# Agent or automation? > A decision tree for one specific process: whether it is solved with a written rule, with an AI agent, with an agent that proposes and a person who signs off, or with neither yet. The interactive tool is at https://c2suite.com/en/resources/agent-or-automation. A person answers it: it does not look at any operations or connect to any system. **It does not return a score, on purpose.** Between an automation and an agent there is no more or less: they are two solutions to two different problems, so a score out of 100 would have to invent an order that does not exist. The result is a recommendation with its conditions and a first step. ## How it decides 1. **You choose an area** (sales, support, finance or operations), and with it the opening questions, which are specific to each one. 2. **Then comes the shared trunk**, the same for every area: whether a case calls for an agent or an automation does not depend on the department. 3. **The tree really branches.** Some questions are skipped depending on earlier answers: with a fixed format it asks about the rule, with free-form text it does not; with no prior review it skips the criteria question and the cost of a mistake decides. 4. **No path uses every question.** There are 25 in the catalog; the longest path has 13 and the shortest 6. 5. **Some answers switch on a condition** that is picked up at the end, even if it was many steps back: personal data, accountability, the cost of a mistake, response time, scattered data or an undocumented process. 6. **The cost of a mistake changes the recommendation.** A path that ends in “agent” but switched on a costly mistake becomes “agent proposes, a person signs off”. 7. **Budget and team size do not branch.** They are asked **after** the recommendation, they are optional and they do not change it: they say whether the result makes sense at that scale, which is a second reading. A small budget does not turn a language problem into a rules problem. 8. **No email is needed to see the result.** The email is asked later, in the contact form, and only if the person decides to continue. ## The four outcomes ### Automation, not an agent What you describe is solved with a workflow: a written rule that always does the same thing with the same input. A language model would add cost, waiting time and a source of errors the rule did not have. In exchange you would get flexibility, and in this case flexibility is a defect. ### Agent proposes, a person signs off The work needs a model: the input is language and the rule cannot be closed. But mistakes are expensive, so the model prepares and a person approves. That approval step is the only way this case reaches production without the first mistake ending the whole project. ### Agent The input is free-form language, the task repeats often enough, someone can review the output while it is tuned, and there are criteria to say whether one output is better than another. This is the case where a model adds something a rule cannot, and where you will be able to tell whether it worked. ### Not yet One of the conditions that make an AI project measurable is missing. The process may still be worth improving, but without repetition, review or criteria, nobody will be able to say whether it worked, and ending without knowing is what keeps a second project from ever happening. ## The 4 areas Each one has its own opening questions, and it also sets the concrete first step in the result. - **Sales** (4 questions of its own: the task, the trigger, where it comes from, who sees it). Prospecting, qualifying, follow-up and logging what gets discussed. - **Support and service** (4 questions of its own: the task, entry channel, how repetitive, who sees it). What a customer asks for after buying: tickets, replies and escalations. - **Finance and administration** (4 questions of its own: the task, the documents, accountability, what it produces). Invoicing, collections, reconciliation and anything someone might audit. - **Operations** (4 questions of its own: the task, where it comes from, when it's due, what it sets off). Scheduling, field work, fulfillment and the paperwork all of that creates. ## The shared trunk (9 questions) The questions that decide, in the order they can appear. - **Input format**: How does the input arrive? - **The rule**: Can the rule be written down in full, exceptions included? - **Same result**: Does the result have to be identical every time? - **Repetition**: How many times a month does it happen? - **Prior review**: Does anyone review the result today before it has consequences? - **Criteria**: Would two people on the team agree on which of two outputs is better? - **Cost of a mistake**: What happens if it gets something wrong and nobody notices? - **Personal data**: Does the input contain personal data? - **Explaining why**: Do you need to explain what the result was based on? ## The conditions it can switch on ### With personal data inside The input contains personal data, so before choosing a model you need to decide where that data can go. There are three real options: remove what identifies people before sending it, contract the model under a data processing agreement that covers it, or run it on your own infrastructure. In Mexico, the 2025 federal data protection law (LFPDPPP) also requires the privacy notice to say so. ### With a record of why Someone can hold you accountable for this result, so being right is not enough: you need to be able to reconstruct what it was based on. In practice that means two things. Store the input and the output of every run, and have the model return where it found each piece of data instead of a bare conclusion. ### With a person in the loop Mistakes have consequences that do not undo themselves, so the output is approved before it takes effect. Put a number on it from the start: the share of outputs approved without changes. Use that number to decide when, and whether, to let it run on its own. ### First, bring the data together What the process needs to know is spread across systems or lives in someone's head. A model does not fix that. It is a separate project that comes first, and without it the agent answers with half the story. The AI Data Quality assessment in this same section measures exactly that. ### And it has to answer fast It has to be solved on the spot, so response time stops being a technical detail and becomes a requirement of the process. It limits which model you can use, and you need a default outcome for when the model does not answer in time. ### And the process is not written down Nobody has written down how it is done today, so there is nothing to compare the model's output against. Documenting it is what turns “it seems to work” into “it works better than before”. ## The scale reading After the recommendation, two optional questions follow: how many people do this work today, and what budget is planned. Budget ranges are in Mexican pesos (MXN). Combined, they give one of these six readings. None of them includes a price of ours: they say whether the result fits the size described. ### First the number, then the budget Without knowing what the manual work costs today, any budget is a guess. The Manual Work Cost calculator in this same section gives you that figure by area and by task. With it, deciding how much to invest stops being a matter of opinion. ### At this scale, the savings don't justify it With only one person on this work, a project of that size will not pay for itself through saved hours. It would need another reason: learning, removing a bottleneck, or preparing for growth that is already decided. If the reason is savings, look for a process that involves more people. ### An automation doesn't cost that much What you describe is a workflow, and that budget points to a much bigger project. It usually means one of two things. Either the real scope includes integrations and data migration that did not come up in the tree, or the budget is too large, and it is better to start with less and keep the rest in reserve. ### Enough for a limited pilot, and that is fine That budget will not cover the whole process, but it covers a piece large enough to learn whether it works: one use case, a few weeks, with measurement in place from the start. It is also the right way to start even with more money, because you choose the first case for how fast it teaches you, not for how big it is. ### Start with the part that repeats most Quite a few people do this work, and the budget cannot cover all of it at once. Instead of spreading it across everything, take the part that happens most often, solve it completely, and use what it gives back to fund the next one. Spreading a little across a lot is how nothing reaches production. ### Enough for a project in stages The scale and the budget match. Even so, split it up: a first, limited stage with its own measurement, and the next stages decided on what the first one shows. A project this size approved all at once is the kind where nobody later knows whether it worked. ## There is no score here The assessments in this section add up answers and return a score. This one can't: between an automation and an agent there is no more or less. They are two different solutions to two different problems. So you follow a path and arrive somewhere. - The result is a recommendation with its conditions, not a score out of 100. - A number here would be made up, and it would get filed away like any number nobody can explain. - In exchange, the whole path stays visible: you can take the recommendation apart step by step. ## Each area has its own questions The area does not only change the examples. It changes the first part of the path. Sales asks about the trigger, finance about the type of document and whether an authority is involved, operations about the shift. These things only exist in that area. - Each area has its own opening questions, followed by the shared trunk. - The trunk is the same for every area on purpose: whether a case calls for an agent or an automation does not depend on the department. - The area also sets the first step in the result, which is the part generic advice cannot give. ## The tree really branches Some questions are skipped depending on earlier answers. With a fixed format it asks about the rule; with free-form text it does not. With no prior review, it skips the criteria question and the cost of a mistake decides. - The number of questions depends on the path. Nobody answers all of them. - The branch you did not take stays on the map, dimmed. Seeing what you ruled out is half the value of a tree. - You can change any answer, and the map goes back to that point. What came after is dropped, because it may not even exist on the new branch. ## The path has a memory Some answers switch on a condition, such as personal data, someone who can hold you accountable, or a costly mistake. That condition changes the result even if the question that set it was many steps back. - The cost of a mistake changes the recommendation itself. An agent with costly mistakes becomes an agent that proposes and a person who signs off. - Personal data does not change whether it is an agent or an automation. It changes where the input can go, and that decision comes before the model. - Without memory, a long tree forgets its own beginning and ends up recommending the same thing to everyone. ## What it is not It does not look at your operations or connect to anything. It is a decision tree you answer yourself, and it is only as good as your answers. It does not quote prices either: the budget question after the recommendation does not change it. - A small budget does not turn a language problem into a rules problem, so branching on money would stop the tree from diagnosing anything. - If the answer is “not yet”, that is a result, not a detour to sell you something else. - You don't need to leave your email to see the result. Until you decide to continue, your answers stay in your browser. ## Frequently asked questions ### What is the difference between an automation and an agent? An automation runs a written rule. With the same input it always does exactly the same thing, and if a case comes up that the rule does not cover, it does nothing. An agent built on a language model interprets input that has no fixed shape and produces something reasonable, which is not the same as something identical. The first is cheap, predictable and auditable. The second copes with variety that a rule cannot list. Most projects that fail chose the second for a problem of the first kind. ### Why doesn't it give me a score like the assessments do? Because there is no scale to measure. The HubSpot Assessment and the AI Data Quality assessment give a score because their areas go from worse to better: more duplicates is worse, more adoption is better. Not here. “Automation” is not a lower grade than “agent”; it is a different thing. A number would force an order that does not exist. ### Can I change an answer halfway through? Yes, on purpose. Each answer in the trail is a button that takes you back to that question, and the map travels back with it. What came after is dropped, because those questions may not exist on the other branch. In the site's other questionnaires the result is fixed. Here, seeing where the other branch leads is the point of the tool. ### Why does it ask for the area before starting? Because each area has its own set of questions, not just its own examples. In sales it asks what triggers the work. In finance, what documents are involved and whether an authority can hold you accountable. In operations, whether it has to be solved within the shift. The area also sets the concrete first step in the result. ### Does the budget change the recommendation? No, and the page says so. It is asked **after** the recommendation, together with how many people do this work today, and both are optional. Together they give a second reading: whether the result makes sense at that scale. The two are kept apart on purpose. The kind of solution does not depend on money. Whether it is worth doing now does. ### Why do you ask how many people do this work? Because a project pays for itself through the hours it gives back, and that depends on how many people are involved. With one person on a process, a large project will not pay back through savings, however clearly the case calls for an agent. With fifteen, a small budget cannot cover everything, so the answer is to start with the part that repeats most. Neither can be said without both figures. ### Do I have to leave my email to see the result? No. The recommendation appears when you finish the path, and you can stop there. The email is only asked later, in the contact form, if you decide to continue. Your path then goes with you, so you don't have to write it out again. ### What happens to my answers? They stay in your browser and do not leave it unless you press the button at the end. If you do, the full path is written into the contact form and saved with what you send. That lets us write to you about your case instead of sending a generic reply. --- Source: https://c2suite.com/en/resources/agent-or-automation You can cite and summarize this content if you credit C2Suite and link to the source URL.