AI Agents
It does more than answer a question. It carries the work through from start to finish and brings the result to you for approval.
What exactly is an AI agent?
Anthropic, the developer of the Claude models, defines agents as systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
Workflows sit on the other side. In a workflow, the model and tools follow predefined steps. Anthropic's advice is clear: find the simplest solution possible, and only increase complexity when needed. Often that can mean building no agent at all.
According to the agent-building guide OpenAI published in 2025, an agent has three core components: the model (the AI that does the reasoning), tools (the systems the agent can use) and instructions (the rules that define how the agent behaves).
How an agent is set up
To trust an agent, the rules around it matter as much as the model.
Data tools
The sources the agent gathers information from. CRM records, spreadsheets, PDF documents, web search.
Action tools
The operations through which the agent actually changes something. Drafting an email, updating a record, handing a conversation to a person.
Instructions
The task definition, steps, tone and hard limits. Well-written instructions are the single biggest factor in how an agent behaves.
Guardrails
Layered controls that block off-topic requests, personal data leaks and risky operations.
Handoff to a person
Control passes to a person when the agent fails repeatedly, or before an irreversible, high-risk action.
Logging and review
Every step the agent takes is logged and reviewed at regular intervals, and the instructions are improved.
Not every job needs an agent. Five common patterns
Anthropic notes that most systems that succeed in businesses are built with simple, composable patterns instead of complex frameworks. Here are those patterns and how they look in a business.
| Pattern | What does it do? | Business example |
|---|---|---|
| Prompt chaining | Splits a task into sequential steps, with each step using the output of the one before | Read the request, identify missing information, draft a quote |
| Routing | Classifies incoming work by type and sends it to the right flow | Sorting incoming email into sales, support or accounting flows |
| Parallelization | Runs parts of a task at the same time, or gets several views on the same task | Checking a contract from different angles at once |
| Orchestrator-workers | One model breaks the work into subtasks and combines the results | Compiling market research from many sources |
| Evaluator-optimizer | One model produces, another evaluates against criteria and asks for fixes | Reviewing quote text against company style rules |
What can go wrong in agent projects?
Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027. It also warns about a market trend it calls "agent washing", where existing assistants, RPA products and chatbots are rebranded as "agents". According to Gartner, only about 130 of the thousands of vendors promoting themselves in this space are truly agentic.
- Start with a narrow task. "Draft a quote from an incoming request" is a better starting point than "manage sales".
- Start with one agent. Move to setups where several agents work together only when you really need them.
- Limit permissions. The agent should access only the tools its task requires.
- Approve everything that goes out. Customer emails, payments and record deletions should require human approval.
- Track costs. Agents call the model at every step, so cost per task should be monitored.
- If a simple solution works, skip the agent. For a task with clear rules, process automation is cheaper and more reliable.
What AI Agents gives you
Multi-step work
It takes on chains of tasks such as gathering information, comparing options and preparing drafts.
You approve
Important decisions and anything sent outside the business go through your approval first.
Works with your tools
It gets work done using tools such as email, calendars, spreadsheets and your CRM.
Setup steps
Task definition
We set out clearly the work the agent will take on and its limits.
Permissions
We decide which tools it can access and where it must ask for approval.
Setup and testing
We test the agent with real examples and adjust its behaviour.
Monitoring
We keep a record of what it does and review it regularly.
Industries it suits
Works well alongside
Process Automation
Filling in the same spreadsheet, writing the same email, searching for the same document every day. Let this work get done without waiting for you.
Learn more ›We plan where to startAI Consulting
There are many places to use AI, but you cannot start on all of them at once. First we work out where you should begin.
Learn more ›Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions. An AI agent plans its own steps towards a goal and uses tools to do the work, for example gathering information, preparing a draft or creating a record.
Can an agent get out of control?
We limit the tools the agent can access and the actions it can take. Every action that goes outside the business needs approval first.
What kind of work is it suited to?
Rule-based work with several steps, such as preparing quotes and reports, research, customer follow-up and sorting incoming requests.
Tell us about your business. We'll plan it together.
In a free 30-minute discovery call, we talk through your processes and work out where AI would save you the most time.