This page is a companion to exception handling records for AI agents, which works through entitlement, promotion and cold-chain cases. Here the job is different: a reusable classification a partner can carry into an unfamiliar business to recognize which exceptions it records, and whether each one leaves the full chain.
Why the hard cases carry the evaluation
A test made only of routine cases rewards an agent that memorizes the happy path. Most real operational cost sits elsewhere: the invoice that does not match its purchase order, the shipment that arrives out of range, the approval nobody with authority was around to give. Those are the moments when a person had to judge, act within limits and check the result.
That is also what the agent-data market says it wants, though no announcement proves demand for a particular company. On 2026-10-09 micro1 announced a commitment to spend $1 billion over 12 months acquiring enterprise operational data, saying it helps agents learn to navigate workflows and make decisions. Appen's agentic AI page (checked 2026-10-18) lists golden trajectories and failure-mode taxonomies among its services. A failure-mode taxonomy describes how agents go wrong; an exception taxonomy describes how businesses already went wrong and recovered, with records. The second is what a company can license.
The six-class exception taxonomy
Each class is defined by what breaks, not by industry. The two right-hand columns name the record that must exist for the case to be usable.
| Class | What breaks | Typical trigger | Authorized action recorded as | Checked result recorded as |
|---|---|---|---|---|
| Rule conflict | Two valid rules give different answers | A legacy contract term contradicts the current price list | Override with a named approver and reason | Corrected invoice, balance or entitlement after the fix |
| Missing input | A required document or reading is absent | A delivery arrives without its certificate | Hold, or a substitute source with sign-off | Release, rejection or recovery once the input arrives |
| Out of tolerance | A measured value leaves its allowed range | A temperature, weight or dimension outside spec | Disposition with written reasoning | Inspection, test or customer acceptance |
| Duplicate or mismatch | Two records claim the same thing | Two invoices for one delivery; a customer with two accounts | Merge, void or correction citing both records | Reconciliation that closes with no residual |
| Authority gap | The person allowed to decide is unavailable or the limit is exceeded | A refund above a clerk's limit with no manager on shift | Escalation or delegated approval with its own record | Post-review confirming or reversing the decision |
| Deadline breach | A time commitment is missed | A service-level clock or filing deadline expires | Documented remedy: credit, waiver, expedite | Customer response, penalty applied or waived |
Platforms already capture parts of this. Salesforce's classic approval history, for example, shows statuses including Submitted, Approved, Rejected, Reassigned and Recalled, along with the actual approver, which is the "authorized action" column for an authority gap. Regulated industries make some chains mandatory: under U.S. drug manufacturing rules, any unexplained discrepancy must be thoroughly investigated, extended to associated batches, and written up with conclusions and follow-up, which records the action and result for out-of-tolerance and mismatch cases.
Reading the taxonomy in an unfamiliar business (Illustrative)
The classes are generic on purpose. One fictional example per class shows how they map to different companies:
- Rule conflict. A 400-person B2B software company merges an acquired customer base; some accounts carry legacy features their new plan excludes. Support grants an exception with a named approver.
- Missing input. A 320-person building materials distributor receives steel without mill certificates. The yard holds the load; quality releases it when the certificate arrives two days later.
- Out of tolerance. A 250-person pharmaceutical distributor's pallet reads above range after a truck transfer. Quality reviews stability data and decides per lot; the temperature excursion worked case walks through it step by step.
- Duplicate or mismatch. A 600-person facilities services firm receives two invoices for one job. Accounts payable voids one, citing both; the vendor statement then reconciles. See accounts payable exception handling.
- Authority gap. A 180-person clinic group needs a write-off above the billing lead's limit on a weekend. The on-call director approves by delegated authority; Monday's review confirms it.
- Deadline breach. A 900-person managed IT provider misses a four-hour response commitment. The account manager issues a service credit; the client's renewal notes record acceptance.
When a partner hears a story like one of these, the follow-up is the same: does each of the three records exist, in which system, for how many years, and can the three be linked?
When a case falls short of the full chain
The taxonomy also makes gaps easy to name. Three shortfalls recur across classes:
- Problem without action. The alert or complaint is logged, but the decision was made by phone and never written down. Common in authority gaps.
- Action without approver. The override exists, but under a shared login or a generic "admin" account. Common in rule conflicts and duplicates.
- Action without result. The fix is recorded, but nobody linked it to what happened next. Common in deadline breaches, where the customer's reaction sits in email.
A class with only the first two records can still teach how people respond. It cannot show whether the response was right, which is what evaluation needs most.
Qualification: records, owners and authority
Exception records usually span systems owned by different teams: operations or quality for the problem, the approving function for the action, finance or customer success for the result. The decision to discuss licensing belongs to the CEO, often with the COO or CFO preparing the case, not to whoever administers the systems. Customer and supplier contracts, personal data in notes, and regulated records all affect what can be licensed; see why data rights matter.
The company needs 50+ full-time employees at peak (contractors excluded), documented operating history, rights to license and an authorized contact. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English.
Partners establish metadata only: which exception classes the company sees often, which system records each link, how many years exist, and who owns them. Volumes, buyer interest, de-identification and any offer stay unknown until SourceX reviews the opportunity. For context on why operational records are sought, see why enterprise AI needs enterprise data, why operations data is valuable and vertical AI training data.
Next step
Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.
If a client can name two or more exception classes it records with all three links over several years, register as a partner, check who qualifies, and read the terms before you introduce the CEO.