Agenticum · Thoughts · 9 min read

AI Agents: Practical Business Examples, and What They Have in Common

An assistant answers. An agent acts: it opens the ticket, reads the contract, moves the money or files the fix, and comes back when it is done or stuck. Here are eight examples companies have put in writing, the numbers they claim, and the pattern underneath all of them.

Nikolay Zuykov · Agenticum · 27 October 2026

Most "AI agent examples" articles list job titles: sales agent, HR agent, finance agent. That tells you nothing about what the software does on a Tuesday afternoon. So this piece does the opposite. Each example names the company that published it, the task, the figure they reported, and the point where a person still decides. All figures are the vendors' or customers' own claims as published on their pages on 27 October 2026; we have not audited them.

Eight agents doing identifiable work

1. Resolving support tickets

Zendesk reports building custom agents on Claude and reaching one million agent executions in seven weeks. The task is familiar to anyone who has run a help desk: read the ticket, check the order or account, answer or route. The person stays at the escalation point. Microsoft describes the same shape for its Copilot Studio agents built on the "Customer Conversation" harness: a primary agent works with specialised sub-agents, and "if the agent can't resolve a request, it can escalate the conversation to a human representative through a defined path".

2. Reviewing contracts at volume

Spellbook, a legal software company, says it runs 530,000 contract reviews a month with Claude. The agent reads the draft, flags clauses against a playbook and proposes edits. A lawyer still signs. The number matters because it shows the task is repetitive enough to be measured, which is the first sign an agent will pay for itself.

3. Drafting long documents from case files

EvenUp, which prepares demand letters for personal injury claims, says drafting went from 15 hours to 15 minutes with Claude. The agent assembles the document from the file; the professional checks it against the evidence. Note what the claim does not say: it does not say the 15 minutes include the review, and you should assume they do not.

4. Doing the financial admin of a small business

Qonto, a business banking provider, describes delegating financial admin for small businesses to Claude on Amazon Bedrock. This is the example closest to many companies in Dubai and Riyadh: invoices, receipts, categorisation, reminders. The money movement itself stays behind a confirmation, because that is where a wrong action costs real cash.

5. Managing ad campaigns from a conversation

Supermetrics says it lets marketers manage ad campaigns from a conversation with Claude. The marketer asks for a change in plain language; the agent makes it in the ad platform. The guardrail here is budget: an agent that can raise spend needs a ceiling it cannot cross without a person.

6. Turning alerts into fixes

Carvana, the online car retailer, says it turns Slack alerts into production fixes with Claude. An alert fires, the agent investigates the code, proposes a change, and an engineer reviews and merges. This is the pattern behind most "coding agent" stories: the agent does the first hour of a repair, not the last five minutes.

7. Research and analysis inside the office suite

Microsoft ships pre-built agents it calls Researcher, Analyst and Facilitator inside the Copilot plans, plus Sales, Service and Finance agents, and lets companies build their own in Copilot Studio with "built-in testing and human-in-the-loop controls". The useful detail in Microsoft's documentation is governance: admins get an agent inventory with role-based access and cost management, and can see which permissions and data an agent needs before allowing it.

8. An agent for one person's desk

OpenAI's Pro and Business Premium plans now include "Dots, your always-on agent", and Anthropic's Team plan includes Claude Cowork alongside Claude Code. These are agents for a single professional: they watch a folder, prepare a brief, chase a thread. Smaller stakes, same rule: the person reviews before anything leaves the desk.

What all eight have in common

Strip the company names and the same four parts appear every time.

PartWhat it means in practiceWhat goes wrong without it
A task with a verifiable resultA ticket closed, a clause flagged, a transaction categorised, a fix merged. Someone can check it in the system of record."Improve our customer experience" cannot be checked, so the agent cannot be measured and will be judged on vibes.
Tools and permissionsThe agent reads and writes through the same accounts and rights a person would have, and no more.An agent that can see everything answers with things it should not, as Microsoft's own Copilot documentation warns about over-shared folders.
A harnessThe loop that plans, calls tools, checks results and stops. We described it in Inside the Agent Harness.A raw model with tool access behaves like a new hire with admin rights and no manager.
A human at the costly stepMoney leaves, a customer is told something, code goes live: a person confirms.The cases in When customers send AI agents: a booking moved on the wrong day, a confirmation that never arrived.

The vendors have converged on this shape. Microsoft calls the engine a "harness" and offers three of them with different levels of autonomy; Anthropic sells the agent-building kit and the always-on desk agent as separate products; the customer stories above all describe a review step. When every supplier builds the same guardrail, treat it as a requirement, not a feature.

How to pick a first agent

We would start with the task that scores highest on these four questions, in this order.

  1. Can the result be checked in a system you already have? Help desk, accounting software, contract repository, ticket tracker. If the answer lives only in someone's head, pick another task.
  2. Is the task boring and frequent? Fifty times a week beats five times a month. The Spellbook number, half a million reviews a month, is the extreme version of this.
  3. What is the worst single mistake? A mis-categorised receipt is cheap to fix. A wrong price sent to a customer is not. Start where the worst case is cheap, and put the person in front of the expensive actions.
  4. Does the data allow it? If the task touches customer or employee personal data, the agent runs only in a tool the company has contracted, in a known location, under the rules in our one-page AI data policy. Both the UAE and Saudi Arabia now enforce data protection laws with penalties.

Then measure one thing: cost per correctly completed task, with the human review time included. An agent that closes tickets fast but sends ten percent of them back to a person for rework may be more expensive than the queue it replaced. The companies above publish their wins. Keep your own count of the misses, because nobody else will.

What to expect in the first month

The first two weeks are spent on access, not intelligence: the agent cannot read the folder, the API key has the wrong scope, the approval step is in email instead of the tool. Week three is when the task definition gets rewritten, because the agent exposed how vague it was. Week four is the first honest number. If that number is good, the second agent takes a fifth of the time, because the harness, the permissions and the review habit already exist.

Sources checked on 27 October 2026. Customer figures are as published by the vendors and their customers and have not been independently verified. The selection and the recommendations are our analysis at Agenticum.