An AI-Native Document Processing Build for Accounting

How the regulatory frame shapes the architecture

Accounting regulatory expectations on AI have hardened over the last twenty-four months. Supervisors who would once accept "we use AI in this workflow" as a sufficient disclosure now ask for the model card, the validation evidence, the override path, and the customer-disclosure language. Vendors who built for the looser bar are scrambling. We built for the harder bar from the start, because the engagement model we sell accounting teams is one we can defend in front of any reasonable supervisor.

For document processing, that defense rests on five artefacts the Build phase produces. The model card documents the deployed system: what it does, what it does not do, the training data lineage, the evaluation methodology, the known failure modes. The validation evidence is the labelled test set with its full provenance, the periodic eval reports, and the calibration curves. The override path is documented in the operator playbook and instrumented in the reviewer UI. The customer-disclosure language is drafted with your legal team during Build and tested with sample interactions before launch. The control map ties each control to a named owner and a measurable SLA.

The artefacts live in version control alongside the code, not in a shared drive. They are reviewed quarterly during Run and updated when the system changes. When a supervisor asks for them, the export is a single command. This is not theatre — it is the operating posture that lets your team say "yes, we use AI in this workflow, and here is the evidence we run it responsibly", with the evidence available in the time it takes to brew coffee.

Accounting sits inside a regulatory perimeter that an AI-native workflow has to inhabit, not bolt onto afterwards. For document processing, the perimeter includes: data residency rules for the source corpus, model-output traceability for any decision affecting a customer, replayability for the regulator's audit window, and named human accountability for every category of decision. We capture each of those requirements during Discovery, before any code is written, and the Build statement of work names which control implements which requirement. The output is an architecture where compliance is not a phase — it is a layer that lives in the same dashboard as the operating metrics.

The specific controls we ship for accounting engagements track the published expectations of the relevant supervisory bodies. The model registry records every prompt and model version that touched a decision, with an immutable hash. The retrieval index documents source provenance, freshness, and approval status per document. The reviewer queue captures the human owner, the timestamp, and the rationale for every escalation. The attestation pack — exportable on demand — bundles the above for any 30/60/90-day window the regulator chooses. This is the same shape that internal audit teams in accounting have been refining for a decade; we replicate it inside the AI-native operating layer instead of duplicating it in a separate evidence binder.

Where we depart from a traditional risk-and-controls program is in cadence. The classic posture treats compliance as an annual or quarterly attestation; the AI-native posture treats it as a weekly heartbeat. Every Monday during Run we sample low-confidence decisions, calibrate thresholds, and produce a drift report. Every quarter we run a red-team exercise on the most consequential flows. The compliance officer joining one of those Monday reviews sees the same dashboard the operators see, with attestation-ready evidence one click away. That continuity is what auditors recognize as a controlled environment, and it is what lets accounting leadership defend the workflow when the next supervisory examination arrives.

Third-party risk management for AI components in accounting is a growing concern that most workflows handle poorly. document processing engagements typically depend on a model provider, a retrieval store, a vector database, sometimes a fine-tuning service. Each is a vendor in your risk register. We map them all during Build, document substitution paths for each, and demonstrate substitutability in the eval harness — so when one vendor changes pricing, terms, or availability, the workflow can move without a re-architecture.

The concrete first-30-day delivery plan

Most accounting AI projects fail in the first month for the same reason: too much time in scoping, too little in shipping. Our Build phase inverts that ratio deliberately. Week 1 has running code; week 4 has reviewable thin-slice production traffic; week 6 has a defensible accuracy baseline against the labelled test set.

The shape of the first week is opinionated. By end of day Wednesday, the retrieval index is loaded with the first batch of approved sources. By end of day Friday, the intake classifier is hitting the labelled test set with an initial accuracy number. The number is intentionally not impressive — it is a baseline against which weeks 2 and 3 measure progress. Most teams underestimate how motivating that early concrete number is for both the operator team (it stops feeling abstract) and the engineering team (the eval feedback loop is closing).

From week 2 onward the cadence is metric-driven. Every Friday produces a delta report against the labelled test set: which slices improved, which regressed, what the next iteration targets. The operator team participates in the Friday review; their judgment on edge cases becomes the next iteration's prompt or retrieval tweak. By week 6, the system has been through 12-15 evaluation cycles, each with accounting-specific calibration, each tied to a documented change. The workflow that hits production at the end of Build is the workflow that has survived a month of empirical correction, not the workflow that looked good in the architecture diagram.

Our Build cadence on document processing for accounting is bias-corrected against the two failure modes we have seen kill accounting AI projects most often: scoping that drifts week-by-week, and a labelled test set that arrives in week 6 instead of week 1.

We fix the scoping by signing the Build statement of work before any code is written — the deliverables are named, the integration footprint is bounded, the milestones have dates. We fix the labelled test set timing by treating it as the week-1 deliverable. Week 1 is not "scoping week" — it is "labelled-test-set week", because every subsequent engineering decision is measured against that test set.

Week 2: retrieval index live with first batch of approved sources. Week 3: intake classifier scoring against the test set, first calibration report. Week 4: action layer drafting with reviewer approval; first end-to-end case flow. Week 5-6: thin slice in production on 5-15% of routine accounting traffic, first weekly review with the operator team. Weeks 7-10: production envelope widens case-class by case-class, calibration loop tunes against the empirical evidence, exceptional cases route to enriched escalation. By day 60-70, the workflow is operating at its target envelope.