Written by
Nacho Seoane
Published on
Share
Automation fits when you can define the steps in advance. An AI agent can help when the next step depends on what it discovers along the way. There is also an option between the two: a process with clear rules that uses AI for one specific task, such as extracting information from an enquiry.
Consider a sales enquiry. Saving the contact, assigning an owner and creating a reminder can follow rules. Understanding an unclear message requires interpretation. Finding out what is missing, consulting different sources and preparing the next action involves a different kind of work.
TL;DR — AI agent or automation
- Automation follows a defined path; an agent can choose how to proceed. The distinction concerns who chooses the steps and tools. A chat interface alone tells you little about how the work happens.
- Compare 3 options: rules, an AI step and an agent. Extracting information from a message can happen within a fixed workflow, without adding autonomy to the rest of the process.
- Start with 1 complete, recurring task. Preparing a sales enquiry may involve recording it, extracting details, checking history, summarising and reviewing: the 5 steps in this guide's example.
- Permissions are part of the scope. Reading information, drafting a response and sending a message are different actions. Define which are allowed and which need a person's approval.
- Measure 4 things: outcome, time, errors and cost. Include human review and check what was saved in the destination system. A convincing response does not prove the task is complete.
- Our recommendation: test the simplest approach that completes the task. Add an agent when the work requires choosing steps based on context and you can verify that this flexibility helps.
What changes between automation and an agent?
The deciding factor is who determines the path through the work. With automation, you define it. With an agent, the model can choose some of the steps and tools within the boundaries you set.
This follows Anthropic's distinction between predefined workflows and agents that decide dynamically how to proceed. Its recommendation is to start with the simplest solution that addresses the problem. Source: Building effective agents.
| Situation | First approach we would test | What needs to be defined |
|---|---|---|
| Copy a form submission into the CRM and assign an owner | Rules-based automation | Fields, duplicate handling and assignment rules |
| Extract the company, requirement and date from a message | A specific AI step | Output format and how to handle missing information |
| Consult different sources depending on the request | An agent with a limited set of tools | Permitted sources, actions and a stopping condition |
| Approve an exception to commercial terms | Human review | The information needed to make the decision |
These are starting points for a decision. One process can combine all four situations, and interpreting text with AI does not by itself require an autonomous agent.
An example: preparing a sales enquiry
The useful outcome is that someone can handle the enquiry with the context already prepared. That means an organised request, the relevant background and a clear view of what is still missing.
This is an illustrative example, not a client result.
Someone writes: “We want to redesign the website for our two brands and connect it to the system our sales team uses.” The email does not say which system they use, what content already exists or who is involved in the decision.
An initial workflow could be:
- Save the enquiry and link it to the right contact.
- Extract the requirements that the message actually states.
- Check authorised history if there is an existing relationship with the company.
- Prepare a summary of what is known and the unanswered questions.
- Leave the response and next action for the owner to review.
If you always need to check the same history and produce the same summary, a fixed workflow with an AI step may be enough. If the information you find determines which source or tool to consult next, a more flexible approach is worth testing.
The foundation remains a website connected to your sales process: useful data, a clear destination and someone responsible for the next step.
What to check before adding autonomy
An agent needs a clear scope, accessible information and a way to verify its actions. Resolve any missing pieces before expanding its permissions.
These are the questions we would use to define the scope:
- Input: exactly what does it receive, and from where? A free-text message presents a different problem from a form with defined fields.
- Context: where is the authoritative information? Decide which document takes precedence if two sources disagree.
- Actions: can it read, prepare a draft or change a record? Every permission should meet a specific need.
- Exceptions: what happens when information is missing, a tool fails or the case falls outside the agreed scope?
- Outcome: how will you know that the task finished correctly?
OWASP recommends limiting the functions and permissions available and requiring human approval for high-impact actions. That approval must control execution itself; an instruction written in a prompt is insufficient. Source: Excessive Agency.
How to run a test that helps you decide
Test a bounded task and compare the complete outcome with the current way of working. Start by collecting representative examples: routine cases, incomplete requests and cases that should be handed to a person.
For each example, write down what a correct result would look like. For the sales enquiry, that might mean a record without duplicates, a faithful summary, unanswered questions and an assigned task. Good writing alone is insufficient.
Anthropic distinguishes the agent's conversation from the final state it leaves in the system. In this example, saying “I created the task” does not prove that it exists in the CRM. You need to check the destination. Source: Demystifying evals for AI agents.
During the test, record:
- The person's working time, including review and corrections.
- Cases completed correctly and cases requiring intervention.
- Errors that cause rework or affect another system.
- Tool costs and model usage costs.
- What happens when a request is repeated or an integration is interrupted.
Our approach is to expand the scope when the whole process shows a verifiable improvement. A demonstration that works once can help you explore; a repeatable test gives you a basis for deciding.
Frequently asked questions
These questions help define the scope before choosing tools or requesting a quote. The decision depends on the work, its permissions and how you will check the outcome.
Do I need to replace my CRM or existing tools?
You can keep your tools if their integrations support the intended work. First check which data and actions they expose, what permissions they require and whether those features are included in your plan. Sometimes your existing setup is enough. Sometimes it needs one missing connection rather than a new platform.
Is a chatbot the same as an agent?
Chat describes how you interact; an agent describes how work is carried out. Behind a chat interface there may be predefined responses, a document search or an agent that uses tools. Ask what information it consults, which actions it performs and what result you can verify outside the conversation.
What should I include when calculating the cost?
Include development, integrations, usage, supervision and maintenance. Ask for a quote that separates these items and explains which costs change with usage. A meaningful estimate needs the expected volume, the tools involved and the level of autonomy; otherwise, part of the work may be left out of the price.
Can an agent work without approval at every step?
Yes, within a defined scope and set of permissions. You can allow information retrieval or drafting while requiring review before sending messages or changing sensitive data. Enforce that distinction in the tools and test it, including a request that attempts to bypass the boundary. Approval should be a working control.
How do I know whether ordinary automation is enough?
Write out the workflow and check whether rules can determine every step. If the same type of input always leads to the same actions, start there. Use AI for the part that needs interpretation, and test an agent only if the work also requires choosing how to continue.
Final recommendation
Start with rules and add autonomy where the work needs it. Our choice between the 3 approaches would be automation for predictable steps, an AI step for interpreting input, and an agent for investigating and choosing a path. All three need a verifiable outcome, an owner and a way to handle exceptions.
If the first problem is receiving and assigning enquiries, our guide to website forms connected to a CRM helps establish that foundation before adding AI.
In summary
- Compare 3 approaches before buying a solution: rules, an AI task and an agent. Each handles a different level of variation in the work.
- Define 1 observable outcome for each test: for example, an enquiry recorded and assigned. This lets you check the full task and identify where the workflow fails.
- Separate input, context, actions, exceptions and outcome. These 5 decisions turn an agent idea into a scope you can price and evaluate.
- Test 3 groups of cases: routine, incomplete and out of scope. A demonstration with easy inputs does not show how the system handles everyday work.
- Record 4 decision measures: outcome quality, working time, errors and cost. Include corrections and supervision so that apparent savings do not hide extra work.
- Distinguish 3 permission levels: reading, preparing and executing. Approval for consequential actions must control the tool that acts, as OWASP recommends.
- Choose 1 task before expanding the system. Anthropic recommends starting with simple solutions; we expand autonomy after verifying that it improves the complete workflow.
Sources and references
These sources support the technical distinction, permission boundaries and evaluation approach. The sales example is illustrative; the recommendations on scope reflect how we approach projects. Sources consulted for the original article on 30 September 2026.
- Anthropic: Building effective agents. Predefined workflows and agents; published in December 2024 and used here for that conceptual distinction.
- OWASP: Excessive Agency. Permissions, available functions and approval for consequential actions.
- Anthropic: Demystifying evals for AI agents. Checking the outcome an agent leaves in its environment.
Where to start in your business
Choose a recurring task with an owner and an outcome that is easy to verify. Document how it works today and identify the step that requires interpretation. That gives you a basis for deciding whether to connect tools, add a specific AI task or develop an agent.
At Apogeo, we develop AI agents for businesses around that concrete work. To explore a use case, tell us what your team receives, what happens next and where things get stuck. We can use that context to define a solution and a useful test.
