The AI that approves invoices, and the question nobody answered

SSee · The process as it actually runs

Adding AI to one step does not make the process better. It can just move the queue.

An employee raises a purchase, a supplier sends an invoice, and a document service reads it. An AI model recommends an approval, and finance posts the payment. The business added the AI step to cut manual keying and approvals, but kept the old hand-offs around it, and nobody recorded how long the whole journey took before or after.

Walk the path and the bottlenecks are not where the AI sits. Documents wait to be re-keyed, exceptions wait for a reviewer, and larger amounts wait for an approver who is busy. Errors cluster at the same hand-offs: missing purchase-order references, totals that do not match, and unfamiliar supplier formats.

• Queues at the hand-offs — Time is lost between steps, not inside the model.
• Recurring error patterns — Missing PO references, mismatched totals and unfamiliar supplier documents recur and are handled ad hoc.
• AI sits mid-process — It recommends, but nothing in the design says what happens next or who may act on it.
• No baseline — Without a before-measure, nobody can say whether the end-to-end process improved.

AAssess · What to automate, and what could go wrong

Some steps suit automation. Approval authority is not one of them until someone defines it.

Reading documents, matching them to purchase orders and routing work suit automation: high volume, rule-bound and checkable. Approving spend is different. It carries authority, and authority comes from policy and delegation, not from a model's confidence.

The risks are specific. Privacy: documents hold personal and commercial data, so where they are processed matters, and readiness never means sending confidential documents to a public model. Accuracy: models misread and can sound sure. Access: a knowledge assistant that retrieves confidential information without honouring access filters can show people what they could never open directly. Approval authority: if no one states what the AI may approve alone, its limit is whatever the workflow allows. Exceptions: unusual cases need a defined path to a person.

• Privacy and data location — Process documents in an approved private or enterprise environment with data controls, not a public model.
• Accuracy — Extraction and recommendations are probabilistic, so a deterministic check has to sit behind them.
• Approval authority undefined — With no stated boundary, the AI can effectively approve anything the workflow lets through.
• Access and exceptions — Assistants must respect existing access controls, and unusual cases need a named human route.

NoneDefined AI approval boundary — synthetic scenario as deployed

VValidate · Run the transactions

Simulate real cases and see whether automation proceeds, stops, or routes to someone authorised.

Innovera's validation runs representative transactions, clean and awkward, through the workflow under different boundary settings, and records what happens at each step: does it proceed, stop, or go to a person with the authority to decide? The same population is measured manually, so the comparison is end to end rather than step by step.

The pattern to look for: the setting that looks fastest is rarely the one that keeps approvals inside policy. A rules gate that is independent of the model keeps the boundary real, and an audit trail lets you prove it afterwards. Both are checked, not assumed.

• Speed alone misleads — Removing the gate makes cycle time and review effort look excellent while compliance falls.
• Enforce the boundary outside the model — Authorisation is checked in the workflow, not trusted to the model's own judgement.
• Exceptions need an owner — A transaction routed to a person only helps if that person has the authority and the evidence.

EEvolve · The better state

An orchestrated workflow where AI has a defined role, a limit and a record.

The better design treats AI as one governed component in an orchestrated workflow. Models and agents act inside stated boundaries: what they may read, what they may recommend, and what they may approve alone. Anything above those boundaries goes to an authorised person with the evidence in front of them, and the decision is logged.

Productivity is then measured end to end against the baseline: cycle time, cost per completed transaction, first-pass accuracy, exception rate, review effort and approval-control compliance. Governance continues after go-live, because models, suppliers and policies change. For work that crosses organisations, Innovera's InterOps platform, which connects systems and data sources, is one example of a layer that could support permitted cross-organisation workflows. That use is conceptual: confirm current availability.

• Orchestrated workflow — Document reading, AI recommendation, rules, human review and posting are designed as one process.
• Controlled agent boundaries — Each model or agent has a stated scope, a limit and a route to a person.
• Measured productivity — Six end-to-end measures tracked against the pre-AI baseline.
• Audit trail and ongoing governance — Every decision can be reconstructed, and the boundary is reviewed as things change.

100%Approval-control compliance — illustrative simulation with a gate, a defined limit and an audit log

Outcomes shown are illustrative and not verified customer results. The demo uses five scripted transactions, synthetic handling times and an illustrative 82/18 mix of clean and awkward invoices. AI failures are scripted, not produced by a live model, and the figures show what to measure, not what to expect. The InterOps reference is conceptual and its current availability for this use must be confirmed.