A managing partner of a 200-client Estonian accounting firm said this after a long product demo: "If we automate, my accountants have to absorb 2.5 times more clients for the same pay. They will leave. There is no way."
He had run the math correctly. His firm bills per client, pays per hour worked, and competes on price with a dozen similar firms in a mid-sized city. The automation under discussion would compress roughly 40 hours of monthly bookkeeping per client to about 8 — which, from his seat, reads less as margin expansion than as a staffing problem. It is a clean example of a pattern that recurs across established firms: the partners want the technology and still can't make the math work inside their own businesses.
Disclosure: I'm building Artifi, an AI tooling layer for finance teams, so I have a stake in this market. The cases below are field observations, presented as I found them.
Enthusiasm is real; adoption is not
Read only the headlines and accounting looks mid-transformation. The 2024 Karbon State of AI report found 85% of professionals "excited or intrigued by AI" (Karbon), and Thomson Reuters reported that firms with AI strategies are twice as likely to see AI-driven revenue growth (Thomson Reuters).
The second paragraph of those reports complicates the first. 57% of firms provide no formal AI training; only 24–27% believe they have the talent, systems, or regulatory readiness to deploy AI; and 66% of professionals feel overwhelmed by their work technology at least weekly (AICPA). The most telling line in the Karbon data: owners and partners are more likely to embrace AI, while individual contributors are more skeptical. The person who signs the check is enthusiastic; the person who has to use the tool, and whose role changes if it works, is not.
Three firms, three blockers
I engaged three Estonian firms over twelve months. The pattern was consistent, the specifics different.
Firm A (~10 employees, annual reports plus SMB bookkeeping). The partner wanted to offer an AI product to clients but required full UI control, hidden upstream technology, and full revenue capture. He was protecting the firm's main moat — being the trusted face on the relationship — which a visible piece of software undermines. The deal couldn't be structured to fit that.
Firm B (~30 employees, more sophisticated SMB practice). Good product feedback over many meetings, but the bookkeepers — generalists on Merit and Procountor — got stuck on the mental model rather than the documentation. The "ask the agent and iterate" loop is foreign to people who have spent fifteen years pattern-matching on transaction descriptions; that gap doesn't close in a 90-minute session.
Firm C (the one above — 200 clients, 11 accountants). The most AI-literate of the three, and the clearest refusal, because the firm's compensation, pricing, and client expectations are all wired against the efficiency gain.
Three thoughtful partners, zero adoption — each responding rationally to the structure of the business, not to the technology. (These firms are also the "Profile A" of a separate piece on buyer profiles.)
The four structural blockers
1. Billable-hour math penalizes automation. When a large share of routine work can be automated, time-based billing stops tracking value. A Deloitte study found 67% of professional-services buyers now prefer fixed-fee arrangements, up from 41% three years earlier (Chief AI Officer) — but if a firm shifts to fixed fees, it takes the price compression at once without the volume gains. Either way, year one of adoption can look like a revenue hole.
2. Staff resist, rationally. The contributor's role and security sit downstream of the work the model absorbs. Even with the same headcount and a larger book per accountant, Firm C's owner expected his team to leave — and in a tight labor market for bookkeepers, that is a credible response, not obstinacy.
3. The pipeline is thin. First-time CPA candidates and accounting graduates have declined over the past decade (the figures are in the rollup piece). If the AI works, there isn't an incoming class to redeploy toward advisory.
4. Succession is strained. The average partner is in their early fifties and fewer than half of firms have a written succession plan (Baker Thornton). A partner a few years from payout has little incentive to bet the final earnings cycle on a multi-year transformation.
Stacked together, these aren't a culture problem; they're a structural one.
This is textbook Christensen
Clayton Christensen's The Innovator's Dilemma described how capable incumbents fail against disruptive technology not from foolishness but because every internal incentive points toward serving their best existing customers the way those customers currently expect (HBS). The disruption usually arrives at the bottom of the market, looks worse than the incumbent product, and is priced in a way the incumbent can't match without damaging its own P&L.
Mapped onto accounting: AI-driven bookkeeping enters at the bottom, serving SMBs the established firm doesn't prioritize; it looks worse than human service; and it's priced as software ($99–$500/month) rather than labor. The incumbent can't match that price without cutting partner draws, so it doesn't try — and explains, correctly, that its clients want a human relationship. Each statement is true. The same was true of mini-mills, PCs, and digital photography. The framework predicts the rational incumbent response, which is exactly what the partners describe.
Two categories are being built for the long tail
AI-native firms. Pilot offers a $99/month AI-only bookkeeping tier alongside higher plans with CPA review and CFO services (Pilot). Bench, an earlier attempt, shut down in 2024 after accuracy issues (John Galt Finance) — instructive but not disqualifying; newer entrants like Truewind, Fondo, and Juno (which raised $12M to apply AI to tax returns (Crunchbase News)) price as software. Their advantage is no legacy book of human-hour revenue to defend — and, in the land-and-expand pattern, a narrow entry product that can widen over time.
PE-backed rollups. Structurally, the rollup escapes the dilemma: when the holdco automates, the savings accrue to the holdco rather than the staff, the aging partner gets a clean exit, and the sponsor has both a horizon and a mandate to deploy AI across the platform. The deal data is in the dedicated piece; the relevant point here is that this is the ownership structure in which the AI-adoption math works inside the existing firm.
Options that exist for an independent partner
None are easy. The ones that recur in conversations: sell now while multiples are elevated (Inside Public Accounting and CPA Trendlines note 8–12x EBITDA for desirable mid-market firms, beginning to compress as seller supply grows); reposition around fixed-fee advisory and let commodity bookkeeping go (long prescribed, rarely executed, because it requires retraining staff and shedding clients at once); partner with an AI-native platform as a white-label or referral channel (what Firm A wanted, with the trade-off that the platform may end up owning the relationship); or run the firm as a cash machine and let it shrink with the partner's retirement (what many will choose by default).
Outlook
The likely shape over 5–10 years: the Big 4 persist on capital and audit moats; much of the $5M–$100M mid-market consolidates into PE-backed platforms; and the squeezed segment is the 10–50-employee, partner-led, billable-hour firm — pressured by AI-native shops at the bottom and rollup acquisition at the top. The throughline isn't that AI is magical. It's that a business model's internal incentives have drifted out of line with its long-run survival, and the technology is the catalyst that surfaces the contradiction.