BY NUMBER7 AIWhy the human remainder is easy to miss
An AP team may already use email ingestion, OCR, an accounting platform, approval software, spreadsheets, shared inboxes, and an outsourcing partner. Each tool can be working as designed while people still bridge the gaps between them.
Traditional process maps usually highlight the official path: receive, extract, approve, post, pay, reconcile. The human remainder lives in the unofficial path: identify which client an email belongs to, split a batch, confirm a duplicate, find a missing receipt, correct a tax field, decide a ledger, chase an approver, re-enter a failed record, check whether a timeout actually created a bill, or explain the exception to the next person.
Because these actions are distributed across people and systems, leaders often see headcount or backlog but not the actual residual-work inventory.
Formal definition
Human remainder: the set of human actions, decisions, waits, corrections, and handoffs still required for an in-scope transaction to reach an agreed terminal state after the current technology and operating process complete their automated work.
The definition requires four boundaries:
- In-scope population: which transactions count;
- Starting event: for example, first receipt in a controlled channel;
- Terminal state: for example, a verified bill posted to the intended ledger;
- Current stack: the software, integrations, people, and procedures being measured.
Change any boundary and the remainder changes.
The human-remainder inventory
| Work class | Typical action | Evidence to capture | Automation opportunity |
|---|---|---|---|
| Intake normalization | Rename, split, route, or assign files | Source channel, batch, page lineage, client/entity | Classification and boundary detection |
| Capture correction | Fix extracted header or line fields | Proposed value, source region, corrected value | Better extraction plus targeted review |
| Identity/context | Select vendor, entity, client, or master record | Candidates, master data, decision | Entity resolution and disambiguation |
| Evidence assembly | Find PO, receipt, contract, or clarification | Missing artifact, request, response, provenance | Evidence retrieval and request workflows |
| Accounting judgment | Choose ledger, tax, class, department, or period | Proposal, policy, prior context, final choice | Contextual suggestion within controls |
| Control/approval | Determine route or chase authority | Rule, threshold, approver, timestamps | Rule evaluation and sequenced approval |
| Exception resolution | Decide what a discrepancy means | Exception type, required evidence, owner, decision | Smallest-decision queue and resolution playbook |
| Integration/handoff | Re-key data or repair failed transfers | Source record, target record, response | API integration, idempotency, retry control |
| Verification/reopen | Check completion or correct later | Target state, correction, reopen reason | Post-action verification and feedback |
This inventory avoids a common mistake: defining all review as one undifferentiated “human-in-the-loop” step. A two-second confirmation and a twenty-minute evidence investigation are not the same remainder.
Measurement model
Measure actual action events:
Human-remainder minutes = Σ (human action count by class × median active minutes for that class)
Also report:
- share of transactions with at least one human touch;
- touches per handled transaction;
- median and 90th-percentile active minutes;
- elapsed exception-resolution time;
- reopen rate;
- time spent waiting versus working;
- distribution by client, vendor, document family, and exception type.
Avoid one composite “human remainder score” until the organization has defined failure costs and weights. Minutes, touches, and terminal outcomes are easier to audit.
Worked example
Assume an illustrative multi-client team receives 1,000 eligible supplier invoices in a month. Its current tools extract most fields and sync approved bills, but action logging shows:
| Residual work | Actions | Median active minutes | Monthly minutes |
|---|---|---|---|
| Capture corrections | 300 | 2.0 | 600 |
| Missing/context evidence | 180 | 5.0 | 900 |
| Coding decisions | 120 | 3.0 | 360 |
| Approval follow-up | 80 | 4.0 | 320 |
| Reopen/correction | 40 | 6.0 | 240 |
| Total | 720 actions | — | 2,420 minutes |
The visible data-entry problem is only 600 minutes. The larger remainder is evidence, decisions, follow-up, and rework. At roughly 40 active hours per month, the improvement case is not “replace the accounting platform.” It is “remove or compress the high-frequency residual decisions around it.”
This example is illustrative, not a benchmark.
The smallest-human-decision rule
When an item cannot proceed safely, the system should ask for the smallest decision that resolves the named state.
Weak task:
Review invoice.
Better task:
Two active vendor records match tax ID ending 1842. Choose the intended record or mark the invoice as a new vendor.
The second task reduces rediscovery. It presents the reason, evidence, candidates, and allowed actions. Even when automation cannot remove the decision, it can shrink the human remainder surrounding it.
Counterexample: a falling exception rate with unchanged work
Suppose a team reclassifies low-confidence extractions as “routine review” instead of “exceptions.” The dashboard’s exception rate falls, but staff still inspect the same items. Nothing operational improved.
Or suppose exception count falls while each remaining case requires three people and two handoffs. The count improved; the human remainder may have increased. That is why touch events and active minutes must sit beside exception and STP rates.
How to run a human-remainder study
- Select a representative period or sample; include difficult vendors, clients, tax cases, credits, and multi-page inputs.
- Freeze the start event, terminal state, inclusion rules, and system boundary.
- Instrument every human action with actor role, timestamp, transaction ID, action class, reason, and outcome.
- Separate active work from waiting.
- Group actions by exception type and root cause.
- Find the high-volume/high-minute intersection.
- Decide whether to eliminate, automate, pre-assemble, simplify, or better route each action.
- Re-measure the same population definition after the change.
Limitations
The model does not imply that all human work is waste. Judgment, accountability, materiality decisions, policy exceptions, and relationship management can be valuable controls. It also does not price the risk of incorrect automation. A successful redesign reduces low-value human effort while preserving or improving the control outcome.
Time studies can change behavior, miss invisible work, or overstate active time. Combine event data with observation and practitioner interviews.
Sources
- ServiceNow documentation: invoice exceptions — examples of discrepancies that interrupt invoice processing.
- SAP Learning: handling invoice exceptions — examples such as missing receipts, mismatches, duplicates, and tax variance.
- Number7 Research model: human remainder inventory, revision 1.0.
How to cite
Number7 Research. “The Human Remainder in Accounts Payable Automation.” Number7AI, revision 1.0, 7 September 2026. https://number7ai.com/research/human-remainder.
Revision history
| Revision | Date | Change | Reviewer |
|---|---|---|---|
| 1.0 | 7 September 2026 | Initial definition, inventory, measurement model, and example | Number7 AI editorial review |
Next step
Instrument the human actions around 50 representative invoices before selecting the next automation project.