700 engineers in India were sold as AI
In 2025 a story broke that should hang on every operations director’s wall: a Silicon Valley firm sold “AI agents” for support and back office. The product, in practice, was a floor of about 700 engineers in India doing the tasks by hand and returning results with bot-like latency.
It is not a cute anecdote. It is the extreme of a pattern we already see in local RFPs: the AI label sells a disguised human flow, a Zapier with a prompt, or a chatbot that dumps everything on an agent. The buyer pays software prices and receives an agency.
Why the costume works
Because buyers rarely ask to see the process. They ask for a polished demo, a “seconds” SLA, and a GPT logo on the one-pager. If the result arrives fast and in clean Spanish, the source (model, person, or hybrid) stays opaque. In Costa Rica and the rest of LATAM that meets another temptation: importing a fashionable vendor without asking where the data lives, who runs it at 2 a.m., or what happens when the “agent” calls in sick.
- Ask for the real diagram: what a model decides, what a person decides, what sits in a queue.
- Demand an audit trail. If you cannot inspect a decision, it is not a system. It is a ticket shop.
- Measure cost per transaction at real volume, not a ten-case demo.
- Ask about data residency and the plan for when the model or the floor of people is down.
What is worth automating
We do not sell magic at Apify. We sell processes we already know: e-invoices that land in email, ID checks against the TSE, a WhatsApp inbox that is not split across five apps. There, automation can be seen, tested, and rolled back. If someone sells an “agent” that replaces a whole team and cannot show the rail, assume the rail is people.
Useful AI leaves a log. Catalog AI leaves an invoice.
If you are evaluating a product with “AI” in the name, bring us the case. We will test it against an honest flow: software where it belongs, people where they belong, and a cost you can defend to leadership.