Off-the-Shelf Tool or Custom AI Agent? How to Decide
The questions that separate the two cases, what each path costs and when you don’t need either.
Updated on November 6, 2026
When a tool is enough and when an agent pays off
Pay for a tool with a monthly fee when your process looks like that of many other companies and the tool already connects with the software you use. It’s usually faster and cheaper than a custom build. We’ll tell you that even though building agents is what we do.
Commission a custom agent when the process depends on your judgment, your data or your exceptions and no tool can complete it without someone on your team finishing it by hand.
You don’t have to pick one for everything. You can buy the standard pieces, such as the calendar or the accounting software, and commission only the part that connects them the way you work.
If you’re still unsure, try the tool first. That’s what the AI guide for small businesses from Spain’s Chamber of Commerce (September 2026, in Spanish) advises: start with a simple solution before considering a custom build. If it falls short, you’ll know what’s missing and you can ask for a more specific quote.
The three questions that decide it
Answer with one specific process in mind, not the whole company. Don’t ask what each path can do, because an agent can do almost anything a tool can. It can also connect with your software and let a person review every result. Ask whether a product that solves your process already exists and, if it doesn’t, whether building one pays off. Go in this order.
Does any tool do the whole process? Check it in a test with your data and your version of your software, not on the vendor’s website. Check too whether your data can live with that vendor. If the tool does the whole process, you need nothing more.
Can configuring it solve what’s missing? Ask whoever does the work today to tell you about the cases that fall outside the norm. Check whether the tool accepts those rules and those exceptions. If it does, configure it and stay with it.
Does hand-finishing cost more than an agent? Count the hours that go each month into finishing what the tool doesn’t finish. Compare them with the agent rows in the cost table. If the task only shows up now and then, it will hardly pay off, so stay with the tool and finish by hand. If it repeats often and hand-finishing costs more, commission an agent and start with the smallest version that lets you test it.
Whichever you choose, ask what happens if you stop paying one day. Ask what data you take with you, whether the process can keep running and who decides when the system changes. In our projects, the repository and the infrastructure are in the client’s name from day one, and the monthly fee has no minimum term.
What each path costs over 1 and 3 years
A subscription costs more the more people use it. An agent costs more the more systems it touches. An AI subscription for business costs about €20 per person per month, based on official pricing as of October 2026. A custom agent costs €2,500 to €10,000 to build, plus €150 to €500 a month to monitor and maintain it. An MVP, which is a smaller first version for a simple task, can start at €700 plus €70 a month. Off-the-shelf tools charge in different ways, by volume or by result. At those prices, the math looks like this:
| Path | First year | Over 3 years |
|---|---|---|
| Zapier, to connect software (750 tasks a month, $19.99 a month) | $240 | $720 |
| Lyro by Tidio, AI customer service (50 conversations a month, €32.50 a month) | €390 | €1,170 |
| MVP of an agent for a simple task (€700 plus €70 a month) | €1,540 | €3,220 |
| AI subscription for 10 people (€20 per person per month) | €2,400 | €7,200 |
| Agent for a single task (€2,500 plus €150 a month) | €4,300 | €7,900 |
| Fin by Intercom, AI customer service (500 outcomes a month, $0.99 each) | $5,940 | $17,820 |
| Agent connected to several systems (€10,000 plus €500 a month) | €16,000 | €28,000 |
| AI subscription for 50 people (€20 per person per month) | €12,000 | €36,000 |
Don’t choose by the cheapest row alone, because they don’t do the same job. The subscription helps each person while they work. Off-the-shelf tools do a specific task within what each product allows. The agent does a specific task by your rules, and you decide which results a person reviews.
The MVP is there to test whether the task pays off. If the agent grows, its cost moves toward the other agent rows. All the agent rows leave out the model and cloud usage. That spend goes through accounts in your name and rises or falls with use.
The Zapier, Tidio and Fin rows are examples using their official pricing on November 6, 2026, with annual billing where they offer it. They aren’t a recommendation, and those prices change often. The Fin row counts only the outcomes, without the licenses for the customer service software. For the tool you’re considering, do the same math with your volume. Ask too what it costs to connect it with your software and whether WhatsApp messages are charged separately.
No invoice shows the time of whoever finishes by hand what the tool doesn’t finish. Multiply the minutes for each fix by how many times it happens a month. For example, five minutes repeated a thousand times a month is more than 83 hours. Put those hours next to the table before you decide.
The three questions in a real case: a rental agency
A rental agency was getting dozens of WhatsApp messages a day from people interested in renting. Each inquiry took five to ten minutes of manual checking. Here’s how that process answers the three questions.
Does any tool do the whole process? WhatsApp templates didn’t. The agency tried sending automatic messages with them, but they didn’t save much time.
Can configuring it solve what’s missing? A template sends a predefined message. The agency needed a conversation that gathered each person’s details and compared them with its criteria, such as financial solvency, pets or whether they smoke. The agency didn’t try other customer service tools, so we don’t know whether one would have been enough.
Does hand-finishing cost more than an agent? The agency decided it did and commissioned one from us. Today the team gets back more than three hours a day. With 20 working days, that’s more than 60 hours a month.
The agent runs the conversation, gathers the details and sends the team an email summary with suitable or not suitable and the reason. A person reads that summary and decides. The criteria live in a spreadsheet the team edits itself. The summary goes by email because the agency’s management software lets you read data but not write it. That limit comes from the software and affects either path.
As in the rest of our custom projects, the repository and the infrastructure are in the client’s name. If the agency drops the monthly fee, what was built doesn’t switch off. With few inquiries a day, the sensible choice would have been to keep doing it by hand.
What you give up on each path
Each path solves some things and leaves you with others. Before deciding, look at which of these losses affects your process least.
With a tool, you give up control over the product. The vendor decides the price, the features and when they change. If the price goes up tomorrow or something you use disappears, you adapt or you switch tools. Whatever the product doesn’t allow, your team will keep finishing by hand.
With an agent, you give up speed at the start. Zapier and Tidio have free plans, so you can try them the same day. An agent needs a well-defined task, the data and the access. With all that ready, a pilot in two weeks is realistic.
With an agent, you put the money up first. The cost of building comes before the agent has saved you anything. If the task doesn’t pay off, you don’t get it back. That’s why it makes sense to start with the smallest version that lets you test it.
With an agent, you take on the maintenance. Someone in your company has to answer for the criteria and change them when the business changes. You also need someone technical to monitor and maintain it, which is what the monthly fee pays for. Having the repository in your name lets you change providers, but it doesn’t remove that need.
If your process looks like that of other companies and switching tools would take little effort, losing control barely affects you. If the process depends on your judgment and repeats daily, that loss can cost you more than maintaining an agent.
When you don’t need either
Sometimes the right call is not to buy or commission anything yet. These are the six cases.
A step in the process is unnecessary. See whether you can remove it before automating it. Automating an unnecessary step only makes it happen faster.
The task is paperwork with no decisions. If it’s always done the same way, a rule in the software you already use or a macro may be enough. Look first at what you already pay for.
The task rarely happens. A workflow that runs three times a month doesn’t pay for the monitoring an agent demands. Before adding another monthly fee, also check whether it beats continuing to do it by hand.
The data isn’t in order. Neither a tool nor an agent will fix scattered or outdated information. An agent without reliable data answers with whatever seems most likely. Here the first job is getting the sources in order.
Nobody knows which number should improve. It could be hours, requests handled or turnaround times. Without that number you can’t check whether it worked or defend the spend when it’s time to renew.
Nobody will look after it. A tool and an agent both need a person in your company who watches how they’re doing and decides small changes. Without that person, they end up abandoned.
Frequently asked questions
My software is old and has no connectors. What do I do?
First see what it allows: a technical connection (API), a file export or access to its database. A tool with a monthly fee usually needs a ready-made connector. A custom agent can work with any of those routes, though sometimes an intermediate piece has to be built. If the software only lets you read data, the result can go out another way, such as email. If there’s no route at all, what’s left is imitating clicks on the screen. That breaks when something moves.
Can I start with a tool and commission an agent later?
Yes, and it’s what we advise if you’re unsure. The test will tell you what the tool is missing and you can ask for a more specific quote. Before you start, check that you can get your data out of it. You can also keep both and commission only the part the tool doesn’t do.
Zapier or a custom agent?
Ask yourself the same three questions. Zapier is usually enough when the steps are always the same and your software has a connector. A custom agent starts to make sense when each case requires interpreting something by your criteria, the test with Zapier doesn’t solve it and hand-finishing costs more than building it.
Does a custom agent run without anyone reviewing it?
Only if you decide so. At first the agent proposes and a person reviews every result. When the tests and real use show it works in the expected cases, the actions it can take without review are widened. The decisions you don’t want to delegate can always stay with a person.
How long does each path take?
A tool with a free plan can be tried the same day, though connecting and configuring it takes longer. With the data, the access and a well-defined task, a pilot of an agent in two weeks is realistic.
What happens to my data on each path?
On both paths your data goes through some provider, at the very least the AI model’s. With a tool, the vendor usually decides where it’s processed and you accept that in its contract. With a custom agent, you decide where it goes, because the accounts are in your name. If you handle personal data, ask for a data processing agreement either way.
Torn between a tool and a custom agent?
Tell us about the process. If a tool is enough for you, we’ll say so.