Why Most AI Agency Pitches Fall Flat on Their Face in the Real World

Seven. That’s literally how many pitch messages landed in my LinkedIn inbox by lunchtime on Monday. Every single one had the same angle: give us five grand, hook our AI up to your CRM, and fire half your office staff by next Friday.

It’s completely delusional.

Look, I get why people buy into the hype. The demo videos on Twitter look incredible. You upload a clean little CSV, ask a bot a question in plain English, and boom—it spits out a tidy chart with nice little pastel colors.

Now try doing that with an actual business.

Two months back we went out to a distributor in North Dublin. Real company, thirty years in business, sixty employees. Their "tech stack" is an on-premise PostgreSQL box running on an ancient Dell PowerEdge tower in the server closet under the stairs, plus three ladies in accounts—Mary, Joan, and Sarah—who spend half their working lives typing numbers off paper invoices into green-screen terminal windows.

Half those invoices don't even arrive as clean digital PDFs. They come in as terrible mobile photos taken by truck drivers in rainy loading bays, scanned delivery dockets with coffee rings right over the VAT number, and weird Excel sheets where someone put notes in the quantity column.

You plug ChatGPT directly into that mess? Good luck.

An LLM doesn’t know what copper cable costs per meter. It guesses. If an image is blurry and a line says €14,200, the model might happily hallucinate €11,200 and stick it straight into the general ledger without blinking. You won't find the missing three grand until your quarterly audit blows up in your face six months down the road.

When we set up automation for these guys, we didn't start with neural nets or fancy agent frameworks. We started with boring stuff.

We wrote twenty lines of dead-simple Python to check the basic math before anything else runs. Does Line 1 plus Line 2 equal the bottom total? No? Then don’t even send it to a model. Kick it straight to an alert box on Mary's screen.

For the lines that did pass basic sanity checks, we ran them through a small, quantized local model sitting on our own rack in Dublin. We didn't touch public cloud APIs—partly because shipping customer names and supplier pricing across the Atlantic is a GDPR nightmare waiting to happen, but mostly because local inference is faster and costs a fixed €80 a month instead of burning API credits on every failed scan.

And we kept a human in the middle. If the parser wasn't 99% certain about a supplier SKU code, it highlighted the box in yellow and let Joan hit Enter on her keyboard to confirm it.

Took us three weeks to wire up. Mary and Joan didn't lose their jobs; they just stopped having to spend five hours every Tuesday doing mind-numbing data entry, and the company hasn't had a single inventory discrepancy since we flipped the switch.

Next time an agency tries selling you "autonomous cognitive workflow agents," ask them one thing: what happens to your system when a supplier faxes a purchase order upside down? If they stare at you blankly, show them the door.