An invoice arrives in a format your company did not design — a PDF, a scan, an EDI feed, a portal upload. The job is to turn it into structured fields, check those against the purchase order, vendor master, and policy, route the exceptions a person must judge, and post the clean ones to the ERP. The deliverable is a structured, validated payable, ready to pay.
The 8 tasks — the nature of each, and the oversight it needs
Tag each task in plain terms — what kind of work it is and how hands-off it can run — before any mention of AI. The kind of work is what later decides which tool, if any, fits.
Seven of the eight tasks are rule-based transaction-and-routing work or light prediction — work that runs the same way once the inputs are clean. The exception, and the bottleneck, is the one reading task: interpreting a document the company did not design. That single perception-heavy step gates everything downstream — you cannot validate, match, or post what you have not reliably read. Get the reading right and the rest is plumbing.
Financial controls, fraud / duplicate exposure, and audit relevance keep a human in the loop on exceptions and approvals.
The question isn’t only “is there savings” — it’s can I run this better: cheaper, faster, higher quality, better service? Here’s what best-in-class looks like, and how teams get there. (How much of it AI specifically drives — and how proven that is — is Section 04.)
Process discipline first, then automation — AI is one slice of the second column, not the whole answer.
- E-invoicing & PEPPOL adoption
- Supplier portals / self-service
- "No-PO-no-pay" policy
- Vendor master-data hygiene
- Tolerance & exception-rule tuning
- IDP / OCR capture
- Straight-through (touchless) processing
- Automated three-way match
- Duplicate & fraud detection
“Who runs the work” is its own question, separate from AI. AI shows up across these options — sometimes heavily, sometimes not at all. Vendor-neutral; the real options mapped to AP01.
Recall the tasks and their nature from Section 01. AI is one lever, not the whole story — the mix below is simply the result of matching the right kind of solution to each kind of work, weighted by where the work concentrates.
Document AI leads the mix (~40%) not because most tasks are reading, but because the one reading task is the binding constraint — nothing downstream can act until extraction succeeds. Agentic/RPA does the most steps (routing, validation, posting) at ~30% but is wrapped around the document-AI core. The precise percentages are a directional estimate; the dominant call is research-supported (Everest, McKinsey).
The right lever fits your volume, variability, control needs, and appetite to operate a system. Start here.
The autonomy question: agent or copilot?
Whichever delivery model you pick, one choice cuts across them — who presses enter.
AI agent
Runs the steps end-to-end, completes the clean cases on its own, and routes only the exceptions to a person.
AI copilot
Sits beside the person and speeds up each step; the human acts on every decision.
What to evaluate — whichever you choose
- Accuracy on your own inputs — vendor benchmarks are on clean data; test your messiest cases.
- Straight-through / touchless rate — the real efficiency number, not “AI-powered.”
- Exception-handling experience — most of your team's time goes here, not the happy path.
- ERP write-back & integration depth — does it post cleanly to your system of record?
- Data residency — does data leave your environment, and is that acceptable to compliance?
- The accountability surface — what happens, and who owns it, when the model is confidently wrong?
See every lever across your processes
Run your portfolio through the assessment — work profile, improvement potential, confidence, and executor options across all your blocks, scored against 127 enterprise subprocesses.
Open the AI Value Assessment →