An invoice assistant prepares a neat summary in seconds. The finance operator then spends several minutes checking its figures, finding the missing delivery record and correcting a supplier reference. The demonstration looks fast. Whether the work improved is still an open question.
An artificial intelligence (AI) workflow pilot should answer that question before the business makes the assistant part of everyday work. Choose a useful job, establish how it is handled today and decide what evidence would justify continuing. Include the time spent checking, correcting and recovering from mistakes.
Our series has followed a hypothetical supplier-invoice assistant through permissions, untrusted documents and human approval. Here we turn that example into a pilot plan. The design and tests below are proposed; they are not results from a completed app or a live business trial.
Give the AI workflow pilot one decision to answer
“Try AI in finance” leaves too much room to claim success. A more useful question is: can the assistant prepare a source-linked invoice review pack that helps the operator reach the right decision with less total effort?
That question identifies an output, a user and a comparison. It also gives the pilot a stopping point: the operator has accepted the pack, corrected it or returned the case for more information. Payment is a later decision outside this trial.
The proposed pack contains extracted invoice details, the relevant order and receipt records, unresolved differences and a draft clarification where needed. The assistant can help read varied documents and explain discrepancies. Ordinary application rules should calculate totals, control access and record decisions. There is no reason to make a model decide whether a user has permission or whether two stored amounts are equal.
AI workflow design starts with that division of responsibility. If the real problem is that nobody records receipts, a better capture process may achieve more than an assistant. If existing software already assembles a useful pack, test that capability before commissioning another application.
The pilot question works for a bought tool or a custom build. In either case, judge the work that reaches the operator.
Keep the invoice boundary explicit
For this example, the assistant reads only the selected invoice documents and relevant supporting records. It saves proposed fields, explanations and message drafts in a review area. It cannot change supplier bank details, release payment or post transactions to the accounting ledger.
The application owns the review queue, access rules and history. The operator owns the decision. The existing accounting system remains authoritative for financial postings during the pilot. A later accounting handoff would need its own scope, controls and tests.
The Open Worldwide Application Security Project (OWASP) recommends limiting available functions and permissions, and enforcing authorisation outside the model, in its excessive agency guidance. For this trial, that means a drafting assistant receives drafting access. A prompt asking it to be careful is not a substitute for restricting the account and tools it uses.
Write down the approved document set, permitted outputs and excluded actions before connecting anything. AI agent permissions should remain the same throughout a comparison run. Adding an inbox, a new supplier directory or a sending tool changes the trial being evaluated.
Establish a baseline that includes the awkward work
Observe the current process from receiving the invoice to producing a usable review pack. Record preparation time, checking time, corrections and unresolved cases. Separate active work from waiting for someone to provide missing information.
Choose cases that reflect the work you intend to support. Include ordinary invoices and known sources of difficulty: partial deliveries, inconsistent supplier references, unreadable scans and missing purchase information. Record which cases were excluded and why. A trial restricted to clean invoices can support a conclusion about clean invoices; it cannot establish readiness for the whole queue.
Have a knowledgeable operator establish the expected fields and disposition for the evaluation cases. Where the evidence does not settle a question, the expected outcome should be an explicit hold or escalation. Do not manufacture a definitive answer simply to make scoring easier.
Compare the current process and assisted process against the same outcome standard. Watch for learning effects if someone handles the identical invoice twice: the second attempt may be faster because they remember it. Comparable case groups, or alternating which method is used first, can make the comparison fairer. Keep a separate set of unseen cases for evaluation after changing the prompts or application.
A small trial gives limited evidence. There is no universal invoice count or two-week duration that proves readiness. Agree the case mix and review point in advance, then report what was actually covered.
Set AI pilot success criteria before seeing the results
The US National Institute of Standards and Technology (NIST) organises its AI Risk Management Framework core around governing, mapping, measuring and managing risk. Its measurement guidance includes evaluation methods, realistic conditions and limitations. The scorecard below is our practical application to the invoice example, not a NIST scoring formula.
| What to assess | Evidence to collect | Decision it informs |
|---|---|---|
| Useful output | Correct fields, relevant source links and an appropriate recommendation or hold, checked against the expected case outcome | Does the pack help the operator do the job? |
| Total human effort | Preparation, source checking, correction and exception handling, with waiting time recorded separately | Has work been removed or moved into review? |
| Errors and rework | Mistakes by type and consequence, including omissions and cases incorrectly presented as ready | Which cases need a different design or remain excluded? |
| Operating cost | Model usage, document processing, hosting or subscriptions and support effort; setup costs shown separately | Is the demonstrated benefit worth sustaining? |
| Control and recovery | Results of denied-action, pause, restart and history checks | Can the operator contain and correct a failure? |
Set targets for the business context. A modest time improvement may be useful if the pack makes exceptions easier to see. A large average improvement does not compensate for an unauthorised supplier change. Treat a failed permission boundary as a stop condition, rather than averaging it into an accuracy percentage.
For rates, retain the counts and case types behind them. Report both routine and exception performance so an easy majority does not hide expensive failures. No errors observed in a small set means exactly that; it is not evidence of zero future errors.
Recovered staff time is capacity. It becomes a cash saving only when expenditure actually changes. Keep that distinction in any business case.
Test whether the reviewer can detect a wrong answer
Human approval for AI must be useful at the point of decision. The reviewer should see the source invoice, extracted fields, relevant receipt evidence and unresolved differences together. They need a practical way to correct a field, reject a recommendation or ask for more information.
In an isolated exercise with synthetic records, include a plausible but incorrect recommendation. Can the operator identify it from the evidence presented? Record the effort involved and whether the interface makes the discrepancy visible. Do not introduce deliberate errors into live financial work to run this test.
Microsoft's overreliance framework recommends evaluating whether people can understand limitations and verify outputs. It cautions that explanations and sources can increase trust even when they are unreliable. A convincing explanation is therefore part of what the pilot must test.
AI workflow review should measure the combined performance of the person and application. A high acceptance rate alone does not show whether the assistant was accurate or the reviewer stopped checking. Our discussion of human approval boundaries explains what an approval should cover and when changed evidence requires a new decision.
Plan the failures before the demonstration
An ordinary invoice demonstrates useful preparation. The exceptions show whether the proposed boundary holds. Include these cases in the pilot design:
- The receipt is missing. The assistant surfaces the gap and leaves the case pending. It does not invent proof of delivery or treat the invoice itself as receipt evidence.
- A supplier document contains instructions to send information elsewhere. The document remains evidence to inspect, without authority to change permissions or destinations. Test the application's enforced restrictions as well as the assistant's response.
- An invoice is revised after review. The changed version returns for the required decision, with the earlier version and review retained in history.
- The same document arrives twice. The application identifies or flags the possible duplicate without silently discarding a legitimate revision.
- Processing stops halfway through. The operator can determine which drafts were saved and which cases remain pending before restarting.
These are proposed acceptance tests. Ask the builder to demonstrate the actual action path, including denied requests, rather than showing only a polite refusal in chat. The earlier prompt-injection example explains why a supplier document must not become an instruction source.
AI workflow monitoring needs to distinguish received, processing, ready for review, held and failed cases. A missing result should remain visible work, not disappear into a success count.
Move from a test environment to limited use deliberately
Begin with synthetic documents and disconnected test records. This lets you exercise exceptions without exposing real supplier information or affecting accounts. Before using actual documents, establish who may access them, where they are processed, how long they are retained and whether the selected service's data handling fits the business's requirements.
A next stage can run alongside the existing process: the assistant prepares a pack, but the established process remains responsible for the operational decision. Keep those parallel results clearly separated so a test recommendation cannot be mistaken for an authorised record.
Only move to limited operational use after reviewing the earlier evidence. Name the participating users, eligible cases, accountable owner and review date. Bound the workload as well as access, so the review queue cannot quietly grow beyond the operator's capacity.
AI pilot planning also includes an exit. Someone must be able to stop new processing, revoke access, identify affected cases and restore manual handling. Decide what happens to queued work and retained documents when the trial ends. Test recovery before relying on it.
Decide what would justify continuing
At the review point, compare the evidence with the question you started with. Has preparation become easier after including checking and corrections? Do important exceptions remain visible? Can the operator explain a decision and reconstruct what happened?
If the boundary failed, stop and investigate. If the output helps but review takes too long, revise the design and evaluate it again. If the result is useful for a narrow class of invoices, that may be the right scope to retain.
Expansion should be a separate decision about a specific change. Adding another document format differs from granting permission to send messages or write accounting records. Each introduces a different question to test. AI workflow governance means someone is accountable for that decision and the evidence supporting it.
For AI workflows for finance, a successful first pilot can finish with a better review pack and no additional authority. More autonomy is not the measure of success.
The next instalment will use an invoice application build to examine these proposed controls in practice. For now, take one ordinary invoice and one difficult exception to the person who would use the assistant. Agree what a useful result looks like, what must remain under their control and what evidence would make them willing to rely on it.
