AI in operations: six use cases that work, and where a person stays in charge
BY PRAXES··1 MIN READ
TL;DR
Operations is where AI can save a great deal of time, and where a badly planned rollout causes the most disruption. A marketing draft that goes wrong gets rewritten. An invoice approved wrongly, or a supplier order sent twice, costs money and trust before anyone notices.
This guide covers what AI does well in operations today, six use cases worth considering, where AI agents fit, and how to start without putting the operation at risk. For the step-by-step method of redesigning a workflow, our guide to the AI operating model goes further.
AI in operations is the use of AI inside operational work: reading documents, classifying and routing requests, drafting responses, compiling reports, and forecasting demand or workload. It sits alongside the systems you already run, and usually connects them.
The term is easy to confuse with AIOps, which refers specifically to AI used to run IT operations, such as monitoring infrastructure and spotting incidents. This guide is about business operations in the broader sense: finance operations, customer and supplier processes, logistics, and the internal requests that keep a company moving.
These are the areas where AI tends to earn its place. The third column matters as much as the second, because plenty of operational problems are cheaper to solve with rules.
| Use case | What AI does | Is plain automation enough? | Where a person stays |
|---|---|---|---|
| Document handling | Extracts, classifies, and checks data in invoices, contracts, and forms | If the documents follow one fixed layout, often yes | Approving anything that fails a check or exceeds a threshold |
| Request triage | Reads incoming requests, works out what they need, and routes or drafts a reply | For a handful of fixed request types, yes | Complaints, unusual requests, and anything with a financial commitment |
| Reporting | Pulls figures from several systems and drafts the commentary | For the data pull, usually yes | Checking the important figures and the conclusions drawn |
| Frontline knowledge | Answers staff questions from procedures and policies | No, because the questions vary too much | Owning the procedures the answers come from |
| Forecasting and scheduling | Predicts demand or workload from historical patterns | No, this is where analytical AI does its best work | Overriding the plan when something the data cannot see changes |
| System integration | Moves data between systems and reconciles mismatches | Mostly yes, with AI for the messy matches | Resolving records the system cannot match |
The pattern across the rows is consistent. Rules handle the predictable part, AI handles the variation, and people handle the judgement. Our guide to automation versus AI explains how to tell which a step needs.
We start at the constraint: the one step that limits how much the whole operation can deliver. A company grows at the speed of its weakest constraint, however strong its other parts are, so that is where AI has the best chance of showing up on the bottom line.
The idea comes from the theory of constraints, which Eliyahu Goldratt introduced in his 1984 novel The Goal. Every operation is a system of connected steps, like a chain or a pipeline, and its speed, output, and profitability are set by its slowest or most limiting step. Strengthening any other link cannot make the chain stronger, so time and money spent improving steps that are not the bottleneck do not raise total output.
That matters for AI because it is tempting to start wherever a tool demos well. AI pointed at a step that is not the constraint becomes a clever tool sitting on the side: it works, people may like it, and it adds nothing to revenue or to the speed of decisions. The same logic runs the other way. If one step in an otherwise automated chain stays hard, that single step can undermine the entire operation.
To find the constraint, ask where work waits longest, where the queue builds up, and which step everyone has learnt to work around. Sometimes the answer is a policy rather than a machine or a system, such as an approval rule that sends every order to one manager. Policy constraints are the most common kind, and AI rarely fixes one. The fix is usually a change to the rule.
An AI agent is a system that can plan and carry out a sequence of steps across tools, such as checking an order, updating a record, and emailing a supplier, with limited supervision. Customer support is where leaders expect agentic AI to have the most impact, and an air carrier already uses agents to handle rebooking and rerouted bags so human agents can focus on complex cases.
My view is that most mid-sized operations should earn their way to autonomy. Start with AI that drafts and recommends while a person approves. Move to AI that acts within limits while a person monitors. Give full autonomy only to steps where the cost of an error is low and the controls are tested. Our guide to AI agents for business covers the controls in more detail.
Think about a warehouse that automates picking. The conveyor handles the standard orders, and someone still deals with the damaged stock, the wrong labels, and the order that does not match the system. If nobody owns that table, it fills up and the whole line slows down.
AI in operations works the same way. The system handles the standard flow, and anything it cannot process, or processes with low confidence, goes to a named person with a deadline. We design every operational automation with clear points where people review exceptions and approve important actions, because operational decisions carry consequences for customers, suppliers, and staff.
Two measures tell you whether the design is working. The straight-through processing rate is the share of items that pass through with no manual touch. The exception resolution time is how long the items that do need a person wait. A rising straight-through rate alongside a growing exceptions backlog means the bottleneck has moved downstream.
Mostly for document handling, request triage, reporting, answering staff questions from procedures, forecasting, and moving data between systems. The strongest setups use rules for the predictable steps, AI for the variation, and people for judgement and exceptions.
AIOps usually means AI used to run IT operations, such as monitoring systems and detecting incidents. AI in operations is broader and covers business processes such as finance, customer service, procurement, and logistics.
Processes that are repetitive, high in volume, well documented, and checkable. Processes that rely heavily on judgement, or that are still changing, are usually better improved first and automated later. Of the processes that fit, start with the one that is currently the constraint on the whole flow, because improving any other step will not raise output.
Not necessarily. Many improvements come from better use of the systems you already have, connected by automation, with AI added where the work varies too much for rules.
Only for steps where an error is cheap and the controls are tested. Most organisations do better starting with AI that recommends while a person approves, then extending autonomy as the evidence builds.
If you want help finding the workflows worth improving in your own operation, our AI consulting for operations teams page sets out how we work.
Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.
Take the AI Readiness Diagnostic