ClasseCerta

1. One-liner

Maps every client SKU to the right cClassTrib with a signed legal citation — and catches the NCM traps the lookup tools miss.

2. Trend signal — why now?

Brazil’s tax reform (EC 132, LC 214/2025) went live on January 1, 2026 with a dual VAT — CBS (federal) and IBS (state/municipal) — replacing PIS, COFINS, IPI, ICMS and ISS. Every electronic invoice now carries two new mandatory fields: CST and cClassTrib (the “tax CPF” of each product/service). For Lucro Real and Lucro Presumido companies these fields are already mandatory in 2026; for Simples Nacional and MEI the obligation starts January 2027. That’s 82,000 accounting offices staring at a deadline.

The hard part isn’t filling a field — it’s the classification trap. The Fisco published a 1,000+ row cClassTrib table, but you cannot mechanically map NCM → cClassTrib. The same NCM 9619.00.00 carries a 100% reduction for sanitary pads (cClassTrib 200013) but only a 60% reduction for baby diapers (200035). Classification depends on the operation (CFOP) and the product’s application/destination, not the product code alone. Pick wrong and the NF-e gets rejected at transmission (Rejeição 1024) — or worse, it’s silently accepted in the 2026 “informative” year and feeds the CGIBS reference base that sets your client’s 2027 rates against them.

The market knows it’s drowning: a CRC-MG survey found 61% of accountants have not yet mapped the reform’s impact on their clients. PLP 108 sets penalties up to 18% of the operation value for accessory-obligation failures. Reclassifying “tens of thousands of registered goods in under 24 working days” was openly called “tecnicamente inviável” by trade press.

Provenance:

• Signal 1 (demand): “61% dos contadores ainda não mapearam impacto em clientes”; reclassification of tens of thousands of SKUs in <24 working days called technically unfeasible — https://crcmg.org.br/noticias/reforma-tributaria-61-dos-contadores-ainda-nao-mapearam-impacto-em-clientes/ — observed 2026-06-09
• Signal 2 (feasibility): Fisco published the cClassTrib/CST/cCredPres reference tables and LC 214/2025 in machine-readable form (downloadable Excel on Portal NF-e); the legal text is public — an LLM can now read catalog + NCM + statute and propose a code with citation. Wasn’t possible before the tables existed (late 2025) — https://blog.tecnospeed.com.br/tabela-cclasstrib/ — observed 2026-06-09
• Signal 3 (economic): 82,000 active accounting offices / 540,392 registered accountants (CFC, Mar 2026); tax-tech vendors (Tecnospeed, Jettax, e-Auditoria, OMIE, Econet) and AI entrants (SOS Reforma) all charging money to solve exactly this — https://www.crcsc.org.br/noticia/view/41194 — observed 2026-06-09 Category: Regulatory arbitrage

3. The opportunity

A nation-scale forced migration with a legal deadline and a non-trivial classification logic is the textbook regulatory-arbitrage window. But this lane already has runners — SOS Reforma (“AI trained on LC 214/2025, hundreds of accountants, millions of products”), Econet’s EcoClass, OMIE IA Fiscal, e-Auditoria, Jettax. So I’m not pretending the field is empty. The gap is in what they optimize for.

The incumbents are mostly batch lookup engines: feed a SPED file or XML, get back a cClassTrib per item, fast. That solves the volume problem but not the defensibility problem. The four most-cited classification errors (per Contábeis) are all things a naive product→code mapper gets wrong: (1) classifying by NCM only; (2) dumping everything into 000001 full-taxation “to facilitate”; (3) ignoring that cClassTrib depends on CFOP/operation; (4) missing conditional benefits tied to product destination. And none of the lookup tools give the accountant a signed, citation-backed record they can stand behind when the 2026 informative data gets weaponized in a 2027 rate dispute.

ClasseCerta’s wedge: be the classification of record, not the fastest guess. Every line gets the proposed cClassTrib + CST, the LC 214/2025 article it rests on, a confidence flag, and an explicit trap alert when the NCM hides a split benefit or the operation context changes the answer. The accountant reviews, accepts/overrides, and the system keeps a versioned, signed ledger of who classified what, when, and on what legal basis. When the Fisco updates the table (it will, repeatedly), ClasseCerta re-checks every prior classification and surfaces what moved. That’s the part a one-shot batch tool structurally doesn’t do.

4. Target market

• Primary customer: Owner/partner of a small Brazilian accounting office (escritório de contabilidade), 1–10 staff, serving 30–300 SMB clients — mostly Simples Nacional and small Lucro Presumido. Concentrated in the Southeast (>50% of all accountants).
• Why they buy: In their words — “meu cliente tem milhares de produtos e eu preciso reclassificar todos antes da obrigatoriedade, e se eu errar a multa é dele e a culpa é minha.” The accountant is the one the SME is told to “ask” (every guide ends with “procure o seu contador”), so the liability funnels to them. They need to do it fast, defensibly, and at a price that works across many small clients.
• Rough TAM reasoning: 82,000 offices. If even 8% (≈6,500) pay ~R$300/mo, that’s R$23M ARR ($4M). I don’t need all of them — a few thousand is a real business.
• Why now for them: Lucro Real/Presumido fields are mandatory today; Simples clients hit the wall Jan 2027. The window to be their tool-of-record is the next 12 months, after which whoever owns the ledger owns the renewal.

5. Product sketch (MVP)

• Upload a client’s product/service registry (SPED, NF-e XML batch, or CSV export from their ERP) — no integration required to start.
• Per-item proposed cClassTrib + CST, with the LC 214/2025 article/anexo cited as the legal basis on every line.
• Trap alerts: flags when an NCM contains split benefits (pads vs. diapers), when classification depends on CFOP/operation, or when a benefit is conditional on destination/application — the four error classes, surfaced not buried.
• Confidence score + review queue: high-confidence items auto-pass; ambiguous ones routed to the accountant with the competing options and their citations side by side.
• Accept / override / annotate, then export a clean classification map back to the ERP plus a signed PDF/CSV audit dossier per client.
• Table-change monitor: when the Fisco republishes the cClassTrib table, re-check every prior classification and show exactly what changed and which clients are affected.
• Multi-client dashboard: the office sees every client’s classification status and what’s still unreviewed before the deadline.

6. AI angle — what’s load-bearing

Remove the AI and there’s no product — just the same 1,000-row Excel table the accountant already can’t get through. The model does three things a lookup table can’t: (1) reads a messy, free-text product description (“refrigerante lata 350ml”, “serviço de manutenção predial”) and reasons to the right cClassTrib given NCM and operation context; (2) reads LC 214/2025 to attach the specific legal basis and detect conditional/split benefits inside a shared NCM; (3) explains its reasoning so the accountant can sign off rather than blindly trust. The audit ledger and trap-detection are the moat; the LLM is what makes per-item legal-basis reasoning economically possible at thousands of SKUs per client.

7. Localization angle

This is the localization play — it only exists in Brazil and only because of LC 214/2025. Portuguese-native, built on the Brazilian NF-e/SPED data model, priced in BRL for small offices (R$199–499/mo lands where a $49 USD tool wouldn’t), and sold into the Brazilian accountant community (contabeis.com.br, CRC events, regional WhatsApp/Telegram groups). No global SaaS will adapt to this in time; the regulation is the wedge.

8. Business model — path to $1M–$5M ARR

• Pricing: Tiered by client count. Starter R$199/mo (up to 25 clients), Office R$399/mo (up to 100), Pro R$799/mo (unlimited + table-change monitoring + signed dossiers). Annual prepay discount to lock the renewal through the 2027 deadline.
• ACV: R$4,800/yr ($850) blended.
• Rough math to $1M ARR (~R$5.7M): ~1,200 offices on the R$399 tier. Out of 82,000, that’s 1.5% penetration. Plausible inside 12 months given the deadline.
• Rough math to $5M ARR (~R$28M): ~6,000 paying offices (~7%) plus upsell to the monitoring/dossier tier and a per-client usage line for high-SKU clients. Requires becoming a default tool, not a deadline panic-buy.
• Expansion path: Land on the one-time reclassification job → keep them for ongoing table-change monitoring and new-product classification as a recurring need → add a “classification health check” upsell for new clients the office onboards.

9. Go-to-market wedge — first 100 customers

• Contábeis.com.br + community placement. This is where Brazilian accountants live; incumbents already run coupon campaigns there (e.g. SOS Reforma’s CONTABEIS10). Sponsored article + a free “classify your 50 messiest SKUs” trial that produces a shareable trap report. Target: 100 trials → 15–20 paid.
• CRC-regional and SESCON channels. Regional accounting councils are running reform-readiness webinars right now (the 61%-unprepared survey came from CRC-MG). Offer a free live “reclassify a real client catalog on stage” demo at 5–6 regional events. Accountants buy tools their peers are visibly using.
• WhatsApp/Telegram accountant groups. There are hundreds of regional contador groups. Seed the trap-report (the diapers-vs-pads example is genuinely scary to them) and a referral code. Accountants forward tools that make them look competent to clients.
• Cold outreach to the unprepared 61%. Scrape CRC public registries + Google Maps “escritório de contabilidade” by city, send a personalized Loom in Portuguese showing their own sample catalog mis-classified by a naive NCM lookup vs. correctly by ClasseCerta. Expect 3–5% reply on a fear-driven, deadline-bound pitch.

10. Build complexity — justification

Medium. Off-the-shelf: the LLM, the public cClassTrib/CST tables, the LC 214/2025 text, standard SPED/NF-e XML parsers (well-documented formats). Custom work: a reliable classification + trap-detection layer with citations (this is the hard, accuracy-sensitive core), the versioned signed-ledger/audit dossier, and the table-change diff monitor. A 2–3 person team with a Brazilian tax-domain advisor ships a credible v1 in 10–14 weeks. The accuracy bar (a wrong classification carries an 18% penalty) is what makes it Medium not Low — you need a human-in-the-loop review UX and conservative confidence flagging, not blind automation.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Selling a compliance aid; accountant remains the signer of record.
Ethical — no harm / dark patterns	✅	Surfaces traps and confidence rather than hiding them; reduces client penalty risk.
Market exists (evidence above)	✅	Mandatory fields live now; paying incumbents; 61% unprepared.
1–5 person team can build this	✅	2–3 people + domain advisor, 10–14 weeks.
Launchable with <$50K / ₹40L	✅	Inference + standard web stack; main cost is the domain advisor’s time.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	17/20	Hair-on-fire: legal deadline, 18% penalty, liability funnels to the accountant. Not 20 because Simples (the bulk of small-office clients) has until Jan 2027 — urgency is real but staggered.
Demand evidence	15	13/15	Multiple hard signals: mandatory fields, paying incumbents, 61%-unprepared survey, active CRC webinars. A skeptic nods.
Build feasibility	15	12/15	Standard stack + public tables/law, but accuracy-critical core and audit-ledger raise the bar above a weekend build.
Distribution clarity	15	12/15	Named channels (contabeis.com.br, CRC events, contador WhatsApp groups), proven by incumbent coupon campaigns. Conversion uncertain in a crowded field.
Revenue mechanics	15	12/15	BRL pricing benchmarked against existing tools; 1.5% penetration to $1M is realistic. Risk: deadline panic-buy that doesn’t renew.
Time to first revenue	10	8/10	Fear + deadline = fast close; free trial → paid in weeks. Not 9–10 because it needs a credible accuracy demo first.
Defensibility	10	3/10	Honest low. Several funded/AI incumbents already classify cClassTrib; the moat is the audit-ledger + monitoring + workflow lock-in, which is real but copyable. A clone wins on price unless you own the renewal.
Total	100	77/100

13. Qualitative modifiers

Founder-fit tags

domain-expertise-required · technical-heavy

You need someone who genuinely understands Brazilian fiscal classification (or a tight advisor relationship) — the accuracy bar is unforgiving and the trap logic is the product. Technical to build the parsing + LLM + ledger reliably.

Key assumptions to validate (3–5)

1. Assumption: Small offices will trust an AI classification enough to sign off, given the 18% penalty. How to test: Put real catalogs through ClasseCerta and the incumbents in front of 20 accountants; measure whether they accept the citation-backed output and whether trap-detection changes their decision.
2. Assumption: The audit-ledger + table-change monitoring is worth a recurring subscription, not just a one-time job. How to test: Offer monthly vs. one-time pricing to the first 30 trials; measure how many choose recurring.
3. Assumption: ClasseCerta’s classification accuracy beats the naive-NCM-lookup incumbents on the trap cases. How to test: Build a benchmark set of known-hard NCMs (split benefits, CFOP-dependent) and score head-to-head.
4. Assumption: Channels (contabeis, CRC, WhatsApp groups) convert at ≥3% on a fear-driven pitch. How to test: Run one sponsored trial campaign and one cold-Loom batch of 200; measure reply→trial→paid.

Risk flags

1. Competitive / timing risk: This is a hot, crowded lane with a 12-month deadline window. If you’re not in market within ~3 months, incumbents own the ledger and you’re fighting for renewals you don’t hold.
2. Regulatory risk (the good and bad kind): The Fisco will keep republishing tables and decrees (the suspension calendar literally depends on an unreleased decree). Constant table churn is a feature for the monitoring product — but it also means classification logic you ship can be invalidated overnight; you must update fast or your “record” becomes a liability.
3. Liability / trust risk: If an accepted classification later proves wrong and triggers a penalty, blame can splash back at the tool even though the accountant signs. Positioning (decision-support, accountant-as-signer) and conservative confidence flagging are load-bearing, not cosmetic.
4. Post-deadline churn: Once 2027 obligations are met, the one-time reclassification urgency fades. The business only compounds if monitoring + new-product classification become a genuine recurring need.

14. Structured verdict

Score:                  77/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder + Brazilian fiscal-classification domain advisor (2–3 person team)
Time to revenue:        6–10 weeks (deadline-driven; fast close on a credible accuracy demo)
Capital to launch:      R$60–120K ($11–22K) — mostly domain-advisor time + inference
Top 3 assumptions to validate first:
  1. Accountants trust citation-backed AI classification enough to sign — 20-accountant head-to-head test
  2. Recurring (monitoring) beats one-time — monthly vs. one-time offer to first 30 trials
  3. Accuracy beats naive-NCM incumbents on trap cases — benchmark set of hard NCMs, head-to-head
Kill criteria:
  - Abandon if <20% of 30 trial offices accept the AI output without re-doing it manually
  - Abandon if head-to-head accuracy on trap NCMs is not clearly better than a free incumbent
  - Abandon if <10% of trials choose recurring over one-time (no business after the deadline)

15. Next step — 1-week validation sprint

• Day 1–2: Build a 200-SKU benchmark of genuinely hard cases — split-benefit NCMs, CFOP-dependent items, conditional-destination products. Hand-label the correct cClassTrib with a tax advisor. This is the falsifiable yardstick.
• Day 3–4: Run the benchmark through a quick ClasseCerta prototype (LLM + tables + LC 214/2025) and through SOS Reforma / a free NCM-lookup tool. Record accuracy on the trap subset specifically.
• Day 5: Show the prototype + benchmark to 8–10 accountants from a contador WhatsApp group. Go decision: prototype beats incumbents on trap accuracy by a clear margin AND ≥4 of 10 accountants say they’d pay monthly for the citation-backed ledger + monitoring. Falsifiable: if it doesn’t out-classify a free tool on the hard cases, the wedge is fake and this is a PASS.