On LinkedIn I keep reading that agentic coding makes teams 10x more productive. It is good marketing. It has little to do with what we measure in real projects.
In episode 2 of Prompt & Proper, Steffen Ochsenreither from Swiss Post and I take that number apart. And we talk about the other side of the same coin, the one that gets far less airtime: what tokens actually cost, and how solid the economics behind them really are. The episode is in German, and this post captures the core of the conversation.
What Is Actually There: 20 to 30 Percent#
Steffen’s numbers match ours. On simple tasks, documentation being the prime example, you comfortably reach 80 percent efficiency gain. On complex tasks, where you have to think and describe multi-layered workflows, it is closer to 20. Averaged across everything you land at 20 to 30 percent. That is a lot. It is just not 10x.
Coding Is the Smallest Lever#
The more interesting point: coding itself accounts for maybe 10 percent of a full value chain. The rest sits before and after it, in requirements engineering and ideation on one side, in quality assurance, deployment and operations on the other.
Steffen makes a good case for the upstream side. Every requirements engineer works differently, one on paper, the next in PowerPoint, the third on a Miro board. Agents give you a standardized output there and ask questions along the way that nobody else would have asked. Better input speeds up the coding downstream. Writing lines of code is something any AI can do. Translating requirements cleanly is the actual craft.
Anyone who only accelerates the coding creates pressure at both ends. Upstream the requirements cannot keep up, downstream the pile of things to review and test grows. And more output does not mean more value. Steffen puts it plainly:
“Shit in with agentic is not just shit out, it is shit out squared. It just gets more, it does not get better.”
The Wrong KPIs#
Most companies measure the impact of agentic coding badly. Lines of code, tickets closed, token consumption. Steffen’s verdict is blunt:
“Token consumption is not a KPI. Number of pull requests is not a KPI. Number of tickets closed, definitely not a KPI.”
Flooring the accelerator does not get you there faster. The relevant question is: how much of the generated code survives in production?
50 Dollars in One Hour#
Since 1 June, GitHub Copilot bills on a token basis. We saw it coming and built monitoring in time. The Monday morning view at eight was sobering anyway: individual developers had burned 50 US dollars of tokens in a single hour. Extrapolate that to a company our size and across a year, and you get numbers where you might as well close up shop.
For companies that do not monitor this, I see real exposure. The invoice arrives, and by then it is too late to react.
The Math Behind It Does Not Add Up#
Steffen brought numbers. Pushing “War and Peace” through a model, roughly 300,000 tokens, costs about 15 cents today. Absurdly little. But models keep getting more complex and consume proportionally more tokens, even as the price per token falls.
The bigger picture looks like this. This year alone, over 700 billion US dollars of investment flows into AI, around two billion a day. Add the infrastructure, power plants and everything attached to it, and you reach roughly 1,400 billion. Against that sits a global revenue of about 150 billion. With around 1.3 billion AI users worldwide, most of whom pay nothing, you would need roughly 600 dollars of revenue per user. Nobody pays that.
We both expect this bubble to burst. I had expected it earlier. The SpaceX IPO could pull a lot of capital out and take some share prices with it. Companies will go bankrupt and money will be gone. Others will survive and carry into the next era. For token prices this means one thing only: they go up.
A Platform as a Gateway#
In my role as Group Chief AI Officer at Zühlke I am building an AI platform, effectively a gateway. Governance lives inside it, and every request passes through. From there it gets routed, to our on-prem models, to cheaper models, or to frontier models when the task genuinely calls for it. For a consulting firm there is the added need to allocate costs properly, and you only get that right with a platform like this.
Steffen adds the point that closes the argument: the gap between the cheap models and the frontier ones is often only about a month. If you do not always need the newest thing, you work perfectly well with the smaller models. The hard part is deciding that day to day. I notice it in myself. I often reach for Opus and afterwards wonder whether this particular text really needed it. But:
“I have no appetite for rework, and I have no time for rework either.”
The Move to Local Hardware#
A lot of our engineers are currently assembling their own rigs, graphics cards plus a mainboard, and running their agents locally. Nvidia announced a chip together with Microsoft that runs LLMs on your PC. I can well imagine that as soon as next year we see agent swarms on laptops and phones, with the data center reserved for the genuinely large tasks.
The bottleneck here is not hardware, it is trust. I set up an agentic system in a secured environment and then stood there asking what I would actually do with it. For my knowledge management it would be excellent. But I do not trust it enough yet and want to steer it too tightly. Behind that sits an unlearning and re-learning of how you work. That happened to me when I moved from the browser to ChatGPT and from there to Claude Code. How much time that takes is regularly underestimated.
Review Becomes the Bottleneck#
If more gets shipped, more has to be reviewed. Our senior developers increasingly complain that they mostly review code their colleagues produced with AI. After documentation, reviewing is the least loved part of product development, because every time you have to think your way back in.
Steffen describes a case where both the product and the test cases were built with AI. All tests green, the product still did not work. A model wants to please, and a quality gate built on the same assumptions as the product checks nothing.
This is exactly why the software engineering discipline does not disappear. A good developer can read and understand code no matter how it came about. That stays.
The Take-Away#
Look end to end instead of only at the coding. Pick the right model for the right task, even against the reflex to always grab the newest one. And keep an eye on token costs before the invoice arrives.
If you enjoy the episode, subscribe to Prompt & Proper on Spotify, Apple Podcasts, Amazon Music or YouTube and tell us which AI topic you are wrestling with right now. All episodes: promptandproper.ai.