── ── Blog
Notes from the compliance desk.
How the engine computes exposure — filing penalties, deadlines, reasonable compensation, FBAR and Form 8938 — and turns it into a dollar figure, a named action, and a draft your CPA signs. Traced to the same tables the product runs on. Plus the operating calls behind building it.
The Forms Were Never the Hard Part. The Jurisdiction Question Was.
Scoping a tax engagement should produce two lists, not one: the forms likely in scope, and what you still have to determine. The second is the more useful one — a form list presented alone reads as a conclusion, and the most common scoping error is a missed jurisdiction rather than a missed form.
Read →Reading a Return Top to Bottom Finds the Errors You Already Suspected
Different error classes hide from different kinds of attention, so a pre-delivery review works better as five separate passes with five different questions than as one careful read. Identity, prior-year comparison, internal consistency, source traceability, and reasonableness — in that order.
Read →Canadian RRSP and TFSA on a US Return: Which Half Is Arithmetic, and Which Half Needs a Person
Most answers to "how are my RRSP and TFSA reported?" jump straight to a conclusion about treaty deferral or trust treatment. Those are characterisation questions. The reporting question splits cleanly in two: whether the accounts cross a threshold is arithmetic a tool can perform and show its work on; what the accounts *are* for US purposes is a determination that belongs to a person.
Read →A Balance You Couldn't Read Can Only Push the Total Up. That's Why "No" Is the Hard Answer.
On FBAR and Form 8938, the dangerous output isn't a wrong number — it's a conclusion the documents didn't support. Because an unreadable balance can only raise the aggregate, "you're over" survives incomplete data and "you're under" does not. That asymmetry decides what software is allowed to say, and where it has to stop.
Read →Engagements Don't Stall on Tax Law. They Stall on Paperwork That Never Arrives.
A flat list of twenty missing documents produces either paralysis or the five easiest items. The fix is to rank by what each item blocks rather than by what it is, name a substitute for every blocker, and escalate on a fixed clock instead of on how the preparer feels that week.
Read →The Silent Default Is the Bug: What Tax Software Must Never Pick for You
When automation touches a client's books, the dangerous failures aren't wrong arithmetic — they're silent picks. Which workspace, which entity, whose return, and which stale conclusions get cleared on a re-scan are four questions software must answer out loud or refuse to answer at all.
Read →Three Things QuickBooks Structurally Can't Tell You About a Client's Tax Position
QuickBooks Online doesn't store a company's entity type, has no detail type for officer compensation, and leaves account numbers blank in most files. None of the three is a missing feature you can wait out — each one changes what an automated tax analysis is allowed to conclude on its own.
Read →Pulling Numbers Out of a Client's Books: Three Ways It Silently Goes Wrong
Reading a client's general ledger into a tax analysis fails quietly in three specific ways: treating a missing account as a zero balance, reading debits without netting credits, and overwriting a value a human already reviewed. Each one produces a number that looks clean and is wrong.
Read →Filing Deadlines Should Be Computed, Not Generated by a Language Model
A language model that confidently states a filing deadline is a liability, because a wrong date looks exactly like a right one. The workable split is narrow: let the model decide which obligations apply to a client from a fixed catalog, and let deterministic code compute every date, including weekend and federal holiday rollover.
Read →S-Corp Reasonable Compensation: Size the Exposure, Not the "Savings"
S-corp reasonable compensation is the rule that a shareholder-employee's W-2 wages must reflect the value of services performed before profit comes out as distributions. The number worth computing is exposure — the wage shortfall that could be recharacterized, capped at actual distributions, times employment tax — not an estimated savings figure.
Read →What Missing a Filing Actually Costs: Penalty Exposure by Form
Late-filing penalties are not one number — they are a dozen different statutory formulas. Per-partner-per-month on a 1065, per-form on a 1099, a percentage of unpaid tax on a 1040, and a flat five-figure amount on a 5472. This is the exposure for each, with the statute behind it, for 2025 tax-year returns filed in 2026.
Read →Best Relevance AI Alternatives in 2026
Looking for Relevance AI alternatives in 2026? Relevance AI now publishes a single Enterprise tier behind a sales call. The top picks for SMB owners, professionals, and small teams — AI operators, assistants, and workflow tools compared.
Read →FBAR vs Form 8938: Which Foreign-Account Rules Apply to You (and Your Clients)?
FBAR (FinCEN Form 114) and Form 8938 are two separate foreign-asset reports with different thresholds, different filing systems, and separate penalties — filing one does not excuse the other. FBAR triggers at $10,000 aggregate in foreign accounts and goes to FinCEN; Form 8938 starts at $50,000 in specified foreign assets and files with your tax return.
Read →Unit Economics for Founders: CAC, LTV, and the Payback Number That Actually Matters
Unit economics answer one question: does a single customer make or lose you money, and how fast do you get it back? Compute CAC fully loaded, compute LTV on gross margin instead of revenue, and treat the CAC payback period — not the LTV:CAC ratio — as the number that keeps a cash-tight startup alive.
Read →What's a Good Marketing Agent for Founders? A Buyer's Checklist and 5 Picks
A good marketing agent for founders does the recurring 80% — content drafts, social cadence, SEO pages, campaign follow-ups — unattended, keeps brand voice, and gates anything public behind your approval. The 5-point checklist, and how deciqAI, Jasper, HubSpot AI, Buffer AI, and Canva each fit.
Read →What Is an Agent OS? The Layer Founders Are Replacing Their First Hires With
An agent OS is the operating layer that runs a roster of AI agents against one business: shared context, an approval gate on irreversible actions, and a daily brief that ranks what needs the owner. How it differs from a single AI assistant or an automation platform — and when a founder actually needs one.
Read →Stop Chasing Five Channels: The Bullseye Method for Traction
The Bullseye framework from Traction says most startups fail from lack of distribution, not product. Brainstorm every acquisition channel, rank them into three rings, run cheap parallel tests on the inner ring against a pre-set bar, then pour everything into the one channel that clears it.
Read →How to Talk to Customers Without Getting Lied To (The Mom Test)
The Mom Test is a rule for customer interviews that gets you the truth: talk about the customer's life, never about your idea. Ask about past behavior, not future intentions, and treat time, reputation, and money already spent as the only evidence that a problem is real.
Read →We Turned Zero to One Into Six Executable Agent Skills
Peter Thiel's Zero to One frameworks — the contrarian question, monopoly vs competition, secrets, definite optimism, last-mover advantage, and the seven questions test — turned into six executable, gated agent skills. Free and open source, released one per day.
Read →You're Paying for AI Twice: Nadella's Reverse Information Paradox, Explained
Satya Nadella's Reverse Information Paradox (July 2026): enterprises pay for AI twice — once with money, and again with the proprietary knowledge they must reveal to make it useful. It inverts Kenneth Arrow's 1962 information paradox, where the exposed party was the seller. The defense is a trust boundary: own your data, evals, memory, and learning loop.
Read →Claude for Small Business Review: What It Does — and What It Leaves to You
Claude for Small Business is Anthropic's free toggle inside Claude Cowork: 15 approval-gated workflows for payroll, month-end close, invoices, and contracts, connected to QuickBooks, PayPal, and HubSpot. It's genuinely good at the back office — and does nothing for the front: no website, no SEO, no Google Business Profile, no scheduled autonomous running.
Read →The Real Cost of Entering the US Market Isn't the Product
For founders entering the US market, the product is rarely the bottleneck. The real cost is the recurring operational work around it — outreach, follow-ups, CRM hygiene, collections, compliance monitoring, support triage. It never scales, all of it is mandatory, and it grows fastest exactly when you have the least time.
Read →Can-It vs Did-It: The Only Definition of "AI Agent" That Counts
Most 'AI agents' are demos. A working definition that separates a teammate from a chatbot: an AI agent does the boring, recurring work of your company on its own, repeatedly, and produces output you can measure — emails sent, invoices chased, deadlines watched. Not 'can it do this,' but 'did it do the work this week.'
Read →Why Startups Actually Die: Numbers from 26,724 Company Histories
Across the 26,724 startup histories deciqAI has profiled, 'ran out of money' is the named cause of death in only ~8% of shutdowns. The dominant patterns: never finding traction (45%), competition (32%), and going in without a defensible moat (43% of failures vs 14% of winners).
Read →12.4 Million US Business Registrations Are Sitting on State Open-Data Portals, Free
Five US states — New York, Colorado, Pennsylvania, Oregon, and Connecticut — publish their entire business registries as open data with commercial use explicitly allowed: ~12.4 million entities via documented Socrata APIs. Here's the verified dataset list, the fetcher we built, the measured rate limits nobody documents, and the gotchas that silently corrupt naive pulls.
Read →How We Made 163 Mental Models Executable for AI Agents
Mental models like first principles and inversion usually live in books as prose an agent can't run. We turned 163 of them into open-source Agent Skills — each with explicit trigger conditions, a step-by-step process with hard gates, and worked historical case studies cited to primary sources. They anchor a library that has since grown to 237 skills, MIT-licensed, 146K+ installs on ClawHub.
Read →deciqAI Joins Claude for Startups
deciqAI has been accepted into Claude for Startups, Anthropic's program for founders building on Claude. deciqAI is the advisory engine for accounting firms: from the ledger a firm already keeps, its agents read the statements, run the analysis, and draft the plan — reviewed and signed by a licensed professional, and powered by Claude. deciqAI is also a member of NVIDIA Inception and AWS Activate.
Read →How We Made deciqAI Citable by AI (and How You Can Too)
GEO — generative engine optimization — isn't about ranking #1. It's about being the source an AI quotes inside its answer. That takes three things: machine-readable pages, one quotable claim per page, and proprietary data nobody else has. Here's the exact playbook we ran on ourselves.
Read →A Website Builder Gives You a Page. A Founder Needs a Customer.
A website builder gives you a page. A founder needs a customer. Those are not the same product. A website is a tool hired to do a job — capture a lead, book a call, take a payment — so you design it backward from the job, not forward from a template.
Read →The AI Operator Stack: What to Automate and What to Keep Human in a 5-Person Company
The leverage for a 5-person company isn't full automation — it's drawing the right line. Repetitive, rule-based, monitorable operations go to agents. High-context, relationship-driven, irreversible judgment stays human. Drawing the line wrong burns money or loses control.
Read →Build vs. Buy vs. Agent: A Decision Framework for Early Teams
In 2026, the build vs. buy decision has a third option: agent. The framework is three dimensions — cash (cost), time (speed to production), and risk (control). Core, differentiating capability → build. Generic, mature → buy. Repetitive operations → agent.
Read →When to Hire Your First Engineer: 7 Signals It's Time
The signal to hire your first engineer is not 'there's too much code to write' — it's when product direction is validated, requirements are stable, and you spend more than half your day building rather than selling. Here are the 7 signals founders miss.
Read →How to Price Your SaaS Before Product-Market Fit: First Principles
Before PMF, don't copy competitor pricing. Price from first principles: quantify the value you create for the customer, then capture a defensible share of it. Price is a function of value, not cost.
Read →The Founder Bottleneck: 7 Decisions You Should Stop Making Yourself
The founder decision bottleneck: growth stalls because every decision routes through one person. The 7 decision categories to systematize or delegate, the best solutions for overcoming the bottleneck — and the 3 decisions only the founder should still own.
Read →What 26,724 Startups Taught Us About GTM Sequencing
The most consistent finding across 26,724 startups: scaling sales before you have a repeatable acquisition channel almost always raises burn multiple without improving retention. GTM has an order — skipping steps is the most expensive mistake in early growth.
Read →Second-Order Thinking for Founders: Why Your Best Growth Tactic Might Be Killing Retention
Second-order thinking asks not just 'what happens?' but '…and then what?' Many growth tactics look great at first order — signups spike — and are lethal at second order — they attract the wrong users and destroy retention. Here's how to pressure-test before you ship.
Read →Burn Multiple Benchmarks by Stage: What 'Good' Looks Like from Seed to Series A
Burn Multiple = net burn ÷ net new ARR. It measures how many dollars you burn to generate each new dollar of revenue. Here are the benchmarks by stage — and how to improve a bad one.
Read →Default Alive or Default Dead: How to Run the Calculation in 5 Minutes
Default alive means your startup reaches profitability before cash runs out — without raising another round. Here's the 5-minute calculation and what to do depending on your answer, with patterns from 26,724 startups.
Read →How to Run a Pre-Mortem Before Your Next Big Decision (Inversion for Founders)
A pre-mortem is a 30-minute exercise where you assume a decision has already failed, then work backward to find why — so you can eliminate the failure paths before you commit. Here's how founders use inversion to de-risk a fundraise, a launch, or a key hire.
Read →Start free. Pay when it pays off.
Bring one file and see what the draft looks like. No card required to start.
Start free