A demonstration is the start of the question
An AI demonstration can turn a document into structured fields or produce a plausible answer in seconds. That shows a capability. It does not yet show whether the complete business workflow becomes more useful, more dependable or worth operating. The missing evidence usually concerns the surrounding work: preparation, review, correction, approvals and exceptions.
Before selecting an automation approach, establish a baseline for the task as it exists. Use a bounded workflow and representative observations. A supplier-document check is a useful illustrative example: information arrives, someone reads it, validates it and decides whether an internal record can be updated.
1. Volume: what arrives, and when?
Count the units of work and describe their variation. A monthly total can hide end-of-period peaks, large attachments or a handful of unusual formats that dominate effort. Separate document families, languages or request types when their handling differs. Record how demand changes rather than assuming the observed week is typical.
For the supplier workflow, distinguish incoming documents from completed checks. A resubmitted document may represent another processing attempt, not another business outcome. That distinction keeps later measures of throughput and cost honest and helps size a sensible evaluation set.
2. Handling effort: where does time actually go?
Measure active handling separately from elapsed time. Waiting for approval may dominate the calendar while reading consumes only part of a person’s effort. Break the task into preparation, interpretation, validation, entry and correction. Include interruptions or handoffs when they materially affect the work.
Then ask which step the proposed system changes. Faster extraction may leave validation unchanged or make it slower if reviewers must check every field against the original. Measure the proposed workflow end to end, including new review and administration work, before describing a benefit.
3. Quality and exceptions: what counts as acceptable?
Define correctness at the level that matters to the action. A missing supplier name and an incorrect bank detail do not carry the same consequence. Record current errors, rework and exception categories, then agree acceptance rules with the responsible business owner. A model’s confident wording is not a quality measure.
Build an evaluation set that includes ordinary examples and difficult cases such as incomplete pages, conflicting values or unfamiliar layouts. Preserve a separate set for checking changes. Decide which outputs require approval, which must be rejected and who resolves uncertainty. NIST’s AI risk guidance provides a useful basis for considering evaluation and human responsibilities across an AI system’s lifecycle.
4. Usable data: can the workflow access what it needs?
Inventory the actual inputs, their owners and their permitted use. Check readability, structure, currency and representative coverage. For a knowledge assistant, an outdated document can be retrieved successfully and still lead to the wrong answer. For extraction, a clean sample may conceal poor scans in everyday work.
Design access through the full path from source to output. Indexing a restricted document does not make it suitable for every user. Agree processing boundaries, retention and deletion needs before expanding the dataset. Where permissions or source quality are unresolved, record them as dependencies rather than hiding them inside a prototype.
5. Economics: count the complete operating model
Include implementation, integration, model or processing usage, storage, review, exception handling, evaluation and ongoing maintenance. Estimate ranges where demand or quality is uncertain. Compare the proposal with simpler alternatives such as validation rules, improved forms or a clearer approval process. AI should earn its place in the workflow.
Released capacity and cash savings are different outcomes. If people complete the same volume with less handling effort, that may create capacity for other work. It becomes cash savings only when an actual cost changes. Describe the intended use of that capacity and assign someone to verify the outcome.
Turn the baselines into a bounded decision
Bring the five baselines into one short evaluation plan: the workflow, representative inputs, quality threshold, approval boundary and full cost assumptions. Define a stopping condition as well as a success condition. Expand only after evidence supports the next decision.
HexDelta’s AI & Data Solutions work starts with that practical assessment. If you are considering document processing or an internal knowledge assistant, discuss a bounded workflow and the evidence you need before committing to wider automation.