Short answer: a working agent scenario looks the same in every industry — many similar cases with small differences, each requiring you to look in several places, where the sequence of steps depends on what turns up. Below are twelve such processes.
How the scenarios were filtered
Three filters that eliminate most ideas before they reach an estimate.
Frequency. Fewer than ten runs a month and an agent will not pay for itself. Development and testing cost the same whether the process runs ten times or a thousand.
Variation. If the steps are always identical, you need a Power Automate flow: cheaper, more predictable, easier to debug. An agent only wins where the sequence depends on the data.
Data availability. If the decision needs something nobody records — a lost-deal reason, a phone agreement — no agent will retrieve it.
Autonomy levels below refer to the 0–4 scale: 0 is read-only, 2 creates an unposted document, 4 executes without notification.
Finance: four scenarios
1. Preparing collections work · level 1
The agent finds overdue balances against your thresholds, checks each customer for an open dispute, cross-references open orders, groups documents by customer and drafts the outreach. It flags customers where shipment is worth holding.
Requires: payment terms filled in, payments applied to documents, disputed amounts marked. Detail: AI for accounts receivable.
2. Preparing a payment schedule · level 1
Collects invoices due, isolates those where the early-payment discount beats your cost of money, matches against expected receipts based on actual customer behaviour, and produces a proposal for approval.
Requires: vendor payment terms filled in, current document dates.
3. Reconciliation exception handling · level 1
Where automatic reconciliation did not match, the agent looks for the likely cause: an unapplied payment, a partial payment, a bank fee, an exchange rate date difference. It prepares an explanation per line.
Requires: bank integration configured.
4. Period close commentary · level 1
Drafts the narrative section: what changed against last period and against budget, and what drove it. A person checks and adds what is not in the data.
Requires: a stable chart of accounts and consistent dimensions. Context: period close.
Sales: three scenarios
5. Processing inbound orders from email · level 2
Parses the message, matches lines to items, checks prices against that customer's price list and availability in stock, and creates an order in draft. Ambiguous lines go into a separate list rather than being guessed.
Requires: tidy item data and configured price lists. The most common obstacle is that customers use their own names for items, which needs a mapping table the agent can extend.
6. Reacting to customers who have stopped buying · level 1
Computes each customer's typical ordering interval, finds those who have gone beyond their own norm, estimates the revenue at risk and prepares a call list with an explanation of why each name is on it.
Requires: at least a year of history.
7. Discount and credit term control · level 1
Checks orders before release: is the discount within policy, is the credit limit breached, is there overdue debt. Flags what needs a decision.
Requires: a discount policy that exists in writing. If it does not, that is the first thing to produce.
Purchasing and inventory: three scenarios
8. Reacting to supply delays · level 1
Notices a receipt slipping, checks which customer orders are exposed, prepares a supplier query and a list for the sales team of customers who need warning.
Requires: current expected receipt dates on purchase orders.
9. Preparing purchase orders · level 2
Takes planning worksheet suggestions, checks them against actual rather than configured supplier lead times, flags those resting on a one-off demand spike, and creates draft orders — with questionable lines separated out.
Requires: maintained item cards. If reorder points are stale, the agent inherits that. See reorder point and safety stock.
10. Working through dead stock · level 0
Finds items with no movement, quantifies the capital tied up, ranks by money rather than count, and answers follow-ups: is it obsolete, seasonal, or tied to a contract.
Requires: a correct costing method and regular cost adjustment.
Master data: two scenarios
11. Master data maintenance · level 3–4
Completing fields on counterparty records, normalising names, finding duplicates. One of the few scenarios where a high autonomy level is justified: the rule can be written down completely, and a mistake is both visible and reversible.
12. Preparing data for migration or audit · level 0
Finding records that break the rules: empty mandatory fields, invalid codes, contradictory values. A one-off task, but one where an agent saves days of manual checking.
What looks appealing and does not work
- "The agent runs the accounting." Accounting is a sequence of decisions with accountability attached, not a set of operations. Individual steps automate; the function does not.
- "The agent answers customers." It manages routine situations; in a conflict the reputational cost of a mistake outweighs the saving.
- "The agent decides who gets credit terms." A commercial decision with direct financial exposure. The agent prepares the data; a person decides.
- "The agent will find what to improve." It works toward a stated goal. Without one it produces generalities you already knew.
How to choose the first one
Take two or three scenarios from the list that genuinely occur in your company and, for each, count how often it runs per month, how long it takes now, and whether the rules exist in writing.
The first to implement is the one with the best ratio of frequency to unknowns — not the one that sounds most impressive. In practice that is almost always scenario 1 or 2: narrow, well-defined, with a visible result within weeks.
Broader context in the complete guide to AI agents; what this looks like as a service on the AI agents page.
Frequently asked questions
How many scenarios can we implement at once?
Start with one. The second and later ones cost significantly less because the integration layer, authentication and logging already exist. Attempting three at once stretches testing and blurs the success criterion.
Why are most scenarios at autonomy level 1 or 2?
Because the time saving lives in gathering and preparing data rather than in the final click. Moving from zero to prepared work delivers most of the value at the lowest risk.
What if our process is not on this list?
Run it through the three filters: more than ten times a month, sequence depends on the data, and the data is recorded in the system. If all three hold, the scenario is workable even if it is industry-specific.
Can we start at level 3 or 4?
Only for operations where the rule can be written down completely without qualifiers — master data normalisation, for example. Otherwise the level is raised later based on how much prepared work people reject.
What does one scenario cost?
An agent from $6,000, a process-specific one from $10,000, after a discovery from $750. Most of the cost is not development but testing what the agent should refuse to do.