Which Business Tasks Are Worth Automating
In a small or midsize business, the tasks worth automating with AI are the ones that repeat often, follow stable rules, and produce a result someone can check quickly: sorting incoming email, pulling data from invoices or PDFs, drafting replies, monitoring postings on portals like SICOP (Costa Rica's public procurement platform), putting together meeting summaries. What you should not automate are decisions that are sensitive, infrequent, or hard to review: approving credit for a customer, setting the final price on a bid, resolving a delicate complaint, making personnel decisions, or issuing legal opinions. The practical rule: if a mistake shows up within seconds and costs nothing to fix, it's a good candidate; if a mistake surfaces weeks later and costs money, reputation, or a lawsuit, leave it to a person.
The three filters for deciding whether to automate a task
Before you talk about tools, filter the task. You can answer these three questions today, with what you already know about your operation, and they rule out most bad ideas.
- Does it repeat? If the task comes up several times a week and always in the same form, there's enough volume to justify the effort. A task that happens twice a year is almost never worth automating: setting it up and maintaining it will cost you more than just doing it by hand.
- Can it be checked quickly? If it takes someone less time to review the result than to produce it from scratch, automation delivers a net gain. If checking is as expensive as doing it (for example, reviewing a complex calculation means redoing the calculation), there's no real savings, just the feeling of speed.
- Is the mistake reversible and cheap? A misfiled email gets moved to another folder. A bid sent with the wrong amount, or personal data exposed, can't be undone with one click.
If all three answers are yes, it's a candidate. If one fails, the right move is probably to automate only part of the process, not the whole thing.
Tasks that are worth automating in a small business
In practice, almost everything that qualifies falls into four families. What they have in common is that they're preparation tasks, not decision tasks: the AI gets the work ready so a person can resolve it.
- Reading and sorting. Separating emails from new customers out of those from suppliers, tagging WhatsApp messages by type of inquiry, identifying which invoices belong to which project. Mistakes are visible right away, and fixing them means moving one item somewhere else.
- Extracting data from documents. Pulling the cédula jurídica number (a company's legal ID number in Costa Rica), amounts, dates, and deadlines out of PDFs, contracts, and receipts so they can go into a spreadsheet or a system. It's tedious, prone to human error from fatigue, and easy to audit against the original document.
- Monitoring and alerting. Checking every day whether a tender has been posted that matches what your company sells, whether a competitor changed a price, whether a deadline has passed. Here the AI decides nothing: it filters out noise and puts what deserves your attention in front of you.
- Drafting first versions. A first pass at an answer to a frequent question, a meeting summary, a product sheet, a follow-up email. The draft gets reviewed and edited; nobody sends it unread.
A concrete day-to-day example: a company that bids on public contracts can have an agent reviewing new postings and flagging which ones fit its business lines and which have short deadlines. That part is repetitive and verifiable. But deciding whether to bid, at what price, and taking on which risk is still a management decision, made with the tender documents in hand.
Where automation isn't worth it, even if the technology allows it
Just because a model can generate an answer doesn't mean it should make the decision. These are the areas where, in a small business, automating turns out expensive:
- Decisions with direct financial consequences. Setting final prices, approving discounts, extending credit, authorizing payments. Having AI prepare the analysis is fine; having it sign off is not.
- Sensitive conversations with customers or staff. A serious complaint, a collections call, a contract negotiation, anything involving employment matters. The cost of a poorly calibrated message isn't the message: it's the relationship.
- Legal, accounting, or regulated technical opinions. Professional responsibility can't be delegated to a tool. AI can help search and organize; the opinion is signed by whoever is licensed to sign it.
- Processes nobody has documented. If the procedure lives in one person's head and changes case by case, automating it just locks in the mess. Write it down first, even if it's only one page.
- Anything touching personal data without clear controls. Before putting customer information into any tool, you need to know where it's stored, who can see it, and under what consent it was collected.
What Costa Rican data protection law requires before you automate
If the task you want to automate touches people's data (customers, suppliers, employees), the applicable framework in Costa Rica is Law No. 8968, Protection of the Individual with Regard to the Processing of Personal Data, enacted on July 7, 2011 and published in the official gazette La Gaceta No. 170 on September 5, 2011. Its purpose is to guarantee every person, regardless of nationality, the right to informational self-determination and privacy in the processing of their personal data.
The operational detail is in Executive Decree No. 37554-JP, the regulation implementing that law, along with its amendments in force, including those introduced by Executive Decree No. 40008-JP. It defines the procedures for obtaining informed consent, for individuals to exercise their ARCO rights (access, rectification, cancellation, and objection), for registering databases intended for distribution or commercial use, and the oversight powers of the supervisory body.
That body is the Agencia de Protección de Datos de los Habitantes (PRODHAB, Costa Rica's data protection authority), which reports to the Ministry of Justice and Peace. Case files, database registrations, and the current regulations can be consulted on the official portal prodhab.go.cr or in the Sistema Nacional de Legislación Vigente (SINALEVI, the national database of current legislation). A concrete check you can run today: list which outside tool receives your customers' data, what country it's stored in, and whether you have those people's consent on record. If any of those three answers doesn't exist, sort that out before you automate.
How to measure a task before automating it
You can do this during the coming week, without buying anything. Pick a candidate task and keep a simple log for five business days:
- How many times a day it happens and how long each time takes. Actually time it, don't estimate: the estimate and the stopwatch almost never match.
- Who does it and what they stop doing while they do it. The real cost of a repetitive task is usually the interruption, not the minutes.
- How many times it was done wrong and how the error was caught. If nobody caught it, you already have a problem that predates AI.
- What percentage of cases are "normal" and how many are exceptions. If half are exceptions, the task isn't ready yet.
Then run the twenty-case test: take twenty real examples from the past and run them through the tool you're evaluating. Compare result by result against what the person did. If it fails on three out of twenty and all three failures were visible at a glance, it can work with human review. If it fails on three and one of them wasn't noticeable, pause the project until you understand why.
What to expect on timelines and what can't be guaranteed
Let's be direct here. A simple, well-scoped automation can be running in a few weeks, but the period where you're still fixing odd cases lasts considerably longer: count in months, not days. For the first few weeks you'll review one hundred percent of the output, and at the start that can feel like more work, not less. That's normal, and it's part of the cost.
There's also no guarantee that an AI system will always get it right, or that its behavior won't change when the model provider or the data source changes. That's why any serious automation needs two things from day one: a way to review what it did, and a button to shut it off without the operation grinding to a halt. If a vendor promises perfect accuracy, instant results, or "zero human involvement," they're selling an expectation they won't be able to deliver on.
Where to start next week
Pick one task. The most boring one, the most repeated one, the one everyone in the office hates and nobody will miss. Automate it with mandatory human review for at least a month and measure the same things you measured before: times per day, minutes, errors. If the numbers improve, expand; if not, you learned something cheaply. Starting with the most critical process in the business is the most common mistake and the most expensive one.
If you'd like to work through this exercise with someone who has already done it at other companies, at Axiomify we design Axiomify AI agents for specific, verifiable tasks, with human review built in from the start. The first thing we do in any conversation is look together at whether the task you have in mind passes the three filters; if it doesn't, we'll tell you before you invest.
This guide is general information. Timelines, costs and requirements can change; for a specific case, review it with the right person before deciding.