Your busiest preparer spends two to three hours on data entry before doing a single minute of actual tax work on a return. W-2s, 1099s, K-1s, brokerage statements, prior-year organizers — typed, checked, and re-checked by hand while the clock burns billable review time. Now investors are betting that workflow is finished: in April 2026, Juno, a CPA-founded tax prep automation platform, raised a $12 million seed round led by Bonfire Ventures with Impression Ventures and Xfund participating, on the claim that its AI co-pilot automates 90 percent of that data entry across more than 90 document types and cuts total time per return roughly in half. Whether you ever buy Juno or not, the raise marks the moment AI co-pilots for tax prep went from experiment to budget line — and this busy season is when your firm decides how to respond.
This post covers what the funding signals, what co-pilots reliably do today, the compliance frame the IRS has now put around them, and how to evaluate one before your staff start pasting client data into whatever tool they found first.
What the $12 million actually signals
Seed rounds do not validate products; they validate pain. Bonfire's investment thesis is blunt: tax preparation is still dominated by manual workflows and decades-old systems, and the most painful parts — intake, document handling, first-pass data entry — are exactly what large language models are good at. Juno says its platform shrinks a two-to-three-hour prep process to roughly seven to ten minutes of staff time, with a human still in the loop on review, and that it handles the notoriously complex K-1s that consume high-net-worth practices. The company plans to expand from individual returns into business returns — Forms 1120 and 1065 — which tells you where the category is heading: full-firm coverage, not a 1040 side tool.
The signal for your firm is not "buy this vendor." It is that venture money now agrees with what your senior preparers have told you for years: the data-entry layer of tax prep is undifferentiated labor. Firms that keep paying senior-staff rates for it will compete against firms that pay software rates for it and redeploy the savings into review and advisory — the work clients actually notice.
Adoption data backs up the timing. The second annual Blue J and CPA.com AI Tax Research Solution Outlook Report, based on a June 2026 survey of more than 1,000 US tax professionals, found 60 percent now use AI for tax research at least weekly, up from 33 percent a year earlier. The Journal of Accountancy's 2026 tax software survey of 1,808 AICPA members found 65 percent using AI in tax research and 32 percent in client communication. Research led the way; preparation is next, and the Juno raise is capital arriving ahead of that demand curve.
What co-pilots reliably do today
Strip away the marketing and today's tax-prep co-pilots do four things well:
Document intake at scale. Upload a shoebox of PDFs and the tool classifies each one — W-2 versus 1099-NEC versus K-1 versus brokerage statement — and extracts the fields into a structured draft. Juno's headline number is 92-plus document types; competitors cluster in the same range. This is the 90-percent-automation claim, and it is the most mature capability in the stack.
Prior-year variance flags. The co-pilot compares the draft against the client's prior-year return and flags what changed or looks inconsistent: a missing K-1 the client had last year, a W-2 total that moved oddly, a new state. This is genuinely useful review leverage — it points your reviewer at the five lines that matter instead of all five hundred.
First-draft review notes. Some tools draft the review points themselves, in the firm's own checklist format. Treat these as a starting point, not a workpaper: they are only as good as the extraction underneath them.
Advisory surfacing. A few platforms flag planning opportunities visible in the draft — an underfunded retirement account, a Roth conversion window, an estimated-payment shortfall. This is the upsell engine vendors pitch hardest, because advisory is where firms want the reclaimed hours to land.
What they do not do is sign the return. Every serious vendor keeps a human in the loop, and as the next section explains, the IRS now requires that posture in substance if not in those exact words. A co-pilot that drafts and a preparer who reviews is the model; a co-pilot that files while nobody looks is a malpractice fact pattern.
Review capacity still decides quality
Here is the mistake early adopters make: they automate prep, watch per-return hours fall 50 percent, and conclude they need fewer reviewers. The 2026 survey data says the opposite. Financial Cents' 2026 survey of 486 North American professionals found 95 percent use AI at some level but only 11 percent run it firmwide, and just one in five report measurable ROI. The gap between dabbling and returns is review discipline — firms that treat AI output as draft-quality and staff review accordingly get the savings; firms that treat it as finished get the errors.
Think of it as conservation of attention. Before the co-pilot, your preparer spent two hours typing and one hour thinking. After it, the typing is seven minutes — but the thinking still takes an hour, and now it has to be sharper, because the errors that survive extraction are subtler than typos. A transposed digit is easy to spot; a correctly extracted K-1 coded to the wrong activity is not. Budget your reclaimed hours accordingly: some go to taking on more returns, but a protected share should go to deeper review and to the advisory conversations the draft surfaces.
A practical rule: pilot the co-pilot on your simplest 1040s first, measure review time per return separately from prep time, and only expand the document types once review time per return is flat or falling. If review time is rising, your reviewers are catching extraction errors — which means the tool is working as a draft engine and your review step is doing its job, not that the tool has failed.
The compliance frame: OPR Alert 2026-19 and your existing duties
On June 24, 2026, the IRS Office of Professional Responsibility issued Alert 2026-19, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice." Read it before you buy anything. Its core message is that no new AI-specific ethics regime was needed, because your existing duties under Treasury Circular 230 — competence, due diligence, confidentiality, and firm oversight — already apply in full to AI-assisted work. Generative AI can hallucinate, embed bias, and reason opaquely; using it does not lower your standard of care by one notch.
Three existing rules do most of the work:
Circular 230 due diligence and firm procedures. You must verify AI output before relying on it, and firm owners must maintain procedures covering staff training on AI risks, secure data handling, accuracy monitoring, vendor vetting, and documentation. An AI-generated memo sent to a client without human review is a quality-control failure regardless of which tool produced it.
Section 7216. The Internal Revenue Code's disclosure rules carry criminal penalties for preparers who knowingly or recklessly disclose or misuse taxpayer information, with a parallel civil penalty under Section 6713. Uploading a client's return data to a consumer AI tool whose terms let the vendor retain or train on inputs is a disclosure — paying for the tool does not change that. Only enterprise or business-tier accounts covered by a contract with your firm, with training off and retention limited, belong anywhere near client data.
The FTC Safeguards Rule and IRS Publication 4557. Your written information security plan — the IRS provides a template in Publication 5708 — must cover every AI tool as in-scope systems: access controls, encryption, logging, and vendor management. Adding a co-pilot without updating the WISP is the compliance equivalent of hiring a seasonal preparer and skipping the background check.
Set five rules before staff touch any AI tool: no client PII in consumer tools, model training on your data turned off, retention limited by contract, access controlled through SSO and MFA, and an approved-tool list with everything else blocked. Put the list in the WISP and train to it — the alert explicitly names staff training as a firm-procedure duty.
How to evaluate a co-pilot before busy season
Run your evaluation in November and December, not March. A structured pilot takes about 30 days, and you want the verdict before organizers start arriving.
Match the document list to your practice. Vendors quote total document types; you care about your document types. Pull last season's intake log and check the top twenty forms by volume. If 15 percent of your individual returns include K-1s and the tool mangles them, a 90-type count is irrelevant. Test with your ugliest real documents — scanned, skewed, multi-page brokerage statements — not the vendor's sample PDFs.
Score extraction separately from review. For a sample of 20 to 30 returns, record prep time, extraction error count by severity, and review time. The buying question is not "is it perfect" but "is review of its drafts faster than prep from scratch, at equal or better accuracy." Most firms find the answer is yes for W-2 and 1099-heavy returns and closer to break-even for K-1 and multi-state returns in year one.
Verify the data-governance terms in writing. Confirm training is off, retention periods, subprocessors, SOC 2 status, and what happens to your data if you cancel. If the vendor cannot produce these in the sales cycle, it will not produce them during an incident either.
Check the export and integration path. Your co-pilot must push drafts into your existing tax software and your document management system without retyping. A tool that drafts beautifully but exports as a flat PDF has automated nothing — it just moved the typing from your preparer to your reviewer.
Do the per-return math honestly. Price the tool per return at your actual volume, add implementation and training time, and compare against loaded staff cost per return today. Common mistakes: pricing at the vendor's assumed volume instead of yours, forgetting the review hours that remain, and ignoring the cost of running two workflows during the transition season. If the math only works at triple your volume, it does not work.
Watch for cross-client leakage. OPR's alert specifically flags the risk of one client's data surfacing in another client's output. Ask the vendor how tenant isolation works and whether any firm-level learning crosses client boundaries. Get the answer in the contract, not the demo.
Book the experiment like an experiment
A co-pilot pilot is a capital-allocation decision wearing a software costume, so give it its own ledger lines. Book the subscription under a dedicated software account — not lumped into office expense — and track the advisory revenue the reclaimed hours produce separately from compliance revenue. Twelve months later you want to answer "did the tool pay for itself" from the general ledger, not from vibes. If you keep your firm's books in plain text, the structure looks like this — see the Beancount documentation for the full syntax:
2026-01-01 open Assets:Cash:Operating USD
2026-01-01 open Expenses:Software:Tax-Prep-AI USD
2026-01-01 open Expenses:Professional-Development:Staff-Training USD
2026-01-01 open Income:Services:Tax-Prep USD
2026-01-01 open Income:Services:Advisory USD
2026-02-01 * "Juno" "AI tax-prep co-pilot, annual seat"
Expenses:Software:Tax-Prep-AI 1200.00 USD
Assets:Cash:Operating
2026-02-15 * "Client payment" "1040 prep, review billed; entry automated"
Assets:Cash:Operating 450.00 USD
Income:Services:Tax-PrepTwo habits make this pay off. First, split implementation costs — training hours, parallel-run review time — into the training and software accounts during the pilot quarter, so the true first-year cost is visible. Second, when a planning opportunity the draft surfaced turns into an engagement, book it to the advisory account. The ratio of advisory growth to software spend is the ROI number your partners' meeting actually needs.
Your 30-day rollout plan
November is evaluation month. Week one: publish the five data rules, update the WISP's approved-tool list, and pull last season's intake log to build your test-document set. Week two: run two vendors head-to-head on the same 20 returns and score extraction errors and review time separately. Week three: check references from firms your size, verify the security terms in writing, and do the per-return math at your volume. Week four: sign one vendor, train reviewers first — they are the control point — and schedule the preparer rollout for early January with the simple-1040 population.
Juno's $12 million says the capital markets believe data entry is done as a human task. Your busy season will prove or disprove that for your practice exactly once — run the pilot now, keep review capacity sacred, and let the ledger tell you whether the future arrived on schedule.
