Before building anything, I designed a validation experiment to test demand across two audiences.
Two landing pages, two audiences. One targeting accountants: "Multiply your practice without hiring. Your clients invoice themselves via WhatsApp." One targeting entrepreneurs: "Invoice in seconds without depending on your accountant." Each had differentiated copy, pricing tiers, and a ROI calculator.
30 ads across both audiences tested on Meta Ads with multiple copy variants. ManyChat qualification flows filtered and segmented leads before human contact.
Result: $8,914 MXN generated ~90 leads. Zero converted from paid media. Our only paying customer came through personal referral.
Zero conversions told us to stop advertising and start understanding.
Four ethnographic interviews revealed the core pattern: nobody wanted to replace their accountant. They wanted faster service, or an alternative when the accountant isn't available.
The positioning shift: NUMMA is not "replace your accountant." It's "invoice in minutes when your accountant isn't available." The AI complements an existing human relationship, not replaces it.
Research → Design Decisions
"Finding a good accountant is really hard. Mine is so good she's overworked."
Business Consultant
His invoices take 1–2 weeks because his accountant is saturated. Late invoices mean missed corporate payment cycles.
Finding
Speed is the primary value. Wants faster service, not independence.
One question at a time. Critical info first, technical fields after.
"I search for the code on Google and pick whichever one looks right."
Architect
Already invoices himself through facturador.com. But SAT's 50,000+ product codes are impenetrable. Google is his workaround.
Finding
Struggles with 50,000+ SAT codes. Already using AI workarounds.
High confidence: auto-suggest with explanation. Low: show 2–3 options.
"I need to see it before I send it."
Mezcal Producer
Seasonal business: 4 peak months, zero the rest. Every invoice is a significant percentage of annual income. Mistakes are costly.
Finding
Needs visual preview before sending. Every invoice matters.
Invoice preview is mandatory. AI never stamps without user approval.
"I want a tax advisor who actually explains things to me."
Freelance Photographer
Previous accountants had zero transparency. She invoices for other photographers too because they can't do it themselves.
Finding
Zero transparency from accountants. Wants to understand why.
AI explains WHY it chose each code and regime.
What happens when you put an AI agent as the frontstage of a fiscal compliance service?
CFDIs have zero tolerance for error. A wrong SAT code or RFC makes the invoice legally invalid. But conversational AI is probabilistic: it interprets, suggests, and sometimes gets it wrong. Designing a service where AI handles tax-sensitive data means solving for trust in a domain where mistakes have legal consequences.
The agent collects all invoice data even if it detects an invalid RFC. It only blocks at the moment of stamping.
Why not block immediately? Safer, but breaks the conversational flow and frustrates users who don't have the correct data at hand. Validation runs against SAT in real time via Facturama (same stamping infrastructure), not just format checking.
Tradeoff accepted: The user may reach the end and not be able to stamp. Mitigated with explicit warning upon error detection.
Invoice $12,000 to María López, RFC LOPM850612, marketing consulting
⚠️ That RFC doesn't validate with SAT. I'll keep building the invoice — you can fix it before stamping.
👍
Preview ready:
$12,000 + IVA → $13,920
80101504 · Marketing consulting
❌ Can't stamp yet. RFC still invalid. Send the right one and it goes out instantly.
For ambiguous fields (SAT code, CFDI usage), the AI suggests the most likely option and explains why. If confidence is low, it presents 2–3 options.
Why not a manual catalog selector? The catalog has 50,000+ codes. Impossible in WhatsApp. Research with the Architect showed they already used Google AI to navigate the catalog. The friction was the catalog itself.
The principle: Education as Feature. AI explains WHY it chose each code, not just WHAT it chose. This came from the Freelance Photographer: they wanted to understand the reasoning, not just receive a result.
Tradeoff accepted: AI can suggest incorrectly. Mitigated with explicit confirmation and mandatory visual preview before stamping.
Invoice for $8,500 — blueprints
Pretty sure this is 81101505 · Architectural design. Covers floor plans and technical drawings. Confirm, or tell me more?
It's structural calcs, not floor plans
Different category — and honestly, I can't call this one. You pick:
• 81101513 · Structural engineering
• 81101503 · Engineering consulting
First one
81101513 locked in ✓
100% of interviewees depended on their accountant. None had the CSD (digital certificate) required for stamping. Getting one requires e.firma, a Java app, the CertiSAT portal, and 24–72 hours activation.
Why not automated onboarding? "Send your RFC and start invoicing" was impossible without CSD. Before the AI could do anything, humans needed to guide users through a multi-day government certification process.
Tradeoff accepted: Onboarding takes days, not minutes. No way to eliminate it; can be made less painful with step-by-step guidance.
Hey, I want to start invoicing
Welcome! To issue legal invoices you need a CSD — your tax e-signature. Do you have one?
No idea what that is
Most people don't. It's free, but it takes 2–3 days and a SAT appointment. It's a lot for a first time.
Want me to hand you to Andrea from our team? She'll walk you through it on WhatsApp. Once you're set, I take over.
Yes please
Done. Andrea will message you within the hour. After setup, you'll be invoicing in under 2 minutes 🚀
WhatsApp was the obvious channel. Mexican entrepreneurs spend 8+ hours daily on it. Building here meant zero downloads, zero passwords, zero learning curve.
The service blueprint maps what the user sees (AI frontstage) against what's invisible (human backstage). The CSD discovery made the gap visible: the blueprint has more human touchpoints backstage than AI touchpoints frontstage. Designing an AI service means designing the entire system, not just the conversation.
We assumed the wrong problem. We designed for automation. The interviews revealed the real issue was service quality: accountants are saturated, opaque, and unreachable. Entrepreneurs want visibility into where their invoice is and immediacy, because in Mexico, a late invoice means a missed payment cycle. We weren't solving a technology problem. We were solving a service delivery problem.
AI frontstage requires human backstage. The CSD discovery proved it: before the AI could do anything, humans needed to guide users through multi-day government certification. Designing an AI service means designing the entire system, not just the conversation.
"Speed trumps autonomy" changes how you position AI. Users don't want AI to give them independence. They want AI to make existing human relationships work better. AI products that complement human expertise face lower trust barriers than those trying to replace it.