A few months ago, almost everyone at Zecat started using Claude Code on their own. People in finance and administration who had never written a line of code were suddenly building little scripts to get through their day. That is exactly what you want to happen, and it is also a problem. This is the story of our first client, and what we are learning while it is still in progress.

Who Zecat is

Zecat is one of the region’s larger promotional-products groups. It has been around since 1994, runs its own plants and offices across Argentina, Chile and Asia, and sells through partners who put brands on physical products. A couple hundred people, a real catalog, real operations, and the kind of back office that runs on a lot of spreadsheets. They are our first client, and they were brave enough to let their whole team start poking at AI instead of waiting for a committee to bless it.

The part most people get backwards

The instinct, when a company wants to “adopt AI,” is to hire someone to build the automations for them. We think that is mostly wrong, and Zecat is why we changed our mind in practice, not just in theory.

The person in finance who has reconciled the same account for five years knows the business rules better than any engineer ever will. They know the edge case with that one provider, the exception that happens every March, the reason a number that looks wrong is actually right. When they build their own automation, they iterate in minutes instead of filing a ticket and waiting. And when it breaks, they debug it faster, because they understand what it was supposed to do. Our job is not to be the bottleneck that builds every bot. It is to make it safe and easy for them to build their own.

What we automated first

We started where the pain was loudest: finance and administration. These were long, manual Excel processes that quietly ate hours every month. Tax perception cross-checks, bank reconciliations, credit-card statement categorization. The kind of work where someone opens two spreadsheets and copies between them, line by line, hoping not to miss one.

Those run on a schedule now. But the detail that made it stick is not the automation, it is how the team trains it. Instead of hard-coding the logic, the categorization rules live in a plain sheet the finance team edits themselves. When a charge comes through that the system does not recognize, it gets flagged, and someone adds one row to the rules sheet: this pattern maps to that concept and that cost center. Next run, it is handled. No deploy, no engineer, no ticket. They own the rules because they are the rules.

The honest risk nobody mentions

Here is the part that does not show up in the demo. A whole company adopting AI tools on its own, with no guardrails, is a real risk. People paste data into whatever tool is open. There are no roles, no secure logins, no review on the thing that touches a finance report. Enthusiasm without governance is how you end up with a leak, or worse, a silently wrong number that everyone trusts because “the automation did it.”

So the answer is not to slow people down or lock it behind IT. It is to give them rails.

Where we actually add value: the center, not every bot

Once the team started building on their own, our role moved. We stopped trying to write every automation and started building the parts that are genuinely hard and genuinely shared: secure logins and roles, MCP servers that expose company systems to AI safely, API integrations, and repo templates so any area can stand up something new that connects to Supabase, Cloudflare or Vercel without wiring auth and secrets from scratch every time. The areas build the cars. We build and guard the roads.

That is also the honest pitch for what a studio like ours does. We do not sell you a platform you rent forever. We build the rails, on infrastructure you own, so your own people can keep shipping after we are gone, and we tune the templates to how your team actually works rather than to a generic default.

What is next

The same playbook moves into marketing and operations next, area by area, each one owning its own automations on top of the shared rails.

The one we are most excited about is partner-facing. Zecat sells through partners, not to end users, so imagine a WhatsApp assistant a partner can talk to: help me think of product ideas for this client, drop their logo on this item, show me how it would look. That puts the catalog in a partner’s pocket and turns a quote thread into a conversation. It is not built yet. But it is exactly the kind of thing that becomes cheap to try once the rails are in place, which is the whole point of building them first.

Why we are writing this

This is our first client, and we would rather tell you what it actually looks like than polish a case study. The lesson generalizes: adoption beats procurement, the domain experts should build, and the engineer’s job is to make that safe and leveraged. If your team already opened Claude Code on their own and you are equal parts excited and nervous, that is the right instinct, not a fire to put out. Tell us which area went first and what they are touching, and we will help you turn it into something safe and real.