Operations is the function that absorbs everything nobody else's system handles. A shipment arrives short and someone has to reconcile it. A vendor goes silent on a document request and someone has to chase. A report is due Monday and someone spends Friday afternoon pulling numbers out of four systems that disagree with each other. None of this is strategic, none of it is optional, and almost all of it is invisible until it fails.
The reason this work resists conventional automation is that it is mostly exception handling. Rule-based tools handle the clean path — the order that matches, the form that arrives complete, the invoice that ties to its PO — and then hand everything else to a person. In most operations teams the clean path was automated years ago. What remains on the team's plate is precisely the residue that rules could not resolve, which is why headcount keeps rising even as the automation stack grows.
AI agents are interesting to operations for exactly this reason. An agent can pursue an objective across several systems, handle the variance that broke the rule, and escalate only when it genuinely cannot proceed. What follows is not an architecture discussion. It is a look at the specific ops workloads where AI agents take real work off the queue, the ones where they do not, and how to tell the difference before you commit.
Chasing: The Work That Never Ends
A meaningful share of any operations week goes to chasing people who owe you something. A supplier who has not returned a signed W-9. A branch manager sitting on an expense justification. A customer whose remittance advice never arrived, leaving an unapplied payment on the books.
Chasing is ideal agent work because it is persistent, low-judgment, and unbounded in duration. An agent can hold a follow-up open for eleven days across four channels without losing the thread, escalate on a defined schedule, recognize that a reply containing an attachment satisfies the request, and close the loop by filing the document and updating the record. It never forgets, never gets embarrassed about the fourth reminder, and never lets a request die because the person who opened it went on leave.
The practical gain is usually not hours saved on sending emails. It is the elimination of the long tail — the twelve requests per month that used to quietly expire and turn into a write-off, a compliance gap, or an aged receivable.
Exception Queues: Working the Residue
Every ops team has a queue with a name like "needs review" that only grows. Most items in it are resolvable, but each requires looking something up somewhere else, and that lookup is why they queue.
Agents are well suited here because the resolution steps are usually knowable but tedious: check whether a second delivery note exists, look at whether this supplier historically bills freight separately, confirm whether the variance falls inside the contracted tolerance. An agent that can read the contract, query the receipts table, and compare against the last six invoices from that supplier resolves a large fraction of the queue without a person opening a single tab.
Two design points matter. First, the agent should classify before it acts — a price variance, a missing receipt, and a duplicate submission are three different problems with three different resolutions, and lumping them into a generic queue is what made the queue unmanageable in the first place. Second, unresolved exceptions must be escalated with the investigation attached. An agent that hands a human a bare exception has saved nothing; one that hands over "quantity variance of 40 units, contract tolerance is 5%, prior three deliveries from this supplier were also short and credited" has done most of the thinking.
Finance-heavy queues respond fastest. Teams that have already tackled banking reconciliation automation usually find their remaining unmatched items are exactly this shape: individually resolvable, collectively expensive.
Scheduling and Dispatch
Scheduling looks like an optimization problem and is actually a communication problem. The hard part is rarely computing a feasible roster or route. It is that a technician called in sick at 6:40 a.m., the replacement needs a certification the job requires, the customer has to be told, and the change has to propagate to three systems before dispatch.
An agent handles the propagation and the negotiation, not the math. It detects the gap, identifies eligible replacements against the actual qualification data rather than a stale spreadsheet, contacts them in priority order, confirms, notifies the customer with a revised window, and updates the schedule of record. The optimizer, if you have one, still does the optimizing.
This is one of the clearest cases where agents beat both rules and humans: rules cannot handle the negotiation, and humans cannot do it at 6:40 a.m. across forty simultaneous routes.
Recurring Reporting
Most operational reporting is not analysis. It is retrieval, reconciliation, and formatting — pulling the same fifteen figures from the same six places, noticing that two of them disagree, chasing down why, and assembling the result into the format the recipient expects.
An agent can own the retrieval and the reconciliation, and more usefully, it can own the anomaly note. The valuable output is not the dashboard; it is the sentence that says days-payable-outstanding moved four days and here are the three suppliers that account for it. That sentence is what the ops manager was going to spend an hour producing.
Be deliberate about scope. Agents should assemble and flag; they should not be the sole source of a number that goes to an auditor or a board without a traceable path back to the system of record. Every figure in an agent-produced report needs a click-through to its origin.
Data Hygiene Across Systems
The least glamorous and often highest-return workload. Duplicate vendor records, addresses that differ by a suffix, cost centers that were deprecated but still receiving postings, employees active in one system and terminated in another.
Agents work well here because the task is comparison at volume with judgment at the margin — deciding whether "Northgate Logistics LLC" and "Northgate Logistics, L.L.C." are the same entity requires looking at tax IDs, bank details, and invoice history, not string similarity. A remediation agent can run continuously, propose merges with evidence, execute the unambiguous ones, and route the genuinely ambiguous ones to a data owner.
The knock-on effect is what makes this worth doing first. Every downstream automation gets more accurate when the master data underneath it stops contradicting itself.
What Goes Wrong in the First Ninety Days
Three patterns account for most disappointing deployments.
Starting with the highest-volume process instead of the highest-variance one. High-volume, low-variance work was already automated or already cheap per unit. The return is in the messy middle: moderate volume, high variance, currently consuming senior people.
No defined escalation contract. Teams deploy an agent without deciding what it must never do alone and what it must always escalate. The result is either an agent that escalates everything, which is a slower version of the old process, or one that acts where it should have asked. Write the boundary down before launch, in the same terms you would write it for a new hire.
Measuring adoption instead of outcomes. Task counts and messages sent tell you nothing. Measure cycle time on the target workload, the share of items closed without human touch, the reopen rate on agent-closed items, and hours returned to the team. The reopen rate is the one that matters most and the one most often skipped — an agent closing items that come back two days later is generating work, not absorbing it.
Where Agents Should Not Go First
Skip anything where the cost of a wrong action exceeds the cost of the manual process by a wide margin and no reversal path exists — irreversible payments, contract terminations, anything with a regulatory filing deadline attached. Skip processes that are about to be redesigned; automating a workflow two months before it changes wastes the effort twice. And skip work that is genuinely relationship-driven, such as negotiating with a supplier during a supply disruption, where the point of the interaction is the relationship rather than the transaction.
Teams that have already quantified the manual load in a specific area tend to move fastest, because they know what they are comparing against. The expense management ROI analysis approach — cost per transaction, exception rate, cycle time — is the right baseline to build before an agent touches anything.
The most reliable starting point is a single queue your team quietly resents. Pick it, instrument its current cost for two weeks, then scope an agent against those numbers rather than against a demo. If you want a second opinion on which queue to pick, our team can review your workload mix and point at the one most likely to pay back first.
