no idea whether their outside firms use GenAI at all.
The prediction, stated carefully because the careless version is already falsified, is not that no AI line item appears. They are appearing now. The prediction is that no AI line item stabilizes: the 2025-26 AI technology fee is the first act of the Westlaw film, transitional and doomed · acceptance, then realization sag, then clients relabeling the line as overhead, then quiet retreat into the rate. The line's own buyers will kill it, and the mechanism is already visible in the only vocabulary buyers maintain.
That vocabulary is the e-billing standard. LEDES and the UTBMS code sets are the buyer side's own taxonomy of everything it is willing to be billed for: task codes for litigation phases, expense codes for photocopies and couriers, code sets for patent prosecution as recently as 2023. As of June 2026 there is no published or proposed UTBMS code for generative AI · not an expense code, not a task code, not a draft. The buyers who write the codes do not want to buy machine time. They want to price humans against it. Hold that fact; it answers the strongest objection this memo faces, in Section 8.
One column in the trade press argues the opposite, by name, and deserves a direct answer.
6 · The foil · the story the transparent bill tells
Eric Dodson Greenberg's Bloomberg Law column, "AI Billing Transparency Tells a Story That's Good for Law Firms" (May 2026), is the best version of the other side, and it concedes this memo's premise in its first move: the bill tells the client who the firm is. "Opacity cedes the narrative to the client's imagination, which will default to the simplest story." Client AI questions, he writes, should be understood "as an open door for a discussion about value" rather than "as a trap door for discount discussions"; firms without usage data "will still have website pages with luminous dots," and little else. Concede his operational core in full: track everything at task level, answer RFPs with usage data, never stonewall a client who asks. Candor and competence questions; this memo touches neither.
What Greenberg never asks is what story the AI-transparent bill actually tells. Section 2 answered: a substitution story, told monthly, under your own letterhead, to the one reader obligated to act on it. The adjacent profession has demonstrated what a sophisticated client does with that story, including the detail Greenberg's thesis cannot survive, in an episode this series read in full in "Nobody Adopts Efficiency. Everybody Adopts a Raise.": the client did not need the invoice line. KPMG International, pressing its own auditor, did not read an AI disbursement on Grant Thornton's bill. It read Grant Thornton's publicized AI rollout, pre-computed the discount, urged the firm in negotiation to share the savings, and signaled it might switch auditors; the audit fee for KPMG's FY2025 accounts fell from $416,000 to $357,000, roughly 14 percent, as first reported by the Financial Times. The client who can see the machine prices the machine; the invoice line just hands them the arithmetic. The repricing happens even without the slot · the strongest form of this memo's claim, and the quiet refutation of Greenberg's. If visibility alone moves the fee 14 percent in one negotiation, a standing machine line on the monthly bill is not a value story. It is a price list for your own replacement, published by you, monthly.
So the dispute with Greenberg is the surface, not the data. Measure everything. Disclose what candor requires. And put the cost where it legally lives, which is never a line on the bill.
7 · The three ledgers
A real cost has to live somewhere. The architecture is three books.
The first ledger is the MSO's books: cost truth. "A Legal Transaction Has a Bill of Materials. Here Is One." published a complete node-level decomposition of a modeled matter · 84 tasks, every node costed, the roughly 40 percent human floor mapped against the roughly 60 percent the machine can reach. Read carelessly, that artifact is exactly the itemization this memo forbids: machine input, named and priced, node by node. So state the reconciliation plainly, because the two memos are one architecture read from opposite sides. The BOM is the internal cost ledger that exists so that the external invoice never has to show it. A firm that cannot decompose its own production cannot price, audit, or defend anything; a firm that prints the decomposition on the client's bill testifies against itself monthly. Cost truth inside. Identity outside.
The second ledger is the firm's P&L: cost routing. The AI spend is the MSO's input cost, and it reaches the firm inside the MSO's fixed management fee · plumbing settled in "The AI Dividend Has an Address, and It Is Written in the Fee Clause" and not reopened here. From the firm's side the fee is overhead, exactly as the Westlaw subscription is today: real, paid, recovered through rates, itemized to no one.
The third ledger is the client's invoice: identity. The bill stays 100 percent human time, recovered through the densified rate · the dense hour whose economics, measurement, and audit defense are the core argument of "Nobody Adopts Efficiency. Everybody Adopts a Raise.", cited here as canon, not re-derived. The rate rises because the hour carries more, and the record proving it lives where that memo put it.
The Three Ledgers
| The MSO's books | The firm's P&L | The client's invoice | |
|---|---|---|---|
| What it records | Node-level BOM cost truth | The fixed management fee | Hours and rates |
| Where AI appears | Named, costed, decomposed at node level | Inside the fee, undifferentiated | Nowhere |
| Who sees it | Operators and diligence | The partnership | The client |
| Governing authority | The BOM memo's open methodology | The fee clause ("The AI Dividend Has an Address, and It Is Written in the Fee Clause") | Op. 512 and 93-379 overhead treatment; the dense-hour rate record |
| What stabilizes it | It is internal | Contract | The two things clients have always paid for: human time and true third-party pass-throughs |
Beneath the table, one flow, every figure modeled: a $40 machine cost on a single task becomes a BOM node in the MSO's ledger, becomes a component of the fixed fee, is recovered inside the densified rate, and the invoice shows 1.5 partner hours and no machine. Cost truth in the first column exists so the third column never has to show it.
8 · The practice pushes back
Now the objection, from a 30-year COO, at full strength: "This is a shell game with a diagram. My clients' outside counsel guidelines already demand disclosure of AI use and pre-approval of staffing; Op. 512 contemplates disclosure, and 93-379 exists precisely to stop firms profiting on costs the client cannot see. Routing the machine through an MSO fee and recovering it in a raised 'human' rate is the surcharge 93-379 banned, with extra steps. Greenberg is right about procurement: clients are asking in RFPs, and opacity dies whether or not your ledgers are elegant. And your own mechanism boomerangs. If the substitution signal is real, then my clients rationally want to price the human residue · so OCG itemization demands should stabilize, because the slot benefits the buyer. You are predicting the death of a line my clients have every incentive to force onto the bill."
Take it in order, and concede what must be conceded. The candor duties are real and untouched. This memo designs the invoice's line architecture, not the engagement letter's disclosures; a firm may, and often must, disclose that it uses AI, and should answer every RFP with usage data, which is exactly why the first ledger exists. Concede also the hard case: a client whose OCG demands AI cost itemization must be itemized for or declined. No architecture overrides a contract the firm chose to sign.
The 93-379 charge fails on its own terms. What 93-379 forbids is a hidden surcharge on a pass-through and profit smuggled through the expense column. The three-ledger design has no pass-through to surcharge: the AI spend is the MSO's input cost, the firm pays a fixed fee that sits in overhead exactly as the research subscription does today, and overhead recovered through rates is not an evasion of 93-379 · it is the Thomson Reuters survey median and Op. 512's own default treatment of infrastructure-grade tools. The opinion polices the expense column, not what overhead may contain, or every Westlaw contract in America would be a violation.
The boomerang is the strongest sentence in the objection, and it deserves the memo's best fact rather than a wave. If buyers rationally wanted the slot, the place that desire becomes durable is their own standard: the UTBMS code sets, amended as recently as 2023 for patent prosecution. The buyers who write the codes have minted none for generative AI · no expense code, no task code, no proposal · because that is not how the buy side captures machine value. It captures the value through rate negotiation and benchmarks, not slot creation, and KPMG is the live demonstration: no line item demanded, no code requested, a publicized rollout read from outside, a discount pre-computed, 14 percent taken in negotiation. A standing slot costs the buyer audit effort to police a number the seller controls; the rate conversation gets the same money without it. Individual OCG demands will come, and the firm that signs one must honor it. But a buyer side that wanted the slot to stabilize would have written the code by now. It wrote codes for photocopies.
Concede last what the third ledger owes: the densified rate must survive audit, monthly, against clients who measure. The record that defends it is the ground of "Nobody Adopts Efficiency. Everybody Adopts a Raise.", and this memo adds nothing to it except the reason the rate, and not a machine line, is where the defense belongs.
9 · The close
The history runs one way. The research line died on a published curve while couriers sailed through at 86 percent realization; medicine's machine slot, the best-protected ever minted, lost ground every year of its nationally priced life; the e-billing standard has no slot for the machine and none proposed; and the one client who priced a provider's AI did it without any line at all. The robot's cost is real, and it belongs in the books built to hold it.
The firm that puts the robot on the invoice is answering the client's question. The firm that builds the three ledgers is deciding who asks it.
This memo is published by Jopese, a legal management services organization operated by HIRO PARTNERS LLC, a Texas limited liability company. It is offered for educational and analytical purposes only. It is not legal, tax, or investment advice, and it is not an offer to sell or a solicitation of an offer to buy any security or service. Jopese is not a law firm and does not provide legal advice or legal services; legal services are delivered by an independent law firm under a separate engagement in which Jopese does not participate. The invoice excerpts, the Three Ledgers table, the $40 flow illustration, and every dollar figure and percentage not attributed to a published source are hypothetical modeled illustrations and do not describe any actual firm, lawyer, client, agreement, or transaction, including any to which Jopese is a party. References to specific companies, publications, surveys, opinions, and market developments are drawn from public sources and are provided as market commentary, not as an endorsement, a recommendation, or a representation of any relationship. "The substitution line" and "the three ledgers" are the authors' analytical labels, not industry terms of art. Questions: rene@jopese.com.