AI process automation

The problem

“AI integration” today is a synonym for hype: demos with a model stuck on top of a form, “agents” that are a prompt with no safety net, promises to automate “everything” without measuring the return. The typical result is a tool that shines in the demo and fails in production: no error handling, no way of knowing when the model got it wrong, no plan for when it doesn’t know what to answer.

Meanwhile, the work that is worth automating keeps being done by hand: the same piece of data copied from an email into a spreadsheet, the invoice someone transcribes at nine at night, the report put together every week by copying and pasting. Repetitive, structured tasks, where a mistake costs money and where a person loses hours they could spend on what does need judgment. The models are capable of this work. What’s missing is the engineering around them: without guardrails, without human review at the critical points, without a log of what the AI decided, automating is riskier than not doing it.

What I do

• I automate the process, I don’t sell AI for its own sake. First I look at which repetitive task is costing you time and errors. AI comes in where it adds something (reading, classifying, extracting, summarizing); the rest is solved by good software.
• The engineering around it, which is what’s missing in the market: clear guardrails, human review at the critical steps, and a log of every model decision so it can be audited.
• Confidence threshold: when the model isn’t sure, it escalates to a human instead of making something up. The automation has a handbrake.
• I add it to what you already run: integration with your current tools (or through Zapier/Make when you don’t need a full system), instead of an isolated product you have to feed separately.
• Models chosen by task (Claude, GPT), for what they do well rather than for what’s in fashion.

Chatbots, WhatsApp and n8n: when they make sense and when they don’t

These are the three things people ask me for most by name, so I’ll give you my position before you ask. About n8n in particular I wrote a separate piece on what it is and what it’s for, with five workflows and what each one costs per month (in Spanish).

A customer service chatbot pays off when the same handful of questions repeats every day and the answers are already defined: opening hours, order status, requirements, booking an appointment. There a well-scoped bot takes real work off your plate and answers at three in the morning. It stops paying off when your volume of inquiries is low or every case is different, because you end up paying to maintain a funnel that serves people worse than you do, and the customer is going to type “talk to a person” anyway. And if your real problem is that inquiries reach you through five different places, that gets sorted out first, and sometimes that’s enough.

With the other two the answer also depends on the case:

• WhatsApp: yes when the flow has clear states (confirm, reschedule, notify, collect payment) and the message replaces a call that someone makes by hand today. That’s how appointment confirmation works at CM Vida Plena: a state machine, not an open conversation. Not as a free-form conversation channel with a loose model on the other end: the API charges for every message the business delivers, including replies past 1,000 a month, so a chatty bot pays for every turn of a conversation that isn’t going anywhere. And watch out for the intermediary platform: several charge a monthly fee on top of the message cost to give you an interface you don’t need if the flow lives inside the system you already have.
• n8n and workflow tools: yes for connecting services and moving data between the things you already use, especially when the flow is linear and the volume is moderate. It’s quick to build and you can review it yourself. Not for critical business logic: when the process has rules that change, money involved or auditing, a visual flow with twenty nodes becomes impossible to maintain and to test. That’s where code and tests make sense, which is custom software.
• And sometimes the answer is not to automate. If the process runs a few times a month, if it changes every time or if nobody has written it down in steps yet, automating it freezes a mess in place and adds one more piece to maintain. The honest answer there is to put the process in order, measure it for a couple of months and only then decide. That conversation is free and I’ll tell you before quoting.

What’s left out: if what you need isn’t automating a task but building the complete system that contains it, that’s custom software. I’ll tell you which one your case is before quoting.

Real case

I don’t yet have an AI process running in production for a client, and I’m not going to make one up for you. What I do have are real automations that have been operating for years, with the same pattern I apply when the task calls for AI: clear rules, a handbrake and a log.

In the CM Vida Plena system I automated payment reconciliation: the engine matches bank transfers against invoices, and doubtful cases are left for human review. Money involved, zero tolerance for silent errors. AI wasn’t needed here, and I didn’t put it in: that decision is also part of the job.

On my own projects, at scale: wp-seguro, a pipeline that audited 25,399 WordPress sites, classified what it found and automatically notified the owners affected. A repetitive process running on its own, with real volume and real consequences. Along the same lines, the per-branch stock at PetShop MG: my own implementation to show the availability at each location in the online store.

The AI part is in development, in a project of my own: a vision pipeline (Claude Vision) that extracts structured data from messy documents and images, with human review before any action and a log of what the model decided in each case. When it’s in production, it will be here, with figures.

That’s why the first step is mapping: looking at your process and deciding whether it calls for AI or whether good software solves it. If I tell you that you don’t need it, I’ll also tell you before quoting.