The Next Million-Dollar Agent Is Boring and Domain-Specific.

Here’s the divide, in numbers. MIT’s “GenAI Divide” study of 300+ enterprise deployments found 95% of AI pilots delivered no measurable P&L impact — generic tools, slick in demos, brittle in workflows. Meanwhile the domain specialists are printing revenue: Harvey (legal) went from ~$100M ARR in late 2025 to $300M ARR by May 2026 at an $11B valuation; Sierra (customer service) crossed $150M ARR; Abridge (clinical notes) passed $100M ARR at a $5.3B valuation; Decagon and EvenUp sit in the billions. Same foundation models as everyone else. The difference is the shape of the product: one bounded, paid workflow — prior-auth review, contract diligence, claim coding — with an existing definition of “correct.” That last property is everything: a domain agent can be verified, and a verifiable agent can be sold. The general agent sells a demo; the domain agent sells an SLA.

The same MIT data says the how matters as much as the what: specialized, vendor-led deployments succeeded about 67% of the time versus ~33% for internal generic builds — and the winners were the ones embedded into a single workflow with memory and learning loops. That is precisely the playbook below.

What follows is the full build playbook — data, training, answering, evaluation, deployment — with the parts connected the way they actually connect: everything downstream is compiled from one curated corpus, and everything in production flows back into it.

01The data engine: collect, curate, compile

Collect from where correctness already lives: the domain’s documents (policies, contracts, codes, rate tables), its decisions (historical cases with outcomes — the claim and whether it was approved), and above all its expert traces — watch the specialist do the workflow and capture every lookup, every rule applied, every “this one’s weird, escalate.” Ten annotated expert traces are worth ten thousand scraped pages, because they encode the procedure, not just the facts.

Curate like the product depends on it, because it is the product. Deduplicate near-copies, resolve version conflicts (the 2024 policy contradicts the 2026 one — which governs?), attach provenance to every unit, and label a hard core: gold (expert-verified), silver (outcome-verified from history), bronze (plausible, unverified). Most teams drown their gold in bronze; keep the tiers separate forever — they have different jobs.

Compile one corpus into three artifacts. This is the step that makes the economics work — the same curated data becomes:

Then close the loop: every production run emits a trace; triaged traces become new silver; escalations the human resolved become new gold; every incident becomes a permanent eval case. The flywheel is the business — six months in, your corpus is something no competitor can start from.

02Training: climb the ladder, don’t jump to the top

• Rung 1 — scaffold, don’t train. A strong base model + the retrieval index + well-designed domain tools (tools are prompts) + domain skills files (procedure + pitfalls + checks). This rung is embarrassingly effective and infinitely cheaper to iterate; most products should ship here first — and the field data backs it: enterprise practice has converged on hybrid (tune for structure and procedure, retrieve for facts), and Contextual AI reports specialized RAG agents beating generic RAG by nearly 30% accuracy in enterprise deployments before any fine-tuning enters the picture.
• Rung 2 — SFT on gold traces. When the scaffold plateaus, fine-tune on the expert procedures — the model learns the shape of the workflow: which lookup first, when to escalate, how to phrase the output.
• Rung 3 — preference-tune on outcomes. Silver pairs (approved vs. rejected, correct code vs. miscode) are natural preference data — DPO-style tuning is cheap and, as I’ve written from experience, watchable on a single GPU.
• Rung 4 — distill for the margin. Once the big model + scaffold hits target accuracy, distill the workflow into a small model for the 80% of easy cases and route the hard 20% up. Unit economics decide rung 4, not benchmarks.

The discipline: never climb a rung without an eval delta that pays for it. The golden set (Section 04) is the ladder’s judge.

03Answering: multiple paths, one verified result

Domain answers are exactly where single-shot generation is malpractice. The pattern that works: fan out K independent paths — each with its own retrieval slice and plan, so their errors decorrelate — run every candidate through domain verification (schema checks, rule engines, cross-referencing the cited clause actually says what the answer claims), and let a judge reconcile the survivors into one result. What reaches the front end is a contract, not a string: answer + evidence links + agreement score + escalate-or-not (the autonomy dial decides that last field).

The explorer below runs 300 simulated domain queries through the pipeline. Dial the paths up and toggle verification — the accuracy, the cost, and the escalation rate tell you what each piece buys:

And notice what the escalation column really is: not a failure budget but a routing decision — the 4–8% of cases that reach a human arrive pre-packaged with all K candidates, their evidence, and where they disagreed. That turns review from “redo the work” into “break the tie,” which is why a domain agent that escalates well often clears ROI before its accuracy is even best-in-class.

04Evaluation: the golden set is the moat

Watch what the winners actually build. Harvey — the $300M-ARR legal agent — publishes BigLaw Bench, a benchmark built from real billable legal work rather than bar-exam trivia, graded by expert rubrics across structure, style, substance, and hallucination; their open-source Legal Agent Benchmark found frontier models completing less than 10% of real legal tasks end-to-end under an all-pass standard. Read that twice: the market leader’s core public artifact is not a model — it’s an eval. That’s not marketing; it’s the moat announcing itself.

Everything in this playbook is steered by one asset: a golden eval set of a few hundred expert-verified cases, stratified by difficulty and case type, refreshed with every production incident. It gates the training ladder, compares vendor models overnight, feeds the trust ratchet per case-type, and — because it grades traces, not just answers, using the trace-metric families — it catches the agent that gets right answers by the wrong procedure, which in regulated domains is its own kind of wrong.

05Deploy, then reproduce

• Roll out on the ratchet. Shadow mode (agent answers, human decides, agreement measured) → assisted (agent drafts, human approves) → autonomous per case-type as verified streaks accumulate. Demote on failure, per class, automatically.
• Watch it like a system. Fleet dashboards, fingerprint drift, incident replay — the five visualization layers, with drift alarms tuned to the domain’s seasonality (tax agents drift every January by design).
• Reproducibility is a feature customers buy. Pin and version everything — the exact model revision (never a floating alias a provider can rotate underneath you), the system-prompt hash, the tool-registry version, the index build, the rule set — so any past answer can be re-derived exactly, with its evidence trail. Regulated buyers audit on four questions: what went in, what came out, what the system did, and whether it was allowed. In audited industries this isn’t hygiene, it’s a checkbox on the procurement form.
• Budget the regression tax. Every model upgrade replays the golden set and a sample of historical production traces before it touches a customer. Upgrades that move the fingerprint get quarantined to shadow mode first.

06Where the thesis breaks