Documents, not code: the finance AI skeleton

The three finance-AI cases worth copying in 2026 all live in the place AI was least expected: the ERP back office. Rivian automated purchase-order accruals, Lemvigh-Muller pushed over 90% of supplier confirmations to touchless, and Toyota cut equipment diagnosis from 6 hours to 3 minutes. Three unrelated companies, three industries, three different vendors — and the same skeleton: business logic written as documents instead of code, confidence-gated posting, and audit trail as a first-class citizen. The ROI numbers are vendor-reported. The repeating pattern is the signal.

A counterintuitive fact: the most audited work got automated first

Intuition says AI should enter low-risk, low-audit processes first. Reality says otherwise: the public production finance-AI cases of 2026 cluster in purchase accruals, supplier confirmations, and equipment diagnosis — work where every step must leave a trace. The real precondition of automation is not low risk but explicit rules plus mandatory traceability. Explicit rules let the agent execute; traceability lets execution be audited. Audit requirements are not an obstacle to agentification — they are the most natural setting for it.

So the three cases below are not about how strong AI is. They are about a repeatable execution skeleton. All three come from vendor channels (official AWS, SAP, and LangChain case studies), and the figures are self-reported — we treat them as directional evidence and build our judgment on the pattern that repeats.

Case one: Rivian’s purchase-order accrual automation (AWS case study)

Automakers face a classic pain in custom-tooling procurement: tooling development spans 12-24 months, invoices arrive 18+ months after the purchase order, and GAAP requires expenses to accrue across the development period. The finance team must track hundreds of purchase orders, validate delivery schedules, calculate time-proportional accruals, and update SAP — while keeping an audit trail for external auditors. Rivian built a purchase-order accrual automation system on Amazon Bedrock AgentCore: a five-week accelerated proof of concept went straight to production (source: AWS case study).

The important part is not the technology but two design choices. First, all accrual procedures were documented as standard operating procedures and stored in a Bedrock knowledge base — the agent retrieves the relevant procedure per purchase order, and finance managers update policy by editing documents instead of requesting code changes. Second, no journal entry posts without final finance-manager approval; the agent handles volume and drafts, and a human holds the release gate. State and audit trail live in DynamoDB with SOX-compliant timestamps. AWS reports the system eliminated more than 15 days of manual work per close cycle, freeing analysts for variance analysis and decision support.

Case two: Lemvigh-Muller’s touchless procurement confirmations (SAP case study)

Lemvigh-Muller, Denmark’s largest steel and technical-equipment wholesaler, receives thousands of supplier confirmations each month across nine shared inboxes. Each confirmation can be a 10-page-plus email or PDF, and buyers must check quantities, delivery dates, and prices against purchase orders in SAP line by line. Earlier automation attempts had stalled on unstructured data and the need for precise comparison against system records (source: SAP case study).

The company deployed three custom AI agents running on SAP AI Core and SAP Cloud ERP Private: agents read emails and PDFs, extract key data, and compare against SAP purchase orders. Matches continue automatically; deviations route into follow-up workflows or are flagged for buyer review. Within weeks of go-live, more than 90% of supplier confirmations were processed touchlessly (the initial target was 80-85%), with roughly 98% matching accuracy against SAP records. The rollout started with three of nine inboxes, and 23% of suppliers were onboarded in the first weeks. Two caveats keep this honest: the figures come from SAP’s own channel, and this is first-weeks data, not a steady state — but the skeleton — rules agents can execute, deviations to humans — matches Rivian exactly.

Case three: Toyota North America’s 50+ production agents (LangChain case study)

Toyota Motor North America’s ~35-person enterprise AI team deployed agents directly on the production line. Two measurable examples: GearPal lets technicians ask a natural-language question such as “why is this machine down?” and returns relevant diagnostics, service history, and repair guidance — diagnosis dropped from 5-6 hours to 2-3 minutes; R&D GPT lets researchers query all internal R&D materials, compressing research timelines from roughly 3 years to 1 year (source: LangChain case study).

The team reports 50+ agents in production, with new-agent delivery falling from “6 months, 6 engineers” to “4 days, 1 engineer,” and each project must clear a six-to-seven-figure annual ROI threshold. Caveat: this is a LangChain-published partner case, unaudited. What is worth copying is the structural choice — Toyota’s domain knowledge is codified as a reusable skill library injected at runtime rather than hardcoded into individual agents. That is the same principle as documents-as-policy expressed differently: knowledge externalized, behavior explainable, logic reusable.

Our judgment: three bones of the same skeleton

Bone one, documents as policy. Rivian writes procedures into a knowledge base, Toyota turns domain knowledge into injectable skills, and Lemvigh-Muller has agents follow documented rules — business logic is externalized as an editable asset rather than buried in code. This is the root of agent explainability: every action maps back to the document it followed. Without this, auditing and operations are empty exercises.

Bone two, confidence-gated autonomy. High-confidence actions execute; low-confidence or material actions route to a human, who keeps final approval. All three cases keep the human gate. This is not conservatism — it is what makes automation deployable. Auditors do not trust AI; they trust the structure “AI proposes, a human approves.”

Bone three, audit trail as a first-class citizen. Rivian’s SOX-compliant trail is part of the system design, not bolted on afterward; every action, override, and posting is traceable. This bone is listed last because it is the easiest to cut — and the system that cuts it loses all credibility at its first incident.

Our second judgment concerns evidence quality: all three accounts are vendor self-reports, with no independent audit. Writing “Rivian saved 15 days” or “Lemvigh-Muller is 90% touchless” into your business case as established fact is betting your budget on someone else’s marketing numbers. The correct reading: the numbers are directional; the pattern is structural — the skeleton’s recurrence across three unrelated cases is the trustworthy part of the evidence.

Our third judgment: people did not disappear; they moved up. From entry and reconciliation to exception judgment and final approval. This is not just a job change — it is a hidden source of ROI: releasing experts from repetitive work so they handle only the deviations that need experience. That is exactly where Lemvigh-Muller’s buyers, Rivian’s finance managers, and Toyota’s technicians now sit.

Three steps for buyers

Step one: pick a “three-condition” workflow. Rules can be written as an SOP, actions need a trace, and volume is too large for manual handling — accruals, supplier-confirmation matching, and equipment diagnosis are ready examples. Do not discuss platform selection yet; first confirm such a workflow exists. Without it, everything after is empty.

Step two: turn the SOP into knowledge-base documents, and wire in confidence gating plus a human release gate. Let business owners change policy by editing documents, and design “auto-execute above threshold, route to human below” with approval points — this structure itself must get the auditor’s nod.

Step three: design the audit chain from day one. Every action, override, and posting lands in an immutable log; human review overrides feed back into SOP updates, closing the loop. The decision question for you: in your back office, which workflow has rules that can be written as documents and actions that must be audited? If you can answer, it is your first agent. If you cannot, do not put an agent there yet.

References: Rivian purchase-order accrual automation (AWS case study, vendor-reported) — https://aws.amazon.com/blogs/awsforsap/how-rivian-accelerated-finance-operations-with-ai-agents-on-amazon-bedrock/ ; Lemvigh-Muller touchless procurement (SAP case study, vendor-reported) — https://www.sap.com/finland/asset/dynamic/2026/06/be80dd33-587f-0010-bca6-c68f7e60039b.html ; Toyota North America 50+ production agents (LangChain case study, vendor-reported) — https://www.langchain.com/blog/how-toyota-north-america-put-enterprise-ai-on-the-balance-sheet-with-deep-agents-and-langsmith