I tested an 'AI solution.' Two days in I realised it was Anna from Manila
A story about a product the client was paying $18,000 a year for — which turned out to be a spreadsheet with four operators. Six tells that catch this in five minutes.
In July a client called with a small and sad problem. A year earlier he'd bought an "AI solution for processing incoming requests." Was paying $1,500 a month. The solution worked: emails came in, an hour or two later answers went out, customers were happy.
In July the solution started glitching. One day it didn't reply to anyone. The next day it replied, but in eight languages, only one of them correct. On the third — silence again. The client called the provider's support. They said "we'll fix it, technical maintenance." A week later — same thing.
— Understand, — he says, — my customers are starting to take offence. Could you look at it? You're into AI.
I looked. Over two days I realised there was no AI in this solution. At all. Nothing.
Inside the "AI solution" was a Google Sheet. A real Google Sheet. With formulas. Attached to it — four human operators — judging by IPs and timezones, one from Manila, two from Bangalore, and one somewhere in Latin America (timezone matched Argentina, but I'm not sure — could have been Colombia). Emails dropped into the sheet, each operator worked their queue, wrote a reply from an internal template, sent.
The July glitches had a simple explanation: the Manila operator had quit, the others had been overloaded, they weren't coping, and now the internal process was breaking. That's your "technical maintenance."
This isn't evil. There's a completely legitimate tool in business — Mechanical Turk. People clear a queue, the process gets masked as automation, the customer doesn't wonder. Sometimes it's the only way to launch a service that doesn't yet have the volume to justify training a real model.
The problem is that this needs to be said honestly. My client's contract had phrases like "AI-powered platform," "machine learning architecture," and "intelligent routing." That's not "MVP with humans in the loop." That's deception.
And the client paid $18,000 a year for it. Plus the reputational damage to his own customers when the solution fell apart.
Over the next six months I happened to check seven more AI products my clients were using. In three of seven — humans inside. In another two — humans partially (simple cases went to the model, complex ones to a human, the client didn't know). In only two — real end-to-end automation.
That's a field result, not a global statistic. But I think it tracks with what other independent auditors report.
Six tells I now check on the first demo, and which will out Anya-from-Manila in five minutes.
One. Response latency 30 seconds and above. A normal LLM on a typical task answers in 2–15 seconds. If an email gets handled in ninety seconds — especially during Manila or Bangalore working hours — that's a queue.
Two. Email at 3 AM on a Saturday. Tried it once, got a reply twelve hours later when the morning shift came in. The reply included an apology for "a small delay due to system maintenance." System maintenance, apparently, sleeps on schedule.
Three. Too human phrasing. LLMs make predictable errors: template phrases, over-warnings, neatly-formed paragraphs. Humans under stress and KPIs break differently: typos, messed word order, sudden case shifts. When I saw a reply with "sorry for the inconvenience :)" — with the anglicism and the smiley — I already understood.
Four. Same input three times — three noticeably different answers. An LLM at temperature=0 returns nearly identical output. A human — always slightly different, because they're human, their mood, fatigue and timezone shift.
Five. No API. No docs. No streaming. An AI product without a public API in 2026 is like an online store without a cart. Possible. Suspicious. No API means no machine maintaining it.
Six. Too-polished output on complex tasks. The most counterintuitive tell. A perfectly formatted financial report in four seconds is probably two hours of an analyst's work handed to you with a delay and an "AI-generated" label. Fast LLMs don't do that. And those that do — on an expensive model — answer at $2, not $0.10. If you get "expensive-quality" output at cheap prices, the quality is produced by a human, not the model.
When "human in a basement" is the right answer:
— Complex legal review where errors cost money or reputation. — Medical labelling where a regulator requires human-in-the-loop. — Rare languages or formats where models are weak (Basque, dialect-level Filipino, niche industry codes).
In those cases an honest company says so: "we use human review." The price reflects that — it's higher, because a human-hour is pricier than a model-second. That's fine.
Bad — when a company says "AI" and inside it's just Anya.
Questions I now push on before signing anything:
— Show the architecture. Not "tell me about your AI" — show me on a diagram where the model decides, where rules decide, where a human decides. If they can't draw it — they don't know. If they don't know — what's inside isn't what they're selling.
— Give me a latency SLA. Seconds, not minutes. If "up to 30 minutes" — it's not AI.
— Show pricing at a million requests. Real AI scales predictably: more requests = slightly more tokens = slightly more $$. Mechanical Turk scales linearly in people: 10x requests = 10x humans = 10x cost. If price doesn't drop with volume, a team is inside.
— API access. No API — it's not a platform, it's a service with a label.
"AI inside" is not a guarantee. It's a promise that, in my field experience, doesn't hold about 40% of the time.
Verification costs five minutes. Disappointment — $18,000 a year. My client is now using that money to build his own assistant, honestly, on Claude, with visible code and measurable latency. It'll ship in three weeks, and cost him fifteen times less than the previous "solution."
Don't be shy about uncomfortable questions. Honest vendors answer. The rest go quiet.