contents
If US bookkeeping services were a country, its GDP would be #88 in the world, ahead of Uganda, behind Azerbaijan.
The job falls in a category called the tedious-middle. “Too high-stakes to hand to an amateur and too tedious for a specialist to enjoy.” These are the most underloved jobs in the economy. They require a real qualification to enter, but then they bury you in repetitive work. A nurse doing eight hours of discharge calls has to remember which medications can be mixed and which can’t. A junior associate redlining a 200-page contract has to spot the adverse change clause hiding in section 9. A bookkeeper has to know which sales tax line is which, or someone will have to refile in March.
The first wave of venture-scale AI targeted the obvious bottom of the labor pyramid: customer service, copywriting, personal assistants. The second wave is harder to see because it’s going inside jobs that you need a credential to start.
What the middle looks like for AI
The category has three properties.
It requires real expertise. You can’t hand bookkeeping to a random person off the street and trust the output. Same goes for nursing, medical coding, radiology triage, or legal due diligence. There’s training behind each of them, usually a credential, and real risk when the work is wrong.
It is, for the practitioner, not mentally stimulating. The associate scouring through contracts for one clause is not doing what law school promised them, and the bookkeeper categorizing the 500th Stripe payout is not having a good time. On a typical Tuesday, neither of them is using the skills that got them qualified for their job.
And the work has to be correct, because so much downstream depends on it. A medical coding mistake propagates into a denied claim and a 90-day rework. Too high-stakes to hand to an amateur and too tedious for a specialist to enjoy.
These examples are less about jobs being lost to AI and more about where the next venture-scale opportunities live in 2026. It’s different than the bottom of the skill ladder, which has been getting absorbed for a couple years already. Customer service is the cleanest case, with the likes of Sierra and Intercom Fin building real businesses there. Autonomous driving is on its own slower clock for different reasons. Waymo is real, but it’s also nearly a decade in and still city-by-city. In contrast, the specialist top, like brain surgery and the genuinely engaging end of expert work, sits on the opposite side: the role is challenging and the specialist actually wants the work. The middle is where the next wave of accuracy-grade but tedious workflows is opening up, and it’s still under-built relative to the unit economics.
Four proofs at scale
Pilot is the cleanest proof for bookkeeping. Founded 2017, last priced at roughly $1.2 billion in 2021 in an extension led by Whale Rock + Bezos Expeditions on top of Sequoia’s earlier lead, and meaningfully more AI-shaped today than at that round. Bench Accounting shutting down at the end of 2024 is proof: the human version of SMB bookkeeping had broken unit economics before AI was even meaningfully pitted against it. The demand was real, but the cost structure didn’t really work. Other increasingly commoditized services have this in common. Now Intuit Assist is rolling autonomous bookkeeping agents inside QuickBooks, and Intuit itself reported 68% of US small businesses were using AI in 2025.
Harvey is the proof for legal. $11 billion valuation in March 2026 on a $200M raise co-led by GIC and Sequoia. ARR went from about $50M at the end of 2024 to $100M in August 2025 to $190M in January 2026. Used by 140K+ lawyers across 1,500 organizations. Legal due diligence is a sibling tedious-middle workflow: expertise-gated and accuracy-grade, but mind-numbing for the associate doing it.
I’ve written about Hippocratic AI recently as a nursing-adjacent example. A $3.5 billion valuation after a $126M Series C in November 2025. Their product is a voice agent that does the bounded inpatient workflows that nurses describe as the most draining part of the job: pre-op prep calls, discharge instructions, chronic-disease follow-ups. Note that they’re not selling a robotic nurse. The bedside is untouched. What they focus on is the part that wears the practitioner down and demands accuracy without demanding presence. (Moxi, Diligent Robotics’ physical hospital robot, got pulled from MultiCare Health System in 2025 because nurses found it annoying and unhelpful.)
Aidoc is the proof for radiology. 31 FDA-cleared authorizations by early 2026, including a recent foundation-model clearance covering 14 acute indications on abdominal CT triage. The FDA cleared 1,104 radiology AI devices through the end of 2025, roughly 76% of all AI-enabled medical authorizations. Radiologists are still the ones making the diagnoses; the triage aspect is what Aidoc takes off their plate.
”This has been about to happen for 30 years”
QuickBooks shipped in 1992. The “AI will automate bookkeeping” conversation is older than my career. The reasonable skeptic asks why now is different from then. The answer is that LLM agents can finally handle the long tail of edge cases that rule-based automation couldn’t get quite right, like the weird vendor names and the ambiguous spend amounts. Intuit’s own 2025 data shows 68% of US small businesses using AI. Intuit shipped autonomous bookkeeping agents inside QuickBooks during the past 12 months. 70% of US health systems plan to expand AI medical-coding automation in 2026. Harvey went from $50M to $190M ARR in 13 months. The proof is in the pudding.
The other reasonable objection is that the tightest accuracy-gated work is the most regulated. You can’t ship a clinical workflow the same way you would with a Chrome extension. That’s true. But it’s also the moat. The 1,104 radiology AI clearances are proof that the regulatory path exists and that companies who figure it out first get to compound while competitors are still drafting their pre-submission.
What this means for the founder picking a target
If you are picking what to build in 2026 and your ambition is venture-scale, run your candidate workflow through the three conditions before anything else: does it require real expertise, is it unrewarding for the practitioner, and is the accuracy floor high enough that mistakes cascade downstream. If all three hold, it’s a painkiller idea with a reasonably strong moat. If any one is missing, it might still work, but the customer’s willingness to pay or your protection from competitors may not be as robust.
In all these successful examples, the company picked a single workflow inside a credentialed role and replaced just that one component. I think there are plenty of similar opportunities still out there. Compliance reconciliation, audit prep, supply-chain documentation, prior authorization, fraud-investigation casework. Each of them is somebody’s idea of the worst part of their job. If you hear someone say “I wish a computer would just do this part for me,” you’re most likely listening to a TAM that hasn’t been priced yet.