On this page
One signature, more drafts
In October 2025 Deloitte agreed to refund part of the A$440,000 (about US$290,000) it was paid for a report to an Australian government department. The report cited papers that do not exist and put words in a federal judge's mouth. A revised version disclosed that Azure OpenAI had been used[62]. A large firm's review process let it through.
In June 2026 the IRS Office of Professional Responsibility cited that report in its first guidelines on AI in tax practice. They say that "practitioners cannot rely solely on AI; human scrutiny and editing are essential", and that a firm should "fairly credit to the client's account any cost reductions" AI brings[13].
Staffing makes it harder. Accounting degrees fell 6.6% to 55,152 in 2023 to 2024[32]. Starting pay for master's hires rose 17% in two years[31]. Tax professionals e-filed 72.8 million individual returns by April 17, 2026, 53% of all e-filed returns[21], and the deadline did not move.
Agents can now do most of the preparing: sort the uploads, enter the figures, list what changed since last year, draft the workpaper. But each return still has one signing preparer, each audit one engagement partner, and the reviewers between them have no more hours than before. The question for a firm is whether its reviewers, its client consents and its records can keep up with what agents will draft. This playbook is about getting those three ready first.
The short version
AI-enabled vs AI-native accounting firm
Staff use a chatbot for research and emails, practice management summarizes threads, the tax software imports a few more forms. The calendar, the staffing pyramid, the review chain and the billing stay as they were. Each tool's AI acts under its own login, and nobody can show which return lines an agent touched.
Each service line is designed on the assumption that agents do the first pass of preparation, categorization, testing and chasing, and that people review, decide and sign. Review capacity, client consent and the record of who did what are planned for that volume.
Four things set the AI-native firm apart:
- The firm plans review capacity. It sets thresholds, sampling and a daily limit for each reviewer before agent volume arrives.
- Every agent action runs under a named professional's authority. The agent can do no more than that preparer may do, for that client and year.
- Client data stays in its lane. Return information serves only the return, SSNs stay onshore, and one client's file never informs another's draft.
- Pricing and staffing change on purpose, and clients are told how AI is used.
The missing layer: who reviewed what the agent prepared
No firm will replace its tax software, its clients' ledgers or its practice management system to become AI-native. Agents will work across them, and each product's own AI acts under its own login. None of them can say, across all of them, which agent touched which client's return, for which preparer, and who cleared it. The IRS expects firms to document their AI steps[13], and the Safeguards Rule expects them to log what authorized users do[22].
OrchKernel is a control layer that sits between agents and those systems. Agents ask it before they act. It checks what the agent may do and for whom, holds what needs a reviewer, hides fields the agent should not see, and keeps a tamper-evident record. It is not a tax engine, a ledger, a practice management system or an e-file provider, and it never signs anything. The blueprint near the end maps it to the firm's 18 control points.
Where the hours and margin go
People are the cost
Few accounting firms publish their numbers, but two public ones show the shape. At CBIZ, operating expenses were 86.4% of its $2.758 billion revenue in 2025, down from 88.9% in 2024, and "the majority of our operating expenses relate to personnel costs"[27]. Andersen reported cost of services at 71.0% of revenue in 2025, up from 63.1%, in a year that included its listing[28]. No public dataset breaks down costs for small and mid-sized firms; the AICPA MAP Survey does, for members only.
Four months carry the year
CBIZ tells investors its tax and accounting work has "a heavier volume of activity during the first four months" of the year[27], and extended returns make a second peak in the fall. Hiring cannot follow that curve, so the same reviewers absorb it. That is why every change in this playbook is tied to the calendar.
The pipeline is thinner and dearer
Enrollment in accounting programs rose 12.4% to 266,506 in spring 2025, but degrees awarded fell, and new CPA Exam candidates dropped from 42,626 in 2023 to 28,082 in 2024 after a rush to sit the old exam[32]. Starting salaries reached $67,750 for master's hires and $60,834 for bachelor's hires[31].
Government data is mixed: the occupation is projected to grow 5% over ten years[25], while employment at accounting and tax firms fell 0.6% in the year to September 2026[26]. The series measure different things, so read the gap as a hint.
Hourly billing turns saved time into lost revenue
Andersen says "the substantial majority" of its revenue is billed on time and materials[28]. The AICPA says many firms have moved to value pricing while others still bill hours[40]. Where fees are hours times rate, an agent that cuts preparation time cuts revenue unless pricing changes, and the IRS has now said the savings should reach the client[13]. Growth is slowing at the same time: median net client fee growth was 6.7% in the 2025 MAP Survey, against 9.1% in the previous one[31].
Where the hours go, by role
What AI already does, by practice area
Adoption depends on who asks
Thomson Reuters found 21% of tax firms using generative AI at the organization level in early 2025 and 34% by late 2025, with 14% of tax firms using agentic AI[49,50]. It sells AI tax products and screened respondents for familiarity with AI, so treat those as upper bounds. Among 1,808 CPA tax preparers surveyed in June 2026, 65% use AI for tax research and 16% have no plans to use it, down from 24.3% a year earlier[44].
Firm-level change is rarer. In the 2025 MAP Survey, 13% of firms said they had implemented AI and automation successfully[31]. Karbon, which sells practice management software, reports that 98% of accounting professionals use AI (vendor survey)[51]. The 98% counts people who have tried a chatbot; the 13% counts firms that changed how a return or a close gets done.
Clients mostly do not know. In the 2025 Thomson Reuters survey, about 59% of tax-firm clients did not know whether their firm used generative AI[49].
By practice area
Named products are examples from the sources, not recommendations, and their performance figures are their own.
The case for keeping the tax engine
Language models are poor tax calculators on their own. In TaxCalcBench, frontier models calculated fewer than a third of federal returns correctly[46]; its authors work for a tax software company. A 2026 preprint found the best plain model right on 66% of 51 held-out returns, while model-written code checked against symbolic tax rules got all 51 right[47]. The practical reading for a firm: let agents extract, organize, draft and check, and let the tax software calculate.
Where the money went
Tools for firms got the venture money. Basis raised $100 million at a $1.15 billion valuation in February 2026[53]; Rillet raised $100 million at $1 billion in August 2026 and claims more than 600 customers[56]; Black Ore raised $60 million in 2023[54].
AI-led buyers bought firms. Thrive Holdings owns and runs more than 70 businesses, including accounting and IT services, raised over $2 billion at a $12 billion valuation in August 2026[71], and made OpenAI an owner in December 2025[72].
Private equity bought the mid-tier. Grant Thornton (2024)[75], Baker Tilly, which then combined with Moss Adams[76], Schellman (2026)[74] and Crowe (2026), where Crowe LLP keeps "all attest services" and a new Crowe Advisory LLC does tax and advisory[73]. If the two halves share an AI platform, that is a new independence question.
The best-known AI-adjacent firm failed. Bench, a venture-backed bookkeeping and tax firm, shut on 27 December 2024 after burning $135 million[66]. Its founder raised $10 million in 2026 for Synthetic, pitched as "a fully autonomous AI bookkeeper"[68]. MAVI, an offshore talent marketplace, argues AI will remove entry-level roles and deepen the mid-level shortage[70], a founder's view rather than data.
One return's path
An individual return from upload to the log, in a firm at Stage 3. A CAS close and an audit workpaper follow the same shape.
- 01Uploadยท Client
A couple uploads two W-2s, a brokerage 1099, a 1098 and a partnership K-1.
- 02Sort and indexยท AI agent, checked by OrchKernel
Names each file by year, form and payer, and flags what is missing against last year: a second K-1.
- 03Prepare in the tax softwareยท AI agent, checked by OrchKernel
Enters figures for the named preparer it acts for, each linked to its source page. The software calculates.
- 04Change listยท AI agent, checked by OrchKernel
Lists every line that moved past the reviewer's threshold since last year. The K-1 loss is on it; last year it was income.
- Agents stop here. They have no tool to clear a review, sign, or transmit.05Reviewยท Firm staff and the tax software
The reviewer clears the change list and the sampled fields. Nothing moves until every item is cleared.
- 06Signยท Firm staff and the tax software
The preparer responsible for the return's overall accuracy signs it.
- 07Form 8879ยท Client
The clients review the return and sign Form 8879.
- 08Transmitยท Firm staff and the tax software
Staff at the electronic return originator transmit the return.
- 09Recordยท Firm staff and the tax software
The log links each step to a person: who the agent acted for, what it touched, who cleared, signed and transmitted.
The staged path
Six stages, ordered by how far an agent's work travels: inside the firm, to clients on templates, into a return or ledger under review, and only then, for non-attest clients, on its own. Each stage is also tied to the calendar. Launch in the change window from May to early October, prepare and run in shadow from November, and freeze from mid-January to April 15.
- Jan
- Feb
- Mar
- Apr
- May
- Jun
- Jul
- Aug
- Sep
- Oct
- Nov
- Dec
- Busy season: freeze
- No process changes from mid-January to April 15.
- Change window
- Build, test and launch stages.
- Extension peak
- Extended returns. Finish changes by early October.
- Shadow and prepare
- Run new workflows on last season's files.
- Stage 0Get the file in order
- Stage 1Read, sort and draft inside the firm
- Stage 2Talk to clients on approved templatesFirst contact with clients
- Stage 3Agents prepare; people review and signAgent work enters returns and ledgers
- Stage 4Narrow autonomy for non-attest work
- Stage 5The AI-native operating model
- 0
Stage 0: Get the file in order
Clean client records, write the AI policy, settle the consent question.
When: May to June, or October to December
What to do
- Clean client records: entity, year end, states, an attest flag, authorized contacts, and 2848 or 8821 status by year.
- List every AI feature already switched on, and every personal chatbot account staff use for work.
- Write the AI procedures the IRS now expects: approved tools and models per kind of data, review duties, accuracy monitoring[12,13].
- Add AI vendors to the written security plan, and confirm MFA everywhere[19,22].
- Ask counsel which AI uses need Section 7216 consent.
- Name AI among third-party service providers in the engagement letter[30].
Why now
Extraction is only as good as the client file, and the confidentiality duties apply from the first upload.
In place first
- Nothing. Every firm starts here.
What to measure
- Share of clients with complete records
- Share of staff AI use on approved tools, from a short survey
Common mistakes
- Banning AI outright, so staff use personal accounts without telling anyone.
- Buying a preparation tool before the client file is clean.
- 1
Stage 1: Read, sort and draft inside the firm
Nothing an agent produces reaches a client, a ledger or the IRS.
When: Build in the change window; run in shadow in November; live for the new season with review
What to do
- Request lists from last year's return; upload sorting and a document index with a prior-year comparison.
- K-1 tracking: which partnerships have sent theirs.
- Notice intake: type, tax year, amount and response date, routed to a person.
- Research drafts with citations, PBC lists rolled forward, engagement-letter rollovers.
Why now
This is the chase-and-key work that fills the calendar, and none of it leaves the firm. Users rate data import into tax software 3.4 out of 5[44]. It also builds the accuracy record Stage 3 needs.
In place first
- Agents read one client's file per task.
- A model cleared for return information, every read logged.
What to measure
- Extraction accuracy, sampled weekly against source documents
- Days from organizer to ready for preparation
- Reviewer minutes per return: the baseline for Stage 3
Common mistakes
- Counting documents processed without sampling accuracy.
- Letting a research draft reach a client unchecked. That is the Deloitte failure.
- 2
Stage 2: Talk to clients on approved templates
First contact with clientsReminders and status updates, never advice.
When: Start in September or October, before organizers go out
What to do
- Missing-item reminders listing only what this client still owes.
- Chasing 8879 and engagement-letter signatures; status updates.
- Monthly CAS question lists, drafted for the staff accountant to send.
Why now
Late documents push returns into April and onto extension. And about 59% of tax-firm clients did not know whether their firm used generative AI (vendor survey)[49], so this is also the stage to tell them.
In place first
- Manager-approved templates; anything else goes to a person.
- Return information sent only through the portal.
What to measure
- Days to complete document collection
- Unsigned 8879s ten days before the deadline
Common mistakes
- Letting an agent answer a free-text tax question.
- A reminder that names another client's documents. Test merge fields first.
- 3
Stage 3: Agents prepare; people review and sign
Agent work enters returns and ledgersDraft returns, entries and workpapers, each held for a named reviewer.
When: Configure in the change window; run for the first full season after Stage 1 has a season of data
What to do
- Draft returns in the tax software, which calculates, with a change list and sources for each.
- Proposed entries and reconciliations for CAS; draft workpapers and tests of detail for audit.
- Draft notice responses and the extension list.
Why now
Stages 1 and 2 show where extraction is reliable, and the season brings the volume to measure. In one 2025 benchmark, frontier models calculated fewer than a third of federal returns correctly[46], so the agent drafts around the tax engine, not instead of it.
In place first
- Review thresholds and a daily release limit per reviewer.
- No posting to attest clients' books without their recorded approval.
- Agents act for a named preparer, within that person's access.
What to measure
- Reviewer minutes per return against the baseline
- Review notes per return; returns amended; errors found after filing
Common mistakes
- Control dilution: more returns, the same reviewers, less scrutiny each.
- Treating agent output as staff work under Circular 230 without the supervision records that reliance on others requires[11].
- 4
Stage 4: Narrow autonomy for non-attest work
Pre-approved classes of work run without a person first, then get sampled.
When: After a full season of Stage 3 data; CAS changes in the change window
What to do
- For non-attest CAS clients who agree in writing, agents post recurring transactions that match approved rules (vendor, amount band, account), reviewed afterward.
Why now
It needs Stage 3 data showing low correction rates for each class. Independence rules rule it out for attest clients.
In place first
- Correction rates per class over a full close cycle, and an attest flag checked on every call.
What to measure
- Autonomous postings and corrections per class
- Days to close. One field study found AI-using accountants closed 7.5 days sooner, on data from one software partner[45]
Common mistakes
- Extending a posting class to a client that is also an audit client.
- No sampling of autonomous work once it looks routine.
- 5
Stage 5: The AI-native operating model
Staffing, training, pricing and quality management built around agents.
When: Over two to three years, one change window at a time
What to do
- Redraw the pyramid: fewer data-entry roles, more reviewers and advisers.
- Train juniors on checking agent work, so the experience still counts toward licensure.
- Move to fixed or value fees, with a plain statement to clients about AI.
- Put AI controls into the quality management system and its annual evaluation.
Why now
Without these, the time agents save turns into slack or into lower hourly bills. Of firms in the MAP Survey with a plan for freed capacity, 39% would add clients without adding staff and 45% would reduce hours; a third have no plan[31].
In place first
- Stage 3 running across the main service lines, with the log as the record.
What to measure
- Net client fees per professional and realization
- Busy-season hours, turnover and client retention
Common mistakes
- Cutting junior hiring without a plan for where future reviewers come from.
- Changing fees without being able to show the client what AI did on their work.
Your first 90 days
Written for a firm starting in October 2026. Ninety days ends just before busy season, so the plan ends with a freeze, not a launch. Most of October is partner and admin work, so it can run around the October 15 deadline for extended individual returns.
- Days 1 to 30 ยท October
- Name the person who owns the firm's AI procedures under Circular 230 ยง10.36, and link the role to quality management.
- Write the AI use policy: approved tools and models for each kind of data, no client data in personal chatbots, review duties.
- Ask counsel the Section 7216 question, and add AI to the engagement letter's notice about third-party service providers.
- List the AI features already switched on across your systems.
- Pick two Stage 1 workflows for the coming season, such as request lists with upload sorting, and notice intake.
- Days 31 to 60 ยท November
- Run both workflows in shadow on last season's files: the agent drafts, people compare.
- Record extraction accuracy and every correction.
- Start the log of agent actions by client.
- Clean the client records behind your 50 largest returns: attest flag, 2848 or 8821, authorized contacts.
- Days 61 to 90 ยท December to early January
- Turn on the two workflows for the new season, with review.
- Prepare one Stage 2 workflow, 8879 and missing-document reminders, on approved templates.
- Set up the kill switch and test it.
- Freeze changes by mid-January. Decide what to widen after April 15, using the accuracy and reviewer-time data.
What not to fully automate
An agent can prepare each of these. A named person makes the decision and answers for it, however good the agent gets.
Control dilution: the risk sits at the reviewer's desk
Danny Werfel, a former IRS Commissioner, gives the main risk a name. Control dilution is when "the speed and volume AI introduces causes existing review and sign-off processes to become less rigorous in practice, even if they remain unchanged on paper. More returns get processed with less scrutiny per return." He describes the matching failure at the junior level, when "a junior associate accepts the AI suggestion without independent verification"[43].
Nothing in the firm's procedures changes when this happens, which is why it is easy to miss. The bottleneck and the main risk sit at the same desk. The figure below uses round numbers to show the mechanism.
Illustration, not a forecast: one reviewer through one season
Everything agents prepare goes straight to the reviewer.
- Early February60 returns40 h40 min
- Early March120 returns40 h20 min
- Early April180 returns40 h13 min
Release to review is capped at what the reviewer can clear at 40 minutes a return.
- Early February60 returns40 h40 min
- Early March60 returns40 h40 min
60 of this week's returns wait: add a reviewer, extend, or reprioritize
- Early April60 returns40 h40 min
120 of this week's returns wait: add a reviewer, extend, or reprioritize
What the gate does. It does not make reviewers faster. Each reviewer gets a daily release limit, and every return arrives with its change list, sources and a sample of unchanged fields to check. When prepared work outruns review, the queue grows where a partner can see it, and the firm decides: add a reviewer, extend, or reprioritize. Without the gate, reviewers make that choice silently by spending less time on each return.
Measure reviewer minutes per return from Stage 1. If the number falls in Stage 3, be able to say why, and show that post-filing errors did not rise.
When it goes wrong
The loudest AI failures in this sector so far are reports, not returns. Each real case below is tied to the control point that addresses it.
Deloitte Australia: a government report with invented references
The A$440,000 report from the opening cited papers that do not exist and misquoted a federal judge. Deloitte refunded the final installment[62]. The IRS cites the case in its 2026 AI guidance[13].
The control: Sources checked before anything leaves the firm, and the deliverable waits for a named reviewer (control 12).
Deloitte Canada: the same failure, months later
A health workforce plan for Newfoundland and Labrador, costing nearly CA$1.6 million, contained at least four citations to papers that do not exist[63].
The control: Same as above: sources checked and a named reviewer on every deliverable before it leaves the firm (control 12).
KPMG withdraws a report on agentic AI
KPMG pulled a report on agentic AI in June 2026 after citation problems, and organizations it named said its statements about them were untrue or misleading[64].
The control: Client-facing and public content waits for review with sources checked (control 12).
A KPMG Australia partner fined for using AI on an AI training test
The partner was fined A$10,000, and about two dozen others were found doing the same, as reported in February 2026[65].
The control: Competence shows in reviewed work, not course completions; the log shows who reviewed what.
Bench shuts down overnight
The venture-backed bookkeeping and tax firm closed on 27 December 2024 after burning $135 million; thousands of businesses lost access to their accounting and tax documents until a buyer stepped in[66,67].
The control: Clients' books stay in their own ledgers, exportable (control 18).
Tax professionals as targets
The IRS warns that criminals go after tax professionals for client data "to file fraudulent tax returns"[18]. An agent with standing access to every client file is one more target.
The control: Agents never hold credentials; every read is logged by client (controls 4 and 14).
Agent failures to design against (scenarios)
Scenarios, not reported events, each mapped to the control that stops it.
Rules that reach accounting firms
No rule bans AI in this work, and none moves a duty to the software. The 2026 IRS guidance applies existing Circular 230 duties to AI[13], and the AICPA and Werfel have formed a council to refine his risk framework[41].
The hardest open question is the first. Section 7216 allows disclosure to software contractors but says nothing about a hosted model that drafts work. Until counsel or the IRS says otherwise, treat it as a disclosure that may need the taxpayer's consent. The IRS guidance adds that client data must be handled "using only secure, enterprise-approved AI"[13].
Tax return information: use, disclosure and consent
Practitioner duties and the 2026 IRS AI guidance
Signatures, e-file, penalties and records
Professional standards: confidentiality, independence and quality
Data security and state law
Outside the US, for firms with UK or EU work
What they have in common
Across these rules, four requirements keep recurring. The exact obligation differs by rule and by service, but a firm that builds these four for every client has the base the others extend.
- Use client data for that client's work
- Section 7216, the IRS warning about cross-client data and AICPA client notice[1,13,30]. For agents: one client per task, cleared models per kind of data.
- A named person answers for it
- The signing preparer, the ERO, the managing partner for quality[8,14,35]. No duty moves to the agent.
How roles, training and pricing change
Preparers become checkers of agent drafts
The junior job moves from keying to verifying each figure against its source and explaining every change from last year. That is the job that prevents Werfel's failure case[43], and it has to be taught to people who have never prepared a return by hand.
Reviewers become the constraint to manage
Reviewer hours are the scarce input, so they get planned like a budget: release limits, change lists, sampling, and a weekly look at minutes per return. Partners remain the owners of judgment and signature, unchanged in law; the IRS says "final decisions must always rest with qualified professionals"[13].
A new owner for AI procedures
Circular 230 already puts procedures on whoever has "principal authority and responsibility for overseeing a firm's practice"[12], and SQMS No. 1 makes the managing partner ultimately responsible for the quality system[35]. Someone has to own the approved tools, the agent playbooks and the evidence, and put widely used AI tools through change management[42]. In a 20-person firm that is part of a partner's job.
Juniors still need two years that count
The AICPA and NASBA have added a licensure path of a bachelor's degree, two years of experience and the exam, enacted in 14 states by May 2025[33]. If agents absorb the first drafts that used to be that experience, the firm has to design work that still teaches and still qualifies. No source we found defines what qualifying experience looks like when agents prepare the first draft. Only 5% of firms in the MAP Survey had formal training for staff moving into new roles[31]. Cutting junior hiring without a plan means fewer reviewers in five years.
Offshore teams change role
Offshore staff can cost "as little as 25%" of a US hire, according to one conference speaker[34], an anecdote rather than data. Agents now take on much of the data entry those teams did, and the SSN rule limits what offshore staff can see[4], so field-level masking is what keeps them useful for review support.
The fee debate
The IRS guidance says billing for "time that was not actually spent or double billing for AI-assisted tasks may violate ยง 10.27", and that cost reductions should be credited to the client[13]. The AICPA answers that this ignores value pricing and the cost of AI itself; it has asked the IRS for clarification, and its CEO called the language not authoritative[40]. In the 2025 Thomson Reuters survey, 6% of tax firms expected rates to rise significantly and 37% slightly, and 25% planned to pass AI costs through to all clients (vendor survey)[49].
Whatever your pricing, be able to show a client what AI did on their work and how the fee reflects it. That record is cheap to keep if every agent action is already logged by client.
The OrchKernel blueprint for an accounting firm
OrchKernel is the layer between AI agents and the systems a firm runs on, as described in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened.
What it is not. OrchKernel does not replace the tax software, the general ledger, practice management, the portal or the audit platform; they stay the systems of record. It does not calculate tax, transmit returns or sign anything.
The mechanisms
- Approvals
- The action waits for a named person, who sees exactly what will happen: the change list, the entries, the message, the recipient. It runs once, as approved.
- Rules
- Checked on every call, for example a signed 8879 on record before a transmission request, or no posting for attest-flagged clients. A rule allows, holds or denies, with a reason.
- Acting on a named person's authority
- Each agent acts for a named preparer or accountant, with no more access than that person has. Remove the person's access in OrchKernel and the agent loses it too.
- Data access by role and field
- One client's file per task; SSN fields hidden from offshore roles and uncleared models.
- Tamper-evident audit log
- Every request, decision, approval and result, by client, chained so an edited or deleted entry shows.
- Human queue
- Low-confidence extraction, free-text client questions, notices and unknown senders land with a named owner.
- Kill switches
- Stop agents for one service line, one client or the whole firm, at once.
- Connections to your systems
- The firm connects practice management (such as CCH Axcess Workflow, Karbon, Canopy or TaxDome), the portal, document storage, email, clients' ledgers (QuickBooks Online, Xero, Sage Intacct, NetSuite) and audit tools, through MCP servers or REST adapters. Many desktop tax engines offer little or no API, so that connection may be an import or export file. Agents never hold the credentials.
Eighteen control points
The controls a firm needs whatever tools it uses, who owns each, and what enforces it.
Client data and consent
Authority, signatures and filing
Review and independence
Clients and fees
Records, change and stopping
What belongs elsewhere
- The tax calculation
- Stays in the tax engine.
- E-file, Form 8879 and identity checks
- Stay in the ERO's software and portal.
- The workpaper archive
- Stays in the audit and document systems. OrchKernel's log is not the binder.
- Firm processes
- Independence checks, acceptance, peer review and quality documentation. OrchKernel supplies evidence.
- Legal documents
- Consent wording, engagement letters, AI vendor contracts.
- People's credentials
- PTINs, EFINs, CAF numbers and e-Services logins never go to agents.
OrchKernel is source-available under the Business Source License 1.1 and runs on your own servers, so you can read the code that enforces these controls. Running it in production for your own firm is covered; the license page says what needs a commercial license.
Scorecard by stage
Record the baseline before Stage 1, then track the same numbers at each stage. Public benchmarks barely exist for this sector: the MAP Survey detail is for AICPA members, and other benchmark studies are paid. We have not quoted any number we could not source.
Sources and further reading
Sources were read in October 2026; dates are publication or data dates. Last reviewed October 2026.
Primary sources
Statutes, regulations, IRS and FTC publications, Bureau of Labor Statistics data and SEC filings. Statutes and regulations are read in the Cornell Legal Information Institute copy; the EU AI Act in an unofficial copy.
- 126 USC 7216: disclosure or use of information by preparers of returns. Legal Information Institute, Cornell Law School.
- 226 USC 6713: disclosure or use of information by preparers of returns (civil penalty). Legal Information Institute, Cornell Law School.
- 326 CFR 301.7216-2: permissible disclosures or uses without consent of the taxpayer. Legal Information Institute, Cornell Law School.
- 426 CFR 301.7216-3: disclosure or use permitted only with the taxpayer's consent. Legal Information Institute, Cornell Law School.
- 5Rev. Proc. 2013-14: consent to disclose or use Form 1040 tax return information. Internal Revenue Service, Internal Revenue Bulletin 2013-3, January 2013.
- 626 USC 6694: understatement of taxpayer's liability by tax return preparer. Legal Information Institute, Cornell Law School.
- 726 USC 6107: tax return preparer must furnish copy of return to taxpayer and must retain a copy or list. Legal Information Institute, Cornell Law School.
- 826 CFR 1.6695-1: other assessable penalties with respect to the preparation of tax returns. Legal Information Institute, Cornell Law School.
- 9Rev. Proc. 2024-40: inflation adjustments, including section 6695 penalties for returns filed in 2026. Internal Revenue Service, October 2024.
- 10Rev. Proc. 2025-32: inflation adjustments, including section 6695 penalties for returns filed in 2027. Internal Revenue Service, October 2025.
- 1131 CFR 10.22 (Circular 230): diligence as to accuracy. Legal Information Institute, Cornell Law School.
- 1231 CFR 10.36 (Circular 230): procedures to ensure compliance. Legal Information Institute, Cornell Law School.
- 13Introductory guidelines for responsible AI use in federal tax practice. IRS Office of Professional Responsibility, 24 June 2026.Guidance, not a regulation
- 14Publication 1345: handbook for authorized IRS e-file providers of individual income tax returns. Internal Revenue Service, Rev. 12-2025.
- 15Transcript Delivery System (TDS). Internal Revenue Service.
- 16About Form 2848, Power of Attorney and Declaration of Representative. Internal Revenue Service.
- 17e-Services. Internal Revenue Service.
- 18Data theft information for tax professionals. Internal Revenue Service.
- 19Protect your clients; protect yourself. Internal Revenue Service.
- 20Instructions for Form W-12, IRS Paid Preparer Tax Identification Number (PTIN) application and renewal. Internal Revenue Service, revised October 2025.
- 21Filing season statistics for week ending April 17, 2026. Internal Revenue Service, April 2026.
- 2216 CFR 314.4: elements of an information security program (FTC Safeguards Rule). Legal Information Institute, Cornell Law School.
- 23FTC Safeguards Rule: what your business needs to know. US Federal Trade Commission.
- 2417 CFR 210.2-06: retention of audit and review records. Legal Information Institute, Cornell Law School.
- 25Occupational Outlook Handbook: accountants and auditors. US Bureau of Labor Statistics, data for 2025.
- 26Employment Situation, Table B-1 (accounting, tax preparation, bookkeeping and payroll services). US Bureau of Labor Statistics, September 2026, preliminary.
- 27CBIZ, Inc. annual report on Form 10-K for 2025. US Securities and Exchange Commission, EDGAR, filed 26 February 2026.
- 28Andersen Group annual report on Form 10-K for 2025. US Securities and Exchange Commission, EDGAR, filed 27 March 2026.
- 29EU AI Act, Article 4: AI literacy. artificialintelligenceact.eu (unofficial copy of Regulation (EU) 2024/1689), applies from 2 February 2025.
Industry bodies and independent research
The AICPA Code, the Journal of Accountancy and The Tax Adviser (both AICPA & CIMA publications), IESBA, and academic research.
- 30AICPA Code of Professional Conduct. AICPA, updated through September 2026.ET 1.150.040, 1.295.120, 1.295.143 and 1.700.040
- 31MAP Survey finds big jumps in CPA firm starting pay. Journal of Accountancy, September 2025.Summary of the AICPA/CPA.com MAP Survey; the full survey is for members
- 32The accounting graduate pipeline: where do things stand?. Journal of Accountancy, October 2025.
- 33AICPA, NASBA approve new CPA licensure path. Journal of Accountancy, May 2025.
- 34The case for outsourcing: 3 ways to win. Journal of Accountancy, July 2025.
- 35AICPA unveils new QM resources to help firms meet Dec. 15 deadline. Journal of Accountancy, August 2025.
- 36PCAOB postpones effective date for new quality control system. Journal of Accountancy, August 2025.
- 37PCAOB finalizes simplified quality control amendments. Journal of Accountancy, September 2026.
- 38PCAOB publishes guidance related to audit evidence amendments. Journal of Accountancy, October 2025.
- 39AICPA proposes changes to independence rules related to private equity. Journal of Accountancy, December 2025.
- 40AICPA seeks IRS clarity on AI guidelines, CPA fees. Journal of Accountancy, September 2026.
- 41AICPA, former IRS commissioner to lead initiative on AI in tax. Journal of Accountancy, September 2026.
- 42New checklist helps CPAs manage AI cyber risks. Journal of Accountancy, October 2026.
- 43A risk framework for AI use in tax administration and preparation. Danny Werfel, The Tax Adviser, August 2026.
- 44
- 45Human + AI in accounting: early evidence from the field (working paper 4261). Jung Ho Choi and Chloe Xie, Stanford Graduate School of Business, 7 May 2025.Transaction data from one AI software partner
- 46TaxCalcBench: evaluating frontier models on the tax calculation task. arXiv 2507.16126, 22 July 2025.Authors work for a tax software company
- 47Law and order: tax law autoformalization. arXiv 2610.02792, 2 October 2026.Preprint, 51 held-out returns
- 48Final pronouncement: technology-related revisions to the Code. International Ethics Standards Board for Accountants, 11 April 2023.
Vendor sources
Published by companies that sell AI or software to accounting firms, including the Thomson Reuters surveys. Directional, not an industry benchmark.
- 492025 Generative AI in Professional Services report. Thomson Reuters Institute, 2025.Vendor sourceSurvey by a company that sells AI tax products; respondents screened for familiarity with AI
- 502026 AI in Professional Services report. Thomson Reuters Institute, 2026.Vendor sourceSurvey by a company that sells AI tax products
- 51
- 52
- 53
- 54Black Ore emerges from stealth with $60 million in funding. Black Ore, 7 November 2023.Vendor source
- 55
- 56
- 57
- 58
- 59
- 60
- 61
Company and press
Company announcements, news coverage and encyclopedia summaries.
- 62Deloitte to partially refund Australian government for report with apparent AI-generated errors. Fortune, 7 October 2025.
- 63Major N.L. healthcare report contains errors likely generated by A.I.. The Independent (Newfoundland and Labrador), 22 November 2025.
- 64KPMG pulls report on AI usage due to apparent hallucinations. TechCrunch, 13 June 2026.
- 65
- 66Bench burned through $135 million before shutting down. TechCrunch, 5 February 2025.
- 67
- 68Khosla Ventures is betting $10M on Ian Crosby, whose last startup Bench imploded. TechCrunch, 14 May 2026.
- 69With Tabby, a former accountant is using AI to make accountants obsolete. TechCrunch, 21 September 2026.
- 70MAVI bets on the AI boom creating demand for a new kind of accountant. TechCrunch, 28 September 2026.
- 71Thrive Holdings fundraise. Thrive Holdings, 12 August 2026.
- 72Thrive Holdings x OpenAI. Thrive Holdings, 1 December 2025.
- 73Crowe partners with private equity; Baker Tilly on the move. Journal of Accountancy, June 2026.
- 74Top 50 firm announces new majority private equity investment. Journal of Accountancy, March 2026.
- 75
- 76