Use cases
Start with the work that already keeps someone awake.
44 operations, and the reason each one is harder than it looks.
Finance and accounting
Finance work is mostly exceptions. The invoice that does not match, the expense just over the limit, the vendor whose bank details changed last week. Every one of them has a rule, a threshold and someone who has to answer for the decision.
- Invoice reconciliationThree-way match against purchase orders and receipts, with duplicate payments blocked rather than reported after the fact.
- Expense approvalsPolicy thresholds, the right manager in the chain, and a receipt attached to every exception.
- Vendor onboardingTax forms, sanctions screening and bank-detail verification, where a change of account number is the fraud.
- Collections and dunningPayment plans, escalation ladders, and accounts on legal hold that must never receive another notice.
- Revenue recognitionContract terms and milestone evidence, where recognising a month early becomes a restatement.
- Month-end closeA sequence with dependencies, where one unreconciled account holds the whole thing open.
Customer operations
Customer work is judged on the cases that go wrong. Anyone can answer the easy ticket. The question is what happens on the second complaint, the refund above someone's authority, and the onboarding that stalls because a document never arrived.
- Support triage and escalationUrgency weighed against account value, and anything a customer has raised twice going up automatically.
- Refunds and goodwill creditsWho may authorise how much, and why, on the record, before the money moves.
- Onboarding and provisioningDocuments collected, checks run, accounts opened in every system, and a flag the moment something is missing.
- Renewals and savesDiscount authority by segment, with the concession and its reason attached to the contract.
- ComplaintsRegulatory clocks that start on first contact and do not pause for an internal handover.
- Account changesOwnership, billing and entitlement moving together, so none of the three is left behind.
Operations and fulfilment
Physical operations punish optimism. The route that is legal on paper breaks a driver's hours, the substitution that suits the warehouse loses the customer, and the shortfall you cannot evidence is a shortfall you absorb.
- Order exceptionsStockouts, partial shipments and substitutions, where the right answer depends on the customer, not the SKU.
- Dispatch and routingService windows, driver hours and vehicle constraints that make the obvious route illegal.
- Returns and RMACondition assessment, restocking rules, and refund authority that changes with the reason code.
- Supplier claimsChargebacks and shortfalls, evidenced well enough to survive the supplier disputing them.
- Shift and capacity planningCertifications, rest requirements and overtime rules that differ by site.
- Inventory adjustmentsWrite-offs and recounts with an owner, because unexplained shrinkage becomes an audit finding.
Risk, compliance and audit
Compliance is where unproven automation is most dangerous and most valuable. The work is repetitive enough to automate and consequential enough that nobody will authorise it without evidence. That is exactly the gap we build for.
- KYC and AML reviewEscalation thresholds, and a filing obligation that a missed queue turns into a breach.
- Access reviewsSeparation-of-duties conflicts that only appear when two harmless permissions meet in one person.
- Policy attestationExceptions granted with an owner and an expiry, rather than granted and forgotten.
- Incident reportingStatutory deadlines counted from discovery, with the evidence assembled as you go.
- Audit preparationThe trail built during the work, not reconstructed from memory the week before.
- Third-party riskReassessment when a vendor's scope changes, not only on the annual cycle.
People and HR
HR work is high-consequence and quietly jurisdictional. The same request has different correct answers in two countries, and the cost of the wrong one is a claim rather than a complaint.
- Hiring and offersCompensation bands, approved headcount, and an approval chain that changes above a level.
- Onboarding and offboardingAccess granted on day one and genuinely revoked on the last, including the systems nobody remembers.
- Leave and accommodationsEntitlements that differ by jurisdiction and contract, where getting it wrong is a claim.
- Performance cyclesCalibration, evidence, and a decision trail that stands up if it is ever challenged.
- Payroll change controlRetroactive adjustments with an approver, because payroll errors compound monthly.
Revenue and go to market
Revenue operations fail on conflict, not on volume. Two reps on one account, a discount nobody was authorised to give, a contract that shipped with a clause legal never saw.
- Quote and discount approvalAuthority by deal size and margin, with the approval attached to the version that shipped.
- Contract routingRedlines that need legal, and the thresholds that decide which ones do.
- Lead routingTerritory, segment and named-account rules, and the conflicts they create.
- Deal registrationPartner claims on the same account, resolved by policy rather than by whoever shouts.
- Commission calculationPlan rules, clawbacks and disputes, where a silent error erodes trust for a year.
Clinical and patient operations
Healthcare operations have the strictest failure mode of any of these: the loop that does not close is a patient nobody called back. It is also the world our benchmark was built in.
- Intake and eligibilityCoverage checked before the appointment, and the questions that must be asked before anyone is seen.
- Prior authorisationPayer-specific requirements, submitted with the evidence the payer actually accepts.
- Referral follow-upLoops that must close, because the failure mode is a patient nobody called back.
- Documentation completenessNotes that satisfy both the clinician and the coder, checked before the claim goes out.
- Denials and appealsReason codes mapped to the remedy, with deadlines that do not forgive a late filing.
IT and platform
Platform work is where automation is most tempting and most load-bearing. An agent with real credentials is either your most useful colleague or your largest unreviewed change.
- Access requestsLeast privilege by default, time-bounded grants, and an approver who understands what is being asked.
- Change managementFreeze windows, rollback plans, and the dependency nobody declared.
- Incident responsePaging the right owner, and a timeline good enough for the post-mortem and the regulator.
- Vendor provisioningData-processing terms and security review completed before the tool holds real data.
- Certificate and secret rotationExpiry tracked and rotation proven, because the outage is always a lapsed credential.
The pattern underneath all of them
The happy path is never the problem.
Look back at any domain above and the difficulty is in the same four places every time: the exception nobody wrote down, the threshold that decides who may act, the authority to make a call, and the record that has to exist afterwards. Generation handles the happy path. These four are why the pilot dies in review.
If you can describe the exception, we can hold the system to it.