cd /news/artificial-intelligence/the-ai-job-that-failed-silently-and-… · home › topics › artificial-intelligence › article
[ARTICLE · art-173498] src=dev.to ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

The AI job that failed silently, and the habit I brought back from network engineering

Real estate operator and developer Paul David Thompson built a Whisper-based transcription pipeline that fed an assistant answering member questions from recorded coaching calls with cited sources, and it ran in production for six months before a backend billing failure silently stopped transcription with no alert. He also built an intake pipeline that cut underwriting time per deal from two hours to 15 minutes and an offer-generation tool in daily use by paying members of his education company, Real Estate 101. Thompson argues AI projects fail on operational gaps—fault monitoring, rollback, and ownership—rather than on model choice.

recommendation

Add fault monitoring and alerting to every scheduled ingestion job on day one; a silent billing failure left a retrieval assistant serving stale answers with no alarm.

more in playbook: Running AI in production
by read5 min views1 publishedNov 20, 2026

I shipped a Whisper pipeline that transcribes my coaching calls and answers member questions from what was actually said, with the source. It ran for six months. Then a backend billing issue stopped the transcription, and nothing alarmed on that failure. The job sat broken until a person noticed.

I'm Paul David Thompson. I buy commercial real estate and I build the software that runs it. Before that I spent eighteen years in enterprise network engineering at a Little Rock internet provider, ending as Director of IP Networking at Windstream. That was a while ago, and I do not claim it as current expertise. What carried forward still matters: change control, blast radius, and rollback.

The reason most AI projects die is not the model.

I still hear a lot of conversation about prompts and model choice. The work that decides whether a tool stays useful is easier to miss. You need to know when a scheduled job stops. You need to undo what an agent just did. And you need to start with a process you already own, instead of a demo that has no owner.

The tools I ship started as tools for my own work: underwriting screens, offer generation, document pipelines, and agents that answer from my own call recordings. Some of those tools are now in daily use by paying users. I log what I build, including what broke, on pauldavidthompson.com.

If you build software for a living, this will look familiar. I'm writing it down because I keep watching AI projects stall on the same operational gaps I used to watch in networks. A real estate operator I was building for (and I am one) was sorting inbound leads by hand, then building offer terms one deal at a time. Two hours of work before a single number reached a seller. The bottleneck was not deal flow. It was the operator.

I built an intake pipeline that scores and routes incoming leads against the operator's own criteria, then generates customized offer structures for the ones that clear. The judgment stayed with the operator. The typing did not.

Underwriting time per deal went from 2 hours to 15 minutes.

I would have started sooner. I also should have layered in light CRM functionality from day one, instead of leaving it for later. Intake, scoring, routing, and a place to see what already happened. That is the unglamorous part, and it is most of the product. The model sits inside that workflow.

Investors stall at the offer. They can find a property and still not know what to send, so nothing gets sent. I built a tool that takes a property and produces three viable offer structures, including seller financed terms, then generates the letter of intent.

That tool is in daily use by paying members of my education company, Real Estate 101. We added the light CRM on this one. The customer feedback surfaced corner cases we had not considered, and made the tool much better.

The note I wrote down for myself was simple. Pull in customer feedback sooner instead of building in a vacuum. The corner cases users surface are what make the tool.

A clean demo will hide the cases that matter. A paying user will not.

The call-transcription agent was the one that taught me this the hard way. Hundreds of hours of recorded coaching calls held answers nobody could find. The knowledge existed and was functionally unsearchable. I built a Whisper transcription pipeline feeding an assistant that answers from what was actually said on those calls, with the source. Retrieval over real proprietary material, not a chatbot bolted onto a FAQ.

It was live. People used it. Then billing on the backend failed, transcription stopped, and the assistant had nothing new to retrieve. No alarm fired. It sat broken for a while before we caught it.

That lesson is now the first thing I build into every scheduled job. Fault monitoring on day one. In a retrieval system the failure is easy to miss. The answers start to look stale or thin, and it is easy to blame the model when the job that feeds the model already died.

This is the same habit I had in network engineering, applied to a smaller stack. You try not to learn that a circuit is down from a customer ticket. You instrument the job, and you decide in advance what "broken" looks like.

When I build AI agents on a company's own data, I want them scoped so you can see what they did and undo it. That is rollback. That is blast radius. An agent that can draft an offer, update a record, or send a message is useful only if you can bound what it is allowed to touch, and reverse it when it is wrong.

I do not treat that as optional. In a real estate business, a wrong offer or a wrong file in the wrong place creates work for humans who are already busy. These tools should take typing off the operator's plate. They should not create a new mess to clean up.

Change control sounds formal when you say it out loud. In practice it means you know what shipped, you know what it can touch, and you have a way back.

I am an operator first. I have a computer systems engineering degree from the University of Arkansas, and those years in networks, but I do not show up as a professional software engineer. I show up with a process I already run, and I build the smallest tool that takes time out of it.

If I were sitting with you, I would ask which process eats hours every week, and whether you own it well enough to judge the output. Then I would ship a narrow tool to the people who already do that work, and I would put an alarm on the job that will fail silently. I keep the experiment log, with honest statuses (idea, building, live, or shelved), at pauldavidthompson.com/experiments. The notes on what I'm building, what worked, what broke, and what it cost live on that site.

If you are attaching a model to a real workflow, build the alarm first. I learned that in networks, and I learned it again when a billing issue stopped a transcription job I thought was still running.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @paul david thompson 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/the-ai-job-that-fail…] indexed:0 read:5min 2026-11-20 · —