Our glossary defines AI agents and agentic AI in plain words. This guide is about use: what has been shown publicly, what suits an agent and what does not.
The examples come from public pages, and none is ours. We name no client and claim no result.
What do public sources name as agent uses?
Anthropic's engineering post Building effective agents (19 December 2024) separates workflows, where models and tools are orchestrated through predefined code paths, from agents, where the model dynamically directs its own process and tool use. It names two applications as particularly promising for agents: customer support and coding agents.
Zapier's AI page (read 12 November 2026) describes agents as AI that takes action across your tools, not only responding to prompts, with examples such as qualifying leads, routing tickets and handling requests, running in the background. Its template library lists patterns including 'Automated Email Reply Drafts Agent' and 'Resolve support tickets with guarded AI orchestration'.
Example 1: a customer support agent
The pattern: the agent reads a request, looks up the customer's order or booking in your system, finds the relevant policy text and drafts or sends an answer, and passes hard cases to a person. It suits support because the work is conversational and the agent needs access to tools and records. It fails when the policy text is out of date, when it is allowed to promise things you have not approved, or when a customer's wording steers it somewhere it should not go. Our guide to AI customer service automation covers the controls.
Example 2: a coding agent
The pattern: given a task and access to a code repository and a test suite, the agent reads files, proposes changes, runs the tests and revises. Tests give it a clear signal of success, which in our reading is one reason the work suits agents. It fails when tests are weak, when the task is vague, or when it changes more than anyone reviews. A developer should read every change before it is accepted, as with a colleague's work.
Example 3: a lead qualification agent
The pattern: a new inquiry arrives, the agent reads it, checks it against your criteria and public company information, scores it and routes it to the right person with a short note. It saves sorting time. It fails when the criteria are vague, when it relies on thin or outdated public information, or when a good lead is scored low because it is written in an unusual way. Sample the scores each week.
Example 4: a ticket routing and triage agent
The pattern: the agent labels each ticket by topic and urgency, adds a summary and sends it to the right queue, as in the public templates named above. A person still replies, so the cost of a mistake is a delay, not a wrong answer to a customer. It is a sensible first agent, and often a workflow is enough. Our guide to AI workflow automation covers the choice.
What do agents do well?
- Open-ended tasks where the steps cannot be predicted in advance, which Anthropic says is where agents fit.
- Reading messy input such as free-text requests, long threads and mixed documents.
- Using several tools in sequence, such as looking up a record, checking a policy and drafting a reply.
- Routine research and preparation, such as gathering facts from several sources for a person to use.
Where do agents fail?
- Cost and speed. Anthropic says agentic systems often trade latency and cost for better task performance, so a simple task may not justify one.
- Being steered. OWASP's prompt injection page says inputs can alter a model's behavior in unintended ways, even when humans cannot see them, and that it is unclear whether there are fool-proof methods of prevention.
- Wrong statements. A model can say things that are not true. In February 2024 a British Columbia tribunal held Air Canada responsible for wrong information its chatbot gave a customer.
- Hard-to-test behavior. Because an agent chooses its own steps, runs differ, and testing is harder than for a fixed workflow.
- Overreach. An agent with broad access can do more harm than one limited to a narrow task.
Sources for these points: OWASP LLM01:2025, read 12 November 2026, and Torkin Manes' summary of Moffatt v. Air Canada, 7 March 2024. The FTC said in September 2024 that there is no AI exemption from the laws on the books, so claims about what your agent does must be true.
What controls reduce the risk?
- Least privilege. OWASP advises restricting the model's access privileges to the minimum necessary for its intended operations.
- Human approval. OWASP advises human-in-the-loop controls for privileged operations, such as sending money, deleting records or contacting customers.
- Narrow goals. Give the agent one clear goal and a short list of tools.
- Logs. Record what it read, what it decided and what it did.
- A stop switch and an owner who reviews a sample of its work every week.
Should you use an agent?
Start with a workflow, a fixed path with a model on the steps that need language, and move to an agent only when a fixed path will not do. That is Anthropic's advice too: find the simplest solution possible and increase complexity only when needed. If you do use one, start with a task where a mistake is cheap and a person reviews the result. Our AI agents page covers how we scope them, our AI automation page covers workflows, and our guide to AI automation examples lists public patterns.
Conclusion
Public sources point to customer support and coding as the strongest uses of agents, and to lead qualification and ticket routing as common ones. Agents cost more, are harder to test and can be steered by what they read, so give them narrow goals, minimal access and human approval for anything that matters.
PromotEdge International is an international strategy, technology, marketing and design company, and our delivery team is in India. This guide is not legal advice. See our AI agents page.
Facts we rely on: see our company facts for what we state about ourselves and where each fact comes from.