What can agentic loops be used for? Real small-business examples

What makes a job a good fit for an agentic loop?

An agentic loop is the cycle an agent repeats: decide a step, take it, check the result, go again. It suits some jobs and not others. If the terms are new, start with our plain-English guide to what agentic AI is.

Four signs point towards a loop.

• Each case is different. You cannot write the steps down before the job arrives. A supplier email about a late delivery touches different jobs from one about a changed quantity, so what to check next depends on what the last check found.
• It needs lookups before it acts. The AI has to consult something first, such as the job list, the ledger, the stock count or a calendar. Each lookup is a tool: an action the AI is allowed to ask for, which software then carries out.
• There is a clear end point. "Every affected job has a draft update" is a finish line. "Improve our customer service" is not. Without a finish line, a loop can keep going until a limit stops it.
• A mistake can be caught before it lands. The result is a draft, a proposal or a list that a person reads first. If a wrong answer would already have reached a customer, the job is not ready for a loop.

The last sign matters most, and it is why every example below has the same shape: the steps, the tools and the point where a person approves. The first sign has a quick test. Could you write down every step before the AI sees the job? If you could, a fixed workflow will do.

Which office jobs suit it?

Three office jobs fit the four signs. In each one the AI reads, looks things up and drafts. It does not send, pay or file anything on its own.

Supplier emails into job updates. This is the example our glossary uses to explain agentic AI, and it fits all four signs. One email says a delivery is late, another changes a quantity and another is only a receipt, so what the AI checks next depends on what it has just read.

• Steps. The loop reads each email, works out which open jobs it affects, looks up what it needs about each one and drafts an update for the team.
• Tools. A way to read the inbox, a search of the open jobs, a lookup of orders and deliveries, and an action that saves a draft message. There is no tool that sends.
• Where the person approves. A person reads each draft update and sends it, corrects it or throws it away. An email is also outside content, and text inside it can be written to look like an instruction, which is called prompt injection. Limiting the loop to reading and drafting limits what a trick can achieve, and a person reads every draft.

Overdue-invoice reminders drafted from the ledger. The dividing line matters here. A reminder that goes out on a set day after the due date is a fixed workflow, covered below. A loop earns its place when each reminder depends on what the lookups turn up.

• Steps. The loop looks up who owes what, then checks each customer's history: whether a payment arrived but was not matched, whether a dispute or a payment arrangement is on file, and whether the job is finished. It drafts a reminder that suits that case and holds them all for approval.
• Tools. A read-only lookup of the ledger, a lookup of customer notes and job status, and an action that saves a draft.
• Where the person approves. A person reads every reminder before it goes. A reminder to a customer who has already paid, or is in a dispute, can do real damage, so nothing is sent until a person says so.

Enquiries that need a stock or availability lookup. Filing an enquiry is a fixed workflow. The reply is where cases can differ: a customer may ask about several items, want a substitute or ask which dates are free, and what to check next depends on the last answer. If every enquiry needs the same single lookup, a fixed workflow will do.

• Steps. The loop reads the enquiry, works out what is being asked, checks stock or availability, and drafts a reply that says what it found. If it finds nothing, the draft says so instead of guessing.
• Tools. A stock or availability lookup, a read of the price list or the calendar, and an action that saves a draft reply.
• Where the person approves. A person reads the draft before it goes, and looks hardest at anything that promises a date or quotes a price, because the customer will hold you to it.

Which trade and field jobs suit it?

Trades run on moving parts: deliveries, crews, weather and sites. Two example jobs suit a loop, described here in general terms for an Australian tiler, electrician, plumber or concreter. They show the pattern. They are not accounts of any one business.

A late delivery reshuffles the week. The supplier says the materials for Thursday's job will arrive a day late. That one fact touches the job waiting for them, the crew booked for it and whatever was meant to follow.

• Steps. The loop looks up the delivery, finds every job that needs it and checks each job's dates and crew. It looks for a job that already has its materials and could move into the gap, then proposes a revised schedule for the week, with a short message for each crew and each customer affected.
• Tools. Read access to the schedule, the orders and deliveries, and crew availability, plus an action that saves a proposed schedule as a draft. It has no tool that edits the live calendar.
• Where the person approves. The owner or supervisor reads the proposal and decides. Nothing on the real calendar changes until they accept it, and the customer messages stay drafts. The proposal is only as good as the records: the AI does not know a crew member is away unless somebody has recorded it, so a person checks it against what they know.

Assembling the paperwork a job needs. Before a job starts, the documents it needs can be scattered: the accepted quote in the quoting system, site details in an email, insurance and licence documents in a folder.

• Steps. The loop takes the job, checks the list of what your jobs need, searches each place for each item and assembles what it finds into a job pack. Where it cannot find something, it lists it as missing instead of filling the gap.
• Tools. Read access to the quote records, the shared folders and the email, plus an action that saves a draft pack and a list of what is missing.
• Where the person approves. A person checks the pack against the originals before anything is relied on or sent on, especially dates and licence or certificate numbers, and chases whatever is missing. The list of gaps matters as much as the pack.

What should stay a fixed workflow instead?

Any job that follows the same steps every time. Three examples are invoice data entry, where the fields are pulled from a supplier invoice into a draft entry, filing enquiries, where each one is sorted by type and its details are pulled out, and reminders that go out on a schedule. Nothing in them needs the AI to choose what to do next.

These are AI workflows: the steps are set in advance, and the AI does specific jobs inside them, such as reading an invoice or sorting an enquiry. A workflow is cheaper to run, more predictable and easier to test than a loop. It runs the same steps in the same order each time, so you can test it against old examples, and when it fails you know which step failed.

Our Insights post on five AI workflows to consider before chatbots covers jobs of this kind, including invoice extraction, supplier price list changes, follow-ups, meeting notes and document classification. It describes them as workflows with review steps built in, and its advice is the same in spirit: start with the repeatable back-office jobs.

The two work together. A fixed workflow can file the routine supplier emails and hand only those that change a delivery or an order to a loop.

Where must a person still decide?

Some decisions should stay with a person however well the AI drafts. Four kinds are worth naming.

• Money going out. Paying an invoice, approving a refund or placing an order. A wrong payment is hard to get back, and a false invoice can look like a real one.
• Messages to customers. A message carries your name, cannot be unsent and can promise something you cannot deliver.
• Anything that cannot be undone. Deleting a record, cancelling a booking or changing a live schedule. If undoing it would be hard, a person decides first.
• Decisions about a person. Hiring, rostering, pay, discipline, or whether a customer gets credit. The AI can gather the facts. A person weighs them and answers for the decision.

How you build the approval matters. The simplest way is to give the loop no tool that does the risky thing. A loop that can only save a draft cannot send it, and one with no payment tool cannot pay. Where a tool does have to act, make it pause and wait for a person to say yes. Our guide to human in the loop covers where a person must stay in the process.

Start with a person approving every result. Loosen that only when you have evidence the loop gets a particular job right, and not for the four kinds above. These jobs also touch personal information about customers and staff, so check where the tool sends it before you connect anything. Our post on where your business data goes when you use AI explains what to ask.

What does it cost to run?

AI model providers usually price usage per token, the chunk of text a model reads or writes, which is often part of a word. You pay for the text that goes in and the text that comes out.

A loop makes many calls, and each call sends the conversation so far again: the instructions, the email and every result from every earlier step. So a loop that takes many steps costs more than one that takes a few, and one that reads long documents costs more again, because those documents are sent again at each step. No single figure fits every business, because the bill depends on the model you use, how much text is involved and how many steps the loop takes.

Limits keep the cost in bounds. A step limit stops the loop after a set number of trips, a token budget stops it once it has used its allowance, and a time limit stops it when it has run too long. Our guide to stop conditions and budgets shows how to set each one. It is written for people who build software, so hand it to whoever sets the loop up for you. If a loop keeps hitting its limit, treat that as a sign: the job may not suit a loop, or the instructions may need work.

Prices change, so check your provider's current per-token prices.

A practical way to begin: pick one office job from above, give the loop only tools that read and draft, set a step limit and a budget, and have a person approve every result. Run it alongside the way you do the job now, and compare. If every case turns out to follow the same steps, build it as a fixed workflow instead.