AI Agents in Production: What Survives Contact with Reality

‍

Every AI agent demo looks the same.

‍

Someone types a request. The agent plans, calls a few tools, and returns a finished piece of work. The room nods. Someone says the word "transformative."

‍

Then the demo ends, and the real question begins: what happens when nobody is watching?

‍

Production is where agents stop being impressive and start being useful. Or stop being anything at all.

‍

Here is what the demo never shows.

‍

The demo never shows the invoice with a scanned signature at a 45-degree angle. The client who replies in Swiss German, French and frustration in the same email. The API that times out at month-end, exactly when the work matters most.

‍

Demos run on the happy path. Businesses run on exceptions.

‍

We build agents for SMEs in finance, Treuhand and other regulated corners of the Swiss economy, places where "the AI made it up" is not an anecdote but a liability. After enough go-lives, a pattern emerges.

‍

The agents that survive production are never the smartest ones. They are the best supervised ones.

‍

A useful mental model: you are not installing software. You are hiring a very fast, very well-read junior employee with no common sense and no memory of yesterday, unless you build one.

‍

You would not give a junior employee the company credit card on day one. So the agent gets permissions, scoped narrowly.

‍

You would not accept "I just did it" as an explanation. So the agent writes an audit trail: every decision, every tool call, every source.

‍

You would not let them send a client letter unreviewed in their first month. So the agent drafts and a human approves, until the error rate earns it autonomy, one task at a time.

‍

None of this is exciting. All of it is the difference between an agent in production and a screenshot on LinkedIn.

‍

The uncomfortable truth: most of the work is not AI work.

‍

Retries, queues, logging, fallbacks, monitoring, access control. The model is maybe twenty percent of the system. The other eighty percent is the boring engineering that decides whether the whole thing survives a bad Tuesday.

‍

This is also why "we tried an agent and it didn't work" usually means "we tried a model without a system around it."

‍

Scope is the other quiet killer.

‍

An agent that does one process end to end, reading the incoming invoice, checking it against the order, booking it, flagging the exceptions, creates real value every single day.

‍

An agent that "helps with everything" helps with nothing in particular, and nobody notices when it is wrong.

‍

Narrow is not a limitation. Narrow is what makes trust measurable.

‍

And drift is real. The model gets updated. Your processes change. The client who always paid late gets acquired. An agent in production is not a project you finish. It is a colleague you keep managing.

‍

That is why we do not sell go-lives. We stay for the maintenance, because maintenance is where agents either compound or quietly rot.

‍

So, are AI agents ready for production?

‍

Wrong question. Plenty of them already are, quietly booking invoices, triaging inboxes and preparing filings inside companies that will never write a blog post about it.

‍

The right question is whether your process is ready for an agent: defined enough to delegate, measured enough to supervise, and honest enough to admit where the exceptions live.

‍

The technology is no longer the bottleneck. The clarity is.

‍

‍

Frequently Asked Questions

What is an AI agent, practically speaking? A software system where a language model does not just answer questions but does work: it reads context, decides on next steps, calls tools such as your inbox, your ERP or your document system, and produces an outcome. The difference between a chatbot and an agent is the difference between asking someone for directions and having them drive.‍

Why do most AI agent pilots never reach production? Because pilots optimise for the demo. They handle the happy path, skip exception handling, keep no audit trail and have no owner. Production requires permissions, human review steps, monitoring and someone responsible when the agent is wrong. Most failed pilots are organisational failures, not technical ones: the model worked, the system around it was never built.‍

How do you make AI agents safe for regulated industries? Four things, none optional: narrowly scoped permissions so the agent can only touch what it needs; a complete audit trail of every step, decision and source; human approval on anything that leaves the company, at least until trust is earned per task; and deterministic checks around the model for anything involving money, deadlines or law. The model proposes. The system disposes.‍

How narrow should a first agent be? One process, end to end, with an output you can measure. Invoice intake. Email triage. Meeting follow-ups. If you cannot describe what "correct" looks like for the task, the task is too broad for a first agent, and probably too vaguely defined for a human as well.‍

What does it cost to run an AI agent in production? Less than the employee it assists, more than a software subscription. The model usage itself is rarely the significant cost. Integration with your systems, the review workflows and ongoing maintenance are. Budget for an agent the way you budget for any system that touches your operations: a build phase, then continuous care, not a one-off purchase.

‍

‍

If you have thoughts, feedback, or questions, we'd genuinely like to hear them. Reach out directly to the author, Dejan Georgiev at:##INLINE1##