An AI agent is a program that uses an AI model to carry out a task in several steps. Unlike a fixed workflow, nobody lays down the steps in advance: the model decides what happens next and which tool to use for it. Often the best solution is still a fixed workflow in which only one or two steps use an AI model.
What is an AI agent, and how does it differ from a chatbot or a workflow?
A chatbot answers, a workflow follows fixed steps, and an agent chooses its own steps.
More detail
A chatbot holds a conversation and answers questions. A workflow carries out steps that a person laid down in advance, for example: read the receipt, enter the amount, file the document. An agent gets a goal and the tools for it. It decides for itself which steps are needed and in what order, and checks along the way whether it is getting anywhere.
The distinction between workflow and agent comes from the article Building effective agents by Anthropic (December 2024): workflows run along predefined paths, while agents direct their own process and their own use of tools.
What can an AI agent do for my business?
It takes on tasks that have several steps and need some judgement: reading, sorting, checking and summarising information.
More detail
Typical areas:
- Customer service: sort incoming requests, prepare replies and answer simple routine questions.
- Sales and marketing: qualify leads, enrich contact data, prepare messages and analyse campaigns.
- IT and engineering: handle simple support requests, monitor systems and help with debugging or documentation.
- Finance and operations: read invoices and receipts, extract data from unstructured documents, flag irregularities and prepare reconciliations.
- In a small team: summarise emails, draft texts, manage calendars and turn data from PDFs or spreadsheets into a clean format.
What these tasks have in common is that they come up often, have several steps and need judgement, not just copying.
How do agents work? The five patterns
They follow five common patterns. Usually one is enough, and often a fixed workflow with one AI step does the job.
More detail
Anthropic describes five patterns for building processes with AI models. Strictly speaking they are patterns for workflows, where the path is laid down in advance. An agent in the narrower sense goes beyond them, because the model steers its own process.
- Step by step (prompt chaining): a task is split into fixed sub-steps, and each step works with the result of the one before. A check can sit between the steps.
- Sort and route (routing): the model classifies an input and sends it to the workflow or tool that suits it best.
- Work in parallel (parallelisation): several parts run at the same time, or the same task runs several times and the results are compared to make the outcome more reliable.
- Split and delegate (orchestrator-workers): one model breaks the task down when the parts cannot be predicted in advance, and hands them to other models. Then it combines the results.
- Check and improve (evaluator-optimiser): one model produces a result, a second evaluates it and gives feedback. This repeats until the result is right.
I start with the simplest solution and add effort only when the case calls for it.
Where does the agent decide, and where does a person?
You decide that. The greater the damage a mistake could do, the more should stay with a person.
More detail
An agent can read, check and report without changing anything. It can also act: create tasks, change records, send messages. The further its rights reach, the more it needs checks and a person who approves.
With my follow-up machine the limit is drawn tightly: it creates tasks and sends me summaries. It writes to no customer and changes nothing in the CRM. Whether and how I follow up is my decision.
Do I need an agent, or is a fixed workflow enough?
A fixed workflow is enough when every case looks the same. An agent is worth it where the steps cannot be laid down in advance.
More detail
Even in my own system most of it is a workflow with fixed rules: deadlines, checking the last activity, tasks. The AI model only adds two notes. It decides nothing.
A workflow without an agent is often enough. I will tell you when that is the case. How I approach workflows is on the AI automation page.
What does an agent cost, and does it pay for itself?
The build starts with a Build Sprint at a fixed price of €4,900. Whether that pays for itself, I work out with you on your task: the hours it costs today against the cost of building and running the agent.
More detail
An agent has to pay its way. So I start with the task and not with the technology: how often it comes up, who does it today and what a mistake costs. If it does not pay, I will tell you so.
Where does my data go?
An agent reads your data and sends what it should assess to an AI model. Which provider and which data depends on the build, and is settled before it starts.
More detail
In my own system that is the last three emails of a deal, and they go to the AI provider Mistral. The system runs on my own server. Its data sits in a database there, the tasks in the task tool and the summaries in my mailbox.
Which data a model may see in your case, and which it may not, is part of the AI Fitness Check.
My View of the AI Hype
There is a lot of hype about AI right now, and almost every business is trying to work out how to use it well. I am no different: I am still learning where it helps. I take a pragmatic approach and do not claim to be an expert in everything new. I try things out first, and I offer clients only what I feel comfortable with. My experience helps me judge which technique fits which need, and also when a task needs no AI.
Three Examples From My Work
-
Check step in a data import
An AI step spots data that has changed unexpectedly. The import itself is a fixed workflow, and only this check uses an AI model.
-
Bookkeeping receipt processing
An AI model helps decide which tool processes a receipt.
-
Follow-up machine
It watches open quotes in the CRM. If there is no activity after 5, 12 or 21 working days, it creates a task each time. Several times a week it emails me a summary. An AI model adds two short notes: where the deal stands and which angle might fit next. It writes to no customer and changes nothing in the CRM.
When an Agent Is Not the Right Answer
I will advise against an agent if one of these is true:
- Every case looks the same, and a fixed workflow handles it.
- A wrong result would be expensive, and nobody checks it before it takes effect.
- Your systems do not allow the access the agent would need.
In that case I will say so in the first conversation, and you save yourself the Fitness Check.
The First Step: the AI Fitness Check
The AI Fitness Check costs €890 fixed. First you go through one task in a session with me, on site or remote. The written assessment arrives within 5 working days of that session. It contains:
- the task, written down step by step
- for each step: what an agent may do alone and where a person decides
- which of your systems the agent would have to read or write, and whether they allow it
- the hours the task costs today against the cost of building and running an agent
- a recommendation, including when a fixed workflow is enough
For that I need the person who does the task today, real example cases and read access or documentation for the systems involved.
If an agent is worth building, the report describes the scope of the Build Sprint. Fixed price €4,900: your agent runs within two weeks of the start. Book the next step within 30 days and the €890 is credited.
And If It Is Only About Answers to Customer Questions?
Then a chatbot is often the shorter way. How I build one and what it can do is on the AI chatbot page.
Common Questions
What is the difference between AI automation and an AI agent?
Automation carries out steps that were fixed in advance. An agent chooses its own steps. Many solutions combine both: a fixed workflow in which one step uses an AI model.
Can an agent write to a customer?
Yes, if you give it the right to. My own system writes to no customer. It creates tasks and sends summaries to me. What goes to a customer is written and sent by a person.
What happens when an agent fails?
With my own system I get an email when something fails, and every run is logged. For an agent built for you, you and I agree who receives a failure message and how fast someone has to react.
Which Task Should an Agent Take Off Your Hands?
Tell me briefly about the task. The first conversation takes 30 minutes and costs nothing, and I will say honestly whether an agent is the right way to deal with it.