The AI Billing Reckoning

This issue: You've been using AI to deliver faster work. Your clients have noticed. Now the IRS made it official — and your own professional association disagrees. This week: the billing dispute at the center of every CPA firm's Q4 pricing decision, why 82% of AI-adopting firms can't answer the question their clients are about to ask, and what happened in legal AI last week that most coverage conflated into one story.

---

Harvey raised $550 million to accelerate the displacement of the $80,000-a-year associate your clients are already asking you to hire fewer of. That was the tool side of the fee compression story.

On June 19, 2026, the regulatory side arrived.

The IRS Office of Professional Responsibility issued Alert 2026-19 — a formal interpretation of Circular 230's "unconscionable fee" prohibition. The ruling's practical implication: when AI reduces an 8-hour task to 2 hours, billing for 8 hours could constitute an unconscionable fee. Practitioners must "fairly credit cost reductions to the client's account."

The AICPA's Vice President of Tax Policy and Advocacy, Melanie Lauridsen, responded publicly. The AICPA is working with the IRS on "clarifying language and FAQs" because the guidance "doesn't appear to allow for a value-pricing model" and "overlooks the costs and risks associated with implementing the technology." The AICPA's position: AI-generated work has real value regardless of how quickly it was produced, and firms incur genuine costs — training, compliance infrastructure, liability exposure — that the OPR guidance ignores.

Two authoritative voices. Two incompatible standards. Neither definitive until the IRS issues its FAQs.

I've watched a version of this conversation arrive at my own agency — a client at annual renewal who had clearly done the math on what AI had done to our delivery costs. She wasn't wrong to run those numbers. Most firm owners haven't run them yet for themselves.

Until the clarifying FAQs arrive, the OPR Alert is the operative rule. The safest position right now is not to concede that AI efficiency automatically reduces what you can charge — it's to document the value delivered so that whatever you charge, you can defend it. That starts with one sentence in your engagement letter: Fees reflect the value of this work, not the time required to produce it. When a client asks — and they will — you have a documented answer, not an improvised one.

---

82% Flying Blind — And the Billing Dispute Makes It Worse

Thomson Reuters surveyed 1,500+ professional services professionals this year. Forty percent of organizations actively use AI — that number doubled in twelve months. Eighty-two percent of those organizations don't measure whether their AI is working.

That is not a secondary problem. That is the billing dispute made structural.

The OPR ruling requires CPA firms to document what AI efficiency actually produced — the value delivered, not just the hours saved. The AICPA's counterargument rests on the same requirement: value-based pricing is only defensible if the value is documented. Both positions assume you know what your AI is doing. Most firms don't.

The TR Future of Professionals 2026 report (1,800 global professionals) sharpens the picture: 74% of professionals use AI weekly, yet 91% are frustrated with value delivery, and only 6% consistently get the improvements they were promised. Not from lack of capability. From lack of measurement.

The business consequence is direct: 32% of corporate clients expect to re-evaluate their professional services relationships within twelve months based on AI delivery. A third of those say over $1 million in annual work is on the table per client. TR calculates the aggregate US exposure at $143 billion in client revenue in active reconsideration.

That number is a measurement problem as much as a capability problem. Firms that can demonstrate AI-enabled value will retain those clients. Firms that can't are carrying client retention risk they cannot even quantify — because they never established a baseline.

(The 24% talent attrition risk is the same equation: your best people move to the firms that measure and build AI competence.)

---

Two Legal AI Products Shipped the Same Day. One Is Ready for You. One Isn't Yet.

September 17 produced two distinct legal AI announcements. Most coverage treated them as one story.

Thomson Reuters and OpenAI announced a partnership. CoCounsel Legal — TR's legal AI built on Westlaw's citation database — now connects to ChatGPT Enterprise via MCP. If your firm is on ChatGPT Enterprise, you can call on CoCounsel's verified, source-grounded legal outputs from inside that interface without switching platforms. TR's framing: "Fiduciary Grade AI travels with the work." For a firm that has been reluctant to evaluate a second legal tool because the team just settled into ChatGPT, this removes the either/or.

OpenAI launched Astra for Law — its own legal AI product, separate from the TR partnership. Astra for Law has a purpose-built legal search index covering 230 million+ U.S. legal URLs (case law, statutes, regulations, court rules), updated daily, plus 26 partner plugins — including Clio. Accuracy benchmark on a 200-question legal research test: 54% correct at highest reasoning effort, versus 38.7% for GPT-6 without the legal index.

Fifty-four percent is a meaningful improvement. It is not a deployment-ready number for unverified client-facing work. OpenAI benchmarks this openly. The implication is the same as every current legal AI tool: verification on every output before it touches a client matter. What Astra for Law changes is access — a small firm on ChatGPT Enterprise will eventually access a legal-indexed model within their existing subscription, once access expands beyond the current Am Law 200 rollout.

For Clio firms: the Clio plugin in Astra for Law is the specific mechanism that will connect OpenAI's legal AI to your matter context when broader access opens. It is not available to small firms today. When it is, the evaluation question is not whether to try it — it's whether your firm's verification protocol is already in place before the first output goes to a client.

The market is forming into two categories: purpose-built legal AI with source grounding versus general-purpose AI with legal configuration. The evaluation skill every firm needs is knowing which category belongs in which workflow — and that decision does not require waiting for Astra for Law.

---

Your clients' in-house teams are now running AI against your invoices — and the question isn't whether you can defend what you charge, it's whether you have the documentation ready before they send the flag.

---

The Measurement That Makes Billing Defensible

One accounting firm cut its month-end close from 7.5 days to under two — and grew its client count by 55% in the same period without adding headcount. The MIT/Stanford study that captured this found the enabling factor was not the AI tool itself. It was the before/after measurement that made the AI workflow repeatable and the results attributable.

You cannot document value delivered if you never measured the baseline.

The 30-day measurement protocol:

Start with one workflow — not your whole practice, one specific deliverable type. Establish the before-number right now: how many hours does this currently take, or how many units does it currently produce? Set a 30-day calendar reminder. Run the same measurement on day 30 with AI active. The delta is your documented efficiency gain — the number that belongs in your engagement letter, your renewal conversations, and your OPR compliance file.

Three workflows worth measuring first, by firm type:

• Law: Document review and research on your most common matter type. Measure hours from intake to draft — Clio, MyCase, or Practice Panther for time data.
• Accounting: Month-end close cycle time or hours per standard tax return tier — Karbon, Canopy, or CCH Axcess for engagement records.
• Consulting: First-draft deliverable time for your most common report type — Harvest or Toggl for time capture.
• Staffing: Days from job order to submitted shortlist for a standard role — Bullhorn, Loxo, or JobAdder time-stamp data.
• Agency: Hours per deliverable on your highest-volume retainer service — Teamwork, Asana, or your project management platform.

The Fiduciary Grade AI question for every tool in your stack:

Ask this of every AI tool your firm uses for client-facing work: (1) Can I trace every output to a primary source? (2) Does this tool flag uncertainty rather than paper over it? (3) If a client asked to see my work, would this tool's output hold up?

All three yes: the tool belongs in client-facing workflows. One no: it belongs in internal workflows with human review before anything reaches a client. The 54% accuracy benchmark on Astra for Law is the calibration example — a published number that makes the question concrete rather than theoretical.

Stop using tools you haven't categorized. Every AI tool at your firm is either appropriate for client-facing work under the current professional liability framework, or it isn't. If you don't know which — that's your Q4 audit.

You don't need a perfect AI stack — you need to know which tools are appropriate for which category of work, documented before a regulator or client asks.

---

Regulatory Watch

All three updates below reflect the same pressure the billing dispute makes explicit: the documentation you haven't built yet is becoming the documentation someone else will demand.

California SB 574 — still pending, Newsom deadline October 12. Governor Newsom has not acted. Three weeks from this issue. If signed, SB 574 requires California attorneys to use only enterprise-grade AI tools with client matter data, personally verify all citations in AI-generated filings, and prevent discriminatory AI outcomes. For California law firms still using consumer-tier ChatGPT or Claude.ai with client data: that is the specific compliance gap SB 574 closes. The Q4 audit starts before Newsom signs, not after. Still pending — we will publish within 24 hours of Newsom's decision.

California SB 813 — signed September 9. The first U.S. state AI verification and disclosure framework. SB 813 creates disclosure requirements for AI-generated content in specified commercial contexts. For consulting firms and marketing agencies serving California businesses: a new advisory service line exists here — helping clients document compliant AI content processes before the 2027 compliance deadline.

Colorado SB 26-189 — effective January 1, 2027. Disclosure and audit requirements for AI use in hiring decisions. Applies to firms using AI-assisted resume screening or candidate evaluation — including tools your HR vendors supply. The compliance obligation belongs to your firm, not the vendor. Q4 is the window to audit your hiring tool stack and document your human oversight process before December 31.

---

One Action This Week

This week: establish your before-metric for one AI-assisted workflow.

Choose the workflow your firm runs most often for a standard client deliverable. Document the current baseline — hours or units before AI — in writing, right now.

• Law: Open Clio, MyCase, or Practice Panther. Pull time entries for your three most recent matters of the same type. Calculate the average hours to draft or research. Write it down.
• Accounting: Open Karbon, Canopy, or your engagement time tracker. Pull the average hours for a standard return or close cycle from last quarter. Write it down.
• Consulting: Open Harvest, Toggl, or your last three project actuals. Calculate average hours on your most common deliverable type. Write it down.
• Staffing: Open Bullhorn, Loxo, or JobAdder. Calculate your average days from job order to submitted shortlist, last quarter. Write it down.
• Agency: Open Teamwork, Asana, or your project tracker. Pull average hours per deliverable for your highest-volume retainer service. Write it down.

Your success criterion: By Wednesday morning, you have one line in writing: "[Workflow]: currently takes [X hours / Y days] per [matter / return / project / placement / deliverable]."

Not "AI is saving us time" — a number that existed before AI. That baseline is the evidence you'll use to document value when a client asks, defend your fees if the OPR asks, and prove to yourself that the AI workflow you're running is actually working.

The 82% of firms that don't measure AI value? Most of them don't have that one line yet. Writing it this week puts you in the 18% before Q4 starts.

---

Forward this to a firm owner who's been using AI for six months but hasn't updated their engagement letter — the OPR ruling is a countdown, not a warning.

Wolters Kluwer just deployed an AI that automatically audits law firm invoices against client guidelines — and it's already live at the in-house legal teams you send yours to.