BUSINESS SPEED
YOUR ASC 606
MEMO LOOKS
PERFECT — that's the problem
On April 14, 2026, Veritone Inc. filed an 8-K restating Q3 2025. The error, in the SEC filing's own language: "recognizing revenue for a transaction prior to meeting step 1 under ASC 606." They booked revenue before establishing that a contract existed. New material weakness in revenue recognition for non-routine transactions. Revenue restated down $3.3 million over nine months. Disclosure controls and ICFR: not effective.
Veritone is not unusual. The talent shortage means fewer experienced staff reviewing more complex transactions. AI is entering this environment as an accelerant — making good processes faster and bad processes worse.
If your department is using AI for ASC 606 work — and it probably already is — the workflow matters more than the tool.
---
A structured workflow for AI-assisted revenue recognition memos
A good workflow makes errors locatable rather than cascading. A bad workflow lets a wrong conclusion in Step 2 silently propagate through Steps 3, 4, and 5. Or, as Veritone demonstrates, a wrong conclusion in Step 1 cascades through everything that follows.
Step 1 — Contract extraction (AI-appropriate). Use AI to surface key terms: consideration amounts, performance milestones, variable consideration provisions, termination rights, licensing terms, renewal options. Verify extracted terms against the executed contract before proceeding. Veritone's failure was at this step — a transaction was treated as a contract before the contract criteria were met. AI's productivity at term extraction does not answer the underlying question of whether a contract exists.
Step 2 — Issue identification (AI-appropriate with verification). Use AI to flag which ASC 606 steps are likely to involve judgment — distinct performance obligations, constraint on variable consideration, allocation methodology. Verify that the AI has not omitted issues present in the contract.
Step 3 — Analysis by step (human-led, AI-assisted). Address each of the five ASC 606 steps as a separate, discrete exchange with the AI. Do not let it work through all five steps in one pass. Write the controller's own conclusion — not the AI's — for each step in the workpaper.
Step 4 — Citation verification (human-performed). Open the Codification for every ASC paragraph reference in the memo. Confirm that the paragraph exists, that it says what the memo claims it says, and that it supports the conclusion. Five to ten minutes per memo. The single highest-value quality control in AI-assisted work.
Step 5 — Full-document consistency review (human-performed). Read the completed memo end-to-end. Check that the performance obligation count in Step 2 matches the allocation in Step 4 and the recognition pattern in Step 5. Check that constraints acknowledged early are maintained throughout.
---
Why this workflow exists — the failure modes it prevents
EY's 2025 Responsible AI Pulse Survey quantified the knowledge dimension directly. When C-suite executives were asked to identify appropriate controls against five common AI-related risks, only 12% answered correctly. The four failure modes below are the operational surface of that gap — they describe where the 88% gets the controls architecture wrong, and where the memo you sign downstream inherits the error. Each failure mode maps to a specific step in the workflow on the prior page.
The drift problem. ASC 606 is a five-step sequential process. Each step builds on the conclusions of the previous step. Over a multi-turn exchange, AI models progressively lose context established in earlier turns. In complex scenarios, the performance obligation count established in Step 2 can silently change by Step 4. The allocation basis shifts. The final memo is internally inconsistent with its own earlier analysis — but each individual section reads correctly in isolation. Step 5 catches this.
Severity normalization. AI produces a memo in the same measured, professional register regardless of whether its conclusions are correct. A material error in the allocation of transaction price receives the same tonal treatment as a straightforward observation about the contract term. The most dangerous outputs are the ones that sound the most reasonable. Step 3 — writing the controller's own conclusion rather than adopting the AI's — forces the controller to reason through the position rather than accept a confident-sounding assertion.
Citation fabrication. AI produces specific ASC paragraph references with identical confidence regardless of whether those citations support the conclusion or exist in the form described. In testing against the actual Codification, AI-generated paragraph descriptions contained errors — the paragraph numbers were real, but the descriptions of what they covered were not. (The specific documented example is in Part 2 of this series.) Step 4 catches this.
The affirmation loop. A controller who drafts a preliminary conclusion and then asks the AI "does this treatment look right?" will almost always receive confirmation. Ask the same model "what is wrong with this treatment?" and it will identify problems with the position it just endorsed. The controller has not received a second opinion. They have received a mirror. Step 3 prevents this by requiring the controller to author the conclusion, not delegate it.
---
AI-assisted memo review checklist
Before signing an AI-assisted memo, verify:
| # | Check | What You Are Looking For |
|---|---|---|
| 1 | Every ASC citation opened at the source | Paragraph exists, says what is claimed, supports the conclusion |
| 2 | Performance obligation count is consistent throughout | Step 2 identification matches Step 4 allocation matches Step 5 recognition |
| 3 | Variable consideration constraint is applied consistently | Early acknowledgment of constraint is not silently dropped in later sections |
| 4 | Principles-based judgment sections are authored by the controller | AI may have provided a starting point; the conclusion must be the signer's |
| 5 | The memo does not use identical reasoning language for different fact patterns | Indicates template-matching rather than specific analysis |
| 6 | The controller authored the conclusion independently | AI was not used to validate the controller's position; a confirmation from AI is not a second opinion |
| 7 | Contract terms referenced match the executed agreement | AI may have used standard terms that differ from the actual contract |
| 8 | All AI use is documented in the workpaper | Who queried, what was queried, what was used, who verified |
---
What this means for your close cycle
The controller who uses AI to draft a revenue recognition memo is not taking a shortcut. They are making a decision about their audit trail, their documentation quality, and their personal liability on the management representation letter.
The workflow and checklist above redistribute the time spent on a memo — from drafting to verifying — rather than reducing it. The total hours are roughly the same. The defensibility delta is substantial. A Veritone-style restatement is the cost of skipping verification; five minutes per citation check is the cost of avoiding it.
---