Why I Built Geminiano: Testing AI Autonomy in Tax Compliance

In early 2024, Anthropic ran an experiment called Project Vend: a small shop in their San Francisco office run entirely by an AI named "Claudius." The results were fascinating but messy—the AI struggled with pricing, got manipulated by employees, and had what the researchers described as "identity crises."

Reading that report, I kept thinking: what if the domain was less fuzzy? Retail requires judgment calls—pricing, negotiation, reading customer intent. But some domains are almost entirely rule-based. Tax compliance, for instance. Once you know the law, there's usually one correct answer.

So I built Geminiano: an AI agent that handles my Italian freelance tax workflows. I'll be testing it throughout 2026 against my actual tax obligations, and I want to document what happens—successes, failures, and everything in between.

Why Tax Compliance?

I chose this domain for a few reasons:

It's verifiable. Unlike chatbot conversations or creative tasks, tax calculations have objectively correct answers. I can compare every output against my accountant's ("commercialista") work.

It's high-stakes but bounded. Errors have real consequences (legal and financial), but the problem space is finite. Italian freelancer tax law is complex, but it's not infinite.

I can be my own test subject. I have a real tax situation, real invoices, real deadlines. No simulations needed.

Aspect	Project Vend (Retail)	Geminiano (Tax Compliance)
Decisions	Pricing, negotiation, customer service	Fixed rules, deterministic calculations
Rules	Heuristics, "common sense"	Codified law
Risk	Financial loss from bad deals	Legal consequences from errors
Success Metric	Profit maximization	Zero errors, perfect compliance

The Core Question

Where's the right line between AI flexibility and strict guardrails?

The agent needs to understand natural language ("create an invoice for Rossi Srl for January's work") and guide me through multi-step workflows. But it absolutely cannot improvise on tax calculations or compliance checks.

I've split this into three categories:

• Read operations: Checking status, retrieving rules — fully autonomous
• Compute operations: Tax estimates, invoice previews — autonomous but logged
• Write operations: Issuing invoices, recording payments — requires my explicit approval

The hypothesis is that this tiered approach lets the AI be helpful without being dangerous. But I don't know if it'll hold up under real use.

The Architecture

Geminiano is a hybrid system. The LLM (Gemini 3 Flash) handles conversation and orchestration. Deterministic tools handle anything involving numbers or rules:

All tax rules live in versioned JSON files I can audit. Every state change gets recorded as an immutable event. The goal is complete traceability—if something goes wrong, I want to know exactly why.

What Could Go Wrong

I'm genuinely uncertain about several things:

Edge cases I haven't anticipated. Tax law has endless special situations. Will the guardrails hold when the agent encounters something novel?

Trust calibration. Will I over-rely on the agent after a few successful runs? Under-rely and defeat the purpose?

The "almost right" problem. A completely wrong answer is obvious. An answer that's 95% right but subtly wrong in a way that matters—that's the dangerous failure mode.

Maintenance burden. Tax rules change. Will keeping the policy files updated be more work than just doing the taxes manually?

What I'm Hoping to Learn

By the end of 2026, I want to answer:

1. Can a hybrid AI system achieve zero errors on real tax workflows over a full year?
2. Does human-in-the-loop approval actually catch mistakes, or does it become rubber-stamping?
3. What's the real time savings vs. doing it manually or with traditional software?
4. Which failure modes show up in practice vs. the ones I anticipated?

The Bigger Bet

If this works, it suggests a model for AI in other regulated domains—healthcare administration, legal compliance, financial planning. Not AI replacing professionals, but AI handling the deterministic grunt work while humans focus on judgment calls.

If it fails, that's also valuable. Understanding why AI struggles with rule-based domains would be genuinely useful to know.

Either way, I'll write about what happens. The project is open-source, and I'll post updates as I learn more.

---

This is the setup. The experiment runs through 2026. Check back for results.