Chatbot and automation: one-time payment or monthly subscription? The 24-month math

If your automation depends on APIs you do not control —Meta Cloud API, Mercado Libre, Google Calendar, an AI model— the monthly subscription ends up being the sensible option even if the number on the proposal says otherwise, because those platforms change on their own and someone has to fix it when they do. If the workflow only touches your own systems (your database, your spreadsheet, your mail server) and there is someone inside who knows how to read a log, one-time payment is correct and the subscription is money thrown away. In round numbers: a one-time project of USD 2,200 ends up costing between USD 4,100 and 5,300 over two years once you add the real maintenance; a subscription of USD 249 per month adds up to USD 7,524 in the same period. The difference is not the software. It is who answers the phone the day it breaks.

The 24-month math, with a case worked end to end

Take a concrete case: a three-chair dental practice, around 600 WhatsApp conversations per month. The workflow answers frequent questions with RAG over the list of treatments and the accepted health insurance plans, books the appointment in Google Calendar, sends the reminder 24 hours in advance with a utility template, and creates the patient in the CRM.

One-time payment option (custom project):

• Discovery, architecture and build: USD 2,200 one-time, with 90 days of support included.
• Your own 4 GB VPS (Hostinger or Hetzner): USD 14 per month, USD 336 over 24 months.
• Meta, utility and service conversations: USD 18 per month, USD 432 over 24 months.
• AI with a mini model for 600 conversations: USD 9 per month, USD 216 over 24 months.
• Subtotal if nothing ever breaks: USD 3,184.
• Real maintenance: between 4 and 7 interventions in two years, of 2 to 6 hours each, at USD 45-70 an hour for a senior freelancer. That is USD 900 to 2,100 more.
• Realistic total: USD 4,084 to 5,284.

Monthly subscription option (pre-designed solution, live in 72 hours):

• Setup: USD 900.
• Monthly fee: USD 249 including hosting, monitoring, updates and support. USD 5,976 over 24 months.
• Meta and AI billed at cost, separately: USD 648.
• Total: USD 7,524.

The number that matters is not which of the two is cheaper, it is how much separates one from the other: between USD 2,240 and 3,440 over two years, that is, between USD 93 and 143 per month. That is the figure you have to evaluate, not the USD 249 of the monthly fee. The right question is not whether you can pay USD 249 per month; it is whether you would rather pay USD 120 per month on average so that you are not the one doing the diagnosis at 10 pm on a Friday.

What you are really paying for in a monthly fee

A USD 249 subscription breaks down like this, and every line has a market price if you buy it separately:

1. Hosting and database. A 4 GB VPS costs USD 8-15 per month; a managed Postgres like Supabase Pro, USD 25. It is the cheapest part and the one everybody uses to argue that the subscription is expensive.
2. Monitoring with someone behind it. The tool is free: self-hosted Uptime Kuma costs zero. What costs money is having someone look at the alert on a Sunday at 9 pm and know what to do with it.
3. Updates when a third party changes. Meta versions the Graph API and retires old versions every couple of years. AI providers announce the retirement of a model with 3 to 12 months of notice. n8n publishes releases every week and some nodes change shape between versions. All three announce it through changelogs and developer newsletters, not to the inbox of whoever hired the bot.
4. Support with an SLA. The difference between a response in under 4 business hours and whenever I can is not a clause, it is the number of appointments you lose.
5. Content iteration. You changed the price of the first consultation, added a health insurance plan, closed on Saturdays. If nobody updates the RAG knowledge base, the bot does not fail: it confidently quotes a price that no longer exists. That is worse than being down.

The risk of one-time payment is not that it goes down: it is that it fails silently

Let us follow the same dental practice. The mechanism is the one we have seen repeated most often; the numbers are still those of the case modeled above.

Month 8. Meta recategorizes the reminder template: the text had a promotional phrase at the end and moved from utility to marketing; quality drops and the template ends up paused. The n8n send node has Continue On Fail enabled —almost everybody enables it so that one invalid number does not cut off the morning batch— so the daily execution keeps showing up green. Zero messages sent, zero alerts.

Nobody finds out for three weeks. What you see from the other side is the absenteeism: out of 60 weekly appointments, the no-show rate goes from 8% to 18%. That is 6 appointments lost per week at an average ticket of USD 45: USD 270 per week, USD 810 over the three weeks. Add 4 hours of a developer to find the cause at USD 60, and the incident cost USD 1,050. Next to the USD 93-143 per month that separated the two models, that single event ate up seven months of the difference.

The expensive failure is never the one that brings the system down. It is the one that returns success.

What breaks an automation without anyone touching it

None of these causes is triggered by a change of yours. They all come from outside or from the simple passing of time, and that is why one-time payment does not cover them:

• Meta recategorizes or pauses a template. The price per conversation changes and, if quality drops, sending stops.
• The Mercado Libre refresh token. Mercado Libre is the dominant marketplace in LATAM, and its access token lasts 6 hours while the refresh token is single-use and expires after 6 months. If the workflow was stopped and did not renew it, you have to reauthorize by hand from the seller account.
• Google OAuth in Testing mode. If the Google Cloud project never completed verification, the refresh token is invalidated every 7 days. Calendar sync works for a week and dies, and the symptom looks like a problem with the bot.
• The VPS disk. n8n saves every execution in the database. Without EXECUTIONS_DATA_PRUNE=true and a reasonable EXECUTIONS_DATA_MAX_AGE, the database only grows. The practice in the example runs around 20 executions a day and takes years to have a problem; an e-commerce with 600 a day fills an 80 GB disk in 8-12 months. n8n does not come back up and nobody notices until a customer writes and nobody answers.
• The AI model that gets retired. The day they shut down the endpoint, every message returns an error. Migrating is not changing the model name in a field: prompts tuned for one perform differently on another and you have to test the edge cases again.
• The unofficial WhatsApp session. If the chatbot runs on Evolution API with Baileys instead of the Cloud API, the session disconnects because of a phone change or an app update, and someone has to physically scan a QR code. We go into it in Evolution API self-hosted.

When one-time payment is the right option and the subscription is money thrown away

Five situations where charging you a monthly fee would mean selling you something you do not need:

• The workflow does not touch any third-party API with OAuth. It moves data between your Postgres, your spreadsheet and your email. There is no token that expires and no platform that changes its rules.
• It is a job with an end. Migrating 12,000 contacts from a spreadsheet to Pipedrive, unifying the catalog before uploading it to Tiendanube (the e-commerce platform of reference for SMBs in LATAM), generating a year-end close. It runs, it finishes and it does not come back.
• There is someone inside who goes in to look at a failed execution. You do not need a senior developer: the IT person who already administers the ERP is enough, if they have the access and the assigned time.
• Low volume. If the workflow moves 30 operations per month, a USD 249 subscription is more than USD 8 per operation. Automate it once, document it and leave it alone.
• Procurement does not approve a monthly vendor with production access. It happens in mid-sized companies and it is a legitimate constraint. It is solved with one-time payment plus a block of hours contracted as a professional service.

The rule we use: if the workflow depends on more than two third-party APIs you do not control, one-time payment has an expiry date. It is not that it is wrong; it is that it is better to know that date before signing, not to discover it in month 8.

The middle ground: one-time payment plus maintenance, without hosting

There is a third option that almost nobody quotes and that we ask for most often when the client has their own team: a one-time project, infrastructure in the client account, and reduced maintenance of USD 79-129 per month covering monitoring, updates for third-party changes and support, without hosting. Over 24 months that is USD 1,900-3,100 on top of the setup: between a third and a half of what the subscription fees add up to, and the server is yours.

The variant for low volume is a block of 10 hours per year, used when something breaks and not lost if nothing happens.

What does not work is the fourth option, the one that shows up on its own when nothing gets decided: one-time payment and then I will call whoever built it. In practice that call takes between 3 and 15 days to turn into work, because the person is already on another project, and it starts with two hours of remembering how this thing was put together.

Ownership breaks differently in each model

Both require the same baseline —workflows exported as JSON in your Drive, credentials in your own accounts, the notice period in writing— but they fail in opposite places and it is worth knowing which one to watch.

With one-time payment the typical hole is the infrastructure: they hand you the workflows, but the VPS and the domain ended up contracted with the provider card and email because nobody defined whose name they were bought under. A year later you need to get into the server and you depend on someone answering. Ask for the transfer of those accounts on delivery day, not when you need it.

With the subscription the infrastructure belongs to the provider by design, so what has to be agreed is the exit: which version of the workflow you take with you the day you cancel —the last one in production or the one that existed when you signed— and with how many days of notice. It is the clause almost nobody reads and the only one that matters when you want to leave. We break it down in full in if you hire an agency, are the workflows yours?.

How we resolve it in the 15-minute audit

Four questions with thresholds, in this order:

1. How many third-party APIs does the workflow touch? Zero or one, one-time payment. Three or more, subscription.
2. How much revenue goes through it? If the process touches more than USD 3,000 of monthly revenue, do not leave it without monitoring, whichever model you choose.
3. Who goes in to look at a log? If the answer is whoever built it, you already have a subscription; only without a price, without a term and without an SLA.
4. How much does 48 hours of downtime cost you? If that number exceeds six months of the monthly fee, the discussion is over.

If you want the numbers for your case —real volume, which systems it touches, how much the process weighs in your revenue— book the free 15-minute audit or write to us at info@striqtech.com. If the conversation shows that your workflow does not justify a monthly fee, we tell you right there and quote the one-time. To put together the full project budget it helps to first read how much an AI WhatsApp chatbot costs and the 90-day roadmap.