Summary
We price a business return at $99 to the firm, with unlimited prepare-and-examine iterations. The firm keeps the client, the signature and its usual fee; our software and team do the preparation. This paper explains the cost structure behind that number, why it falls as the platform grows, and what per-return pricing means for a CPA firm's capacity, margin and income.
What firms pay and charge today
- Software. Professional tax software charges firms roughly $47 to $148 per business return depending on product and volume, before review time (Intuit, published ProConnect Tax and Lacerte pay-per-return pricing, November 2026).
- Fees. Small firms charge clients an average of $750 to $1,250 for a Form 1120 (National Society of Accountants, 2024 Income and Fees survey, published in the 2024 Profit and Practice Report Summary). Founder-focused services sell an 1120 package directly to startups for roughly $850 to $2,450 a year.
- Time. The binding constraint is not software cost. It is preparer and reviewer hours, and the pipeline of new CPAs has been shrinking: first-time CPA exam candidates fell from 42,626 in 2023 to 28,082 in 2024 (Journal of Accountancy, October 2025).
A $99 return is priced against the time it replaces, not the software it sits beside.
Where the cost of a return goes
The cost of an agent-prepared return is dominated by model calls in the examination loop. Within a round, the shape is roughly:
| Component | Share of round cost | Lever |
|---|---|---|
| Output tokens | about 70% | Scope each fix to the schedule it touches; never regenerate the whole return |
| Fresh input | about 20% | Cache the stable prefix |
| Cached input | about 10% | Already a small fraction of the fresh price |
Two engineering decisions follow.
Stable prefix first. Every prompt is ordered with the material that does not change — chart of accounts, form instructions, authorities, the ledger — first, and the round's delta last. The cached portion is contiguous across every round of a return and largely shared across returns of the same entity type.
Code before models. Anything a deterministic check can catch is caught in pre-flight, which costs nothing in model time.
Two models, routed by role
Not every judgment needs the most capable model. NGTAX.AI routes each role to a pinned model version:
| Model | Work |
|---|---|
| Claude Sonnet 5.5 | Options on findings with one clear answer that touch no election, position or book-to-tax item |
| Claude Opus 5.5 | Classification and preparing the return; tax treatments and elections; independent review and the isolated examiner; resolving elections, positions and book-to-tax differences; offline rule harvesting |
The most capable model is reserved for the judgments that move the return and for the examination itself. Offline work, such as harvesting new checks from resolved findings, runs through batch processing at a lower price than real-time calls.
Why the cost falls with volume
The rule harvest (paper 4) converts recurring findings into deterministic checks. As the check library grows, fewer rounds are needed per return. Our planning model — an estimate until there is production data — projects model cost per return falling from under $5 for the first returns to about $1 at scale, as rounds fall from about seven toward a floor of one or two.
Every evaluation and production run is also bounded by a hard budget: a per-run spend cap and a call cap that stop the run rather than overspend, and a refusal to accept a response from any model other than the one requested.
What it does for a CPA's income
We are built to partner with US CPAs, not to replace them or compete for their clients.
- The fee stays with the firm. The firm engages the client and bills its usual fee. We bill the firm $99 per return. The client relationship and the signature never move.
- More returns per season. Season capacity is bounded by preparer and reviewer hours, not by demand. When preparation and tie-outs come off the desk, the same people can sign more returns before the deadline.
- Higher margin per return. A fee set against hours of preparation, paired with a fixed per-return cost, leaves more of each fee with the firm. A return that needs twelve rounds costs the same as one that needs two, so the firm never chooses between another review pass and its margin.
- Work the firm used to turn away. The corporate return with messy books, the foreign-owned company with a Form 5472, the client who arrives in March. These become returns the firm can take, at its own fee.
- Help when the season peaks. In managed mode, US-licensed CPAs and Enrolled Agents at NGTAX.AI prepare the return and a reviewer checks it before it reaches the signing CPA. Managed preparation is not on the published price list; firms that want it talk to us at support@ngtax.ai.
- Better files. Every return arrives with its findings, resolutions, positions and provenance — the file a reviewer or peer reviewer wants to see.
NGTAX.AI is a US company. Client data is processed and stored in the United States.
What we have not yet established
We have not yet measured, on live client returns, how many rounds a typical return takes, how accurate each agent is, or how much reviewer time is saved. Our evaluation harness scores agents blind against hand-worked returns with planted errors, and we will publish those results with their numerators and denominators. Until then, the figures here are a cost model, not a measured outcome, and we make no claim about how many additional returns, or how much additional income, any particular firm will see.