I Built a Full ASC 606 Reporting Stack With AI. Here's What That Really Took. (Part 1)

Revenue Recognition · SaaS Finance · AI in Accounting

Part 1 of 2, Quick read (~5 minutes) · Full methodology in Part 2

Full code and output files available on GitHub here and Part 2 link here

The Setup

I built a complete ASC 606 revenue recognition stack for a fictional SaaS company, 50 customers, 61 contracts, 948 monthly transactions, covering every edge case I could engineer simultaneously: three products with different recognition patterns, five modification types, four termination treatments, bundle contracts requiring SSP allocation, implementation fees, usage-based overages, free trials, renewals, and product switches.

That combination matters. Each feature is manageable in isolation. All of them together, running across 24 months at the contract level, is where the complexity lives, and where most finance teams quietly give up on rigour.

Then I produced the full output: a validated recognition engine with 12 logic checks, 8 SaaS performance metrics, a 10-tab Excel model, an interactive dashboard, and a 10-page disclosure package formatted as a 10-K financial statement footnote. The disclosure alone, Notes 1 through 6 covering revenue policy, disaggregation across four dimensions, deferred revenue, contract assets, remaining performance obligations, and capitalized commissions, is the deliverable a Big 4 engagement would produce after 8–12 weeks with a team of four.

I used AI extensively throughout. And I want to be honest about what that actually meant, not the polished version, but the real version

What AI Is Good At — and What It Doesn't Know

AI has been trained on vast amounts of accounting literature, SEC filings, technical standards, and SaaS financial models. When you ask it to write an ASC 606 recognition engine or build a deferred revenue roll-forward, it draws on patterns from thousands of real examples. That's genuinely powerful.

But here's what AI doesn't know: your situation.

It doesn't know whether your company's auditors prefer a conservative or aggressive interpretation of the modification guidance. It doesn't know whether a 6% annual churn floor is reasonable for your customer segment or whether it should be 3% or 12%. It doesn't know if your company prices bundle contracts at a premium or a discount to standalone. It doesn't know your materiality threshold. It doesn't know your renewal history. It doesn't know how your sales team actually structures deals.

This gap, between what AI knows from global training data and what's correct for your specific case, is where most of the work actually lived.

The Part Nobody Talks About: Validation

Here's what the timeline actually looked like on this project.

AI wrote code quickly. The recognition engine, the metrics layer, the Excel model, the dashboard - fast. But then the real work started.

The first LTV:CAC output showed ratios of 200–255×. AI didn't flag this as a problem. It was mathematically correct given the inputs. But it was wrong for this dataset, because it was using a spot monthly churn rate, and in a 50-customer model, most months have zero churn events, giving a near-zero rate and an astronomical LTV. Fixing this required understanding why the number was wrong, deciding that trailing 12-month average churn with a floor was the right methodology, and rebuilding the calculation. That diagnostic process took longer than the original build.

The deferred revenue roll-forward in the first disclosure draft showed "total billings" versus "total revenue recognized" as the two movement columns. Arithmetically, these don't tie, because monthly billing is recognized immediately and never touches the deferred balance. A real 10-K uses additions to deferred versus releases from deferred. AI didn't catch this. I did, after checking every quarter's arithmetic and tracing why none of them tied.

The commission roll-forward was missing a write-off line. When a contract terminates early, ASC 340-40-35-1 requires immediate expensing of the unamortized commission balance. The engine implemented this correctly, but the disclosure didn't show it as a separate movement, which meant the table didn't tie. The tool built the right mechanism but didn't know the disclosure needed to reflect it.

When the churn reversal fix was applied to the engine, the revenue total changed from $2,328,108 to $2,317,039. Three of the four Note 2 disaggregation tables updated automatically. The geography total row didn't, because it was hardcoded. A one-number inconsistency across a 10-page disclosure, invisible until you add up the components and compare to the header.

Across the full project, seven bugs were found and fixed. None were flagged automatically. Every one required knowing what the correct answer should look like before you could see what was wrong.

This is the part that took the most time, and the part AI cannot do for you. AI runs the calculation. It doesn't know if the result is right for your organization, your auditors, your industry, your contract structure.

The Judgements That Still Required a Professional

Beyond validation, there were decisions that required accounting expertise, and that AI would have gotten wrong without direction:

Modification treatment. When a customer upgrades mid-contract, you choose between prospective treatment (new separate performance obligation for the increment) and cumulative catch-up (retroactive adjustment to reflect the new price from day one). Most SaaS companies use prospective. That's a policy decision with a material impact on revenue timing, approximately $300K over 24 months in this model. I made it. AI executed it.

Material right assessment. Renewal options must be assessed under ASC 606-10-55-42 do they give the customer a price advantage they wouldn't otherwise receive? If yes, they're a separate performance obligation. In this model, renewals are at market rates, so no material right. Getting this wrong would have created phantom deferred revenue across every renewal contract.

RPO exclusions. Month-to-month contracts and DataPipeline variable overages, both excluded from Remaining Performance Obligations under different practical expedients. Including month-to-month customers would inflate RPO with obligations that have no enforceable economic substance, the customer can cancel with no penalty. Including variable overages would require estimating future usage, producing a number that tells a reader nothing reliable. AI knows the standard exists. Deciding which exclusions apply, and why, is the judgement.

Commission rate and amortization life. The model uses 10–15% commission on full contract value, amortized over expected customer life by segment (18/24/36 months). These numbers don't come from ASC 340-40, they come from the business. AI need to be taught about what your sales team gets paid or how long your customers actually stay.

What This Means

AI compressed what would have been months of execution into days. That's real and significant. A Big 4 engagement on a project of this scope, dedicated manager, two seniors, staff associate, 8–12 weeks, costs more than most Series A SaaS companies can justify for an internal finance project.

One accounting professional with the right tools can now produce comparable output in a fraction of the time.

But the output is only as good as the professional governing it. AI that gets the modification treatment wrong will get it wrong 948 times, consistently, at scale.

The skill that matters now isn't knowing how to write the code. It's knowing whether the output is right.

Part 2 covers the full technical methodology: project architecture, the recognition engine, all 12 validation checks, every bug found and fixed, all key accounting decisions, and if you're thinking about applying this to your own organization, an honest guide to what it actually takes, where it is most likely to go wrong, and what accounting knowledge you need before the tools become useful.

Full code and output files available on GitHub. [Link]

Tools used: Python (pandas, reportlab, openpyxl), Claude (Anthropic), Excel