Resources ยท Guide

Human-in-the-Loop AI: Designing Review That Scales

Review is a control. Designed well it scales; designed badly it becomes the new backlog. How to tell the difference.

Woman pointing at charts on a laptop screen while a colleague reviews the figures with her in an office

Every vendor says a human stays in the loop. Few say what the human sees, how many items reach them, what they are allowed to decide, and what happens to their corrections. Those four answers decide whether review is a control that scales or a queue that quietly becomes the new backlog. This guide is the design we apply inside DocPro and FinSuite, written so you can apply it to any automation, ours or not.

Review is a control, not a safety net

A safety net catches failures. A control decides in advance which outcomes need a person, and gets that person the right evidence at the right time. The difference matters because a safety net is sized by the failure rate, which you do not control, while a control is sized by policy, which you do.

Treat review like any other finance control. It has a trigger (what sends an item to a person), a population (which items are eligible), an owner (whose queue it is), a decision set (what the reviewer may do), a record (what is kept) and a measure (what it costs and what it catches). If any of those is undefined, you do not have a control. You have a pile.

The Human-in-the-Loop technology page explains where review sits between the model, the rules and the audit trail. This guide is about the design choices inside it.

Decide what triggers review

There are four legitimate triggers, and they should stay separate in the design because different people answer them.

  • Confidence. The model is unsure about one field. The reviewer confirms or corrects a single value.
  • Validation. The fields were read with confidence, but a business check failed: the lines do not add up, the PO does not exist, the GSTIN is inactive. The reviewer resolves a business problem, not a reading problem.
  • Policy. Nothing is wrong, but the rules say a person signs off: invoices above a band, a new vendor, a change of bank details. This is approval, and the approver is often not the reviewer.
  • Sampling. The item passed without a flag and is pulled for inspection to measure the system. The reviewer grades the system, not the item.

Mixing these in one queue is the most common design error we see. A validation exception needs someone who can call the supplier. A low-confidence date needs someone who can read a scan. A policy approval needs someone with authority. One queue means the wrong person sees everything, and everything waits for the one person who can handle all of it.

Thresholds by field and by value

Confidence is produced per field, so thresholds belong per field. A single document-level threshold, hold anything below ninety, hides the fact that a wrong invoice date and a wrong invoice total have very different consequences.

Two rules of thumb. Set the threshold by the cost of being wrong: the total, the tax, the bank account and the GSTIN carry a higher bar than a remarks line or a description. And make the threshold value-aware where the cost scales with the amount. A total read at moderate confidence on a small invoice can post and be sampled later; the same confidence on a large one should stop.

Thresholds are not permanent. Start conservative, measure the override rate per field (how often the reviewer changes the value the system proposed), and move a threshold only on that evidence. A field that reviewers confirm unchanged almost every time is a field whose threshold is too high; you are paying people to agree with the software. A field they correct often is a field the model or the document source needs work on, and no threshold will fix that.

What a threshold should never do is drop the item silently. Every hold is visible, owned and aged.

Sampling: reviewing what the system got right

The items that never reach a person are the ones you know least about. A fixed random sample of straight-through items, reviewed as if they had been flagged, is how you detect drift: a new supplier layout, a scanner setting that changed, a rule that stopped firing after an ERP upgrade.

Keep the sample small and the grading structured. The reviewer records whether each field was correct, which is a different question from whether the invoice was fine. Over a quarter that gives you a measured accuracy by field and by document source, which is the number you actually need when someone asks whether a threshold can come down. Stratify the sample by what matters to you, usually supplier, document type and value band, rather than drawing it flat across everything.

Design the queue, not just the rule

The rule decides what is held. The queue decides whether it gets cleared. Most of the work in human-in-the-loop design is here, and most of the failures are too.

One morning's onboarding batch · Illustrative example

PositionApplicationsWhy
Packs read and cross-checked340—
Cleared to the next stage311All checks passed
Held — document expiring12Inside review window
Held — data mismatch9Typed name differs
Held — image unreadable8Re-capture requested
29 to a reviewer Each arrives with the field that failed named, rather than the whole pack sent back.

A reviewer opening a held case sees one line to settle. That is the difference between a queue that clears and one that grows.

The batch above is an onboarding run, but the principle is the same for invoices or statements. A well-formed held item arrives with the one thing that stopped it: the field, the value the system read, the alternatives it considered, the crop of the document where the value sits, and the check that failed. The reviewer should be able to decide without opening the original, and open it in one click when they need to. A reviewer who has to re-read the whole document to settle one field is doing capture, not review, and that queue will grow faster than it clears.

Other properties of a queue that clears:

  • Ownership. Each queue has a named owner and a backup. Items do not sit in a shared inbox hoping to be noticed.
  • Age. Every item shows how long it has been held and what the service expectation is. The oldest sit at the top, not at the bottom.
  • Grouping. Items with the same supplier or the same failure are presented together, so one sitting resolves a dozen with one cause.
  • Fix once. If one cause blocks twenty items, an inactive vendor GSTIN say, the reviewer fixes the cause once and the twenty are re-run.
  • Priority. Due dates and value decide the order, not the time of arrival.

The KYC and Identity Agent page shows the held-case view for document packs; FinSuite AP Automation shows the payables version, where validation and policy holds live side by side but in separate lanes.

Escalation and the decisions a reviewer can make

Limit the decision set and make every decision a recorded event. For a field review the options are accept the proposed value, correct it, or reject the document with a reason. For a validation exception: resolve it (attach the missing PO, confirm the GRN), send it back to the supplier, or escalate. For a policy approval: approve, reject, or delegate within the rules.

Escalation is a second queue with its own owner, not a comment thread. The usual paths are a second reviewer for a correction above a value band, the process owner for a repeated failure from one source, and the supplier for anything only they can fix. Each escalation carries the item's full history, so the second person starts where the first stopped rather than from the beginning.

One rule we hold to: a reviewer can change the data on an item and the decision on an item. A reviewer cannot change the rule or the threshold from inside the queue. Rule changes are a separate, governed action with their own approval, which is what keeps the control a control. Business Rules describes how that separation is kept.

Close the loop: corrections, rules, thresholds

Every correction is information. Routed well, it does three jobs. It fixes the item. It becomes a signal for the model, where adaptation is supported, or a candidate rule where the fix was a pattern rather than a misreading (this supplier always prints the PO number in the remarks line). And it feeds the override statistics that decide whether thresholds move.

The mechanics are plain. Tag each correction with its cause: misread, document quality, supplier layout, missing reference data, wrong rule. Read the tags monthly. Act on the top causes. A review process that only fixes items and never changes the system is paying rent on the same problem every month.

Measure review load before it becomes the backlog

Measure four things from the first week, before anyone has an opinion about whether the queue is too big.

  • Touch rate: the share of items that reached a person, split by trigger. This is the headline, and the split tells you which lever to pull.
  • Time per decision: the median and the tail, by trigger and by reviewer. Long tails usually mean missing evidence, not slow people.
  • Override rate by field: how often the reviewer changed what the system proposed. High means the model or the source needs work; near zero means the threshold is wasting people.
  • Queue age: how long the oldest held item has waited, and how many are older than the service expectation.

Report these beside the business measures (cycle time, cost per document, exceptions by supplier) so review is seen as part of the process rather than an overhead on it. Then set a budget: the number of review hours the process may consume in a week. When the budget is breached, the response is to fix causes, not to add reviewers.

Where review is the wrong control

Some things should not go to a reviewer at all. Deterministic checks (does the total equal the sum of the lines, does the PO exist, is the GSTIN well formed) are rules, and routing them to a person to double check is a waste of the person. Unsupported document types should be routed to the process that handles them, not held in a queue nobody can resolve. Repeated failures from one source are a supplier conversation, not a review item.

Equally, there are decisions where review stays the right control however good the model gets: changes to vendor bank details, credits in contested tax categories, payments above an authority band. The goal is not zero review. It is review that is deliberate, and AI Security and Governance is where those permanent holds are defined and audited.

How to test this on your own documents

Pick one document type and one week of volume. You need a reviewer, a stopwatch and a spreadsheet.

  1. Run the week through capture with field-level confidence and your validation rules. Record what would have been held, and under which trigger.
  2. Count the touch rate by trigger. If validation dominates, the fix is reference data and rules. If confidence dominates, start with the fields and the document sources that drive it.
  3. Have one reviewer clear the held items with only the evidence a good queue would show. Time each decision, and note every time they had to open the original.
  4. Pull a random sample of the unflagged items and grade them field by field.
  5. Tag every correction with its cause and read the tags.

A week of this tells you the shape of the queue you would be running, and whether it clears. The DocPro document test runs the first step on one of your documents in a few minutes; per-field confidence is shown in a demo. For the payables version of the whole loop, Accounts Payable Automation walks through where each trigger sits; for the document side, Intelligent Document Processing covers what confidence means and what it does not.

Where PTAS AI fits

Each part of the loop, and the page that owns it

Part of the designWhere it is handled
Confidence per field and review of a single valueDocPro
Validation exceptions and policy approvals in payablesFinSuite
Rule and threshold changes as a governed actionBusiness Rules
The control model: trigger, owner, record, measureHuman-in-the-Loop
Audit of every hold, decision and overrideAI Security and Governance

Next step

See How Many Items Would Reach a Person

Upload one of your documents and see how each field reads. Or walk us through your approval policy and we will map the queues with you.