Agentic AI
The marginal cost of intelligence is falling sharply, and that changes not only the price of existing tasks but the boundaries of what can be done at all. The gain arrives when the work around the technology is reorganised.
Contact usEffect that can be measured
Falling marginal costs move the boundary of the possible
When the price of intelligence falls towards zero, what changes is not only the cost of existing tasks but the range of possible ones. Work that could never pay for itself becomes viable. That is where the strategic value sits, and it is found by systematically examining the work, not by demonstrating the technology.
An agent is software with room to act, and room to act requires governance. We scope the task, define the data access and measure the effect against a baseline before anything goes live. Logging and traceability are not bureaucracy but the precondition for a use that can be explained, defended and improved.
Productivity often dips before it rises, because the organisation has to learn to use the new capacity. That adoption is part of the work: roles, workflows and accountability are reorganised with the technology. The EU AI Act sets the frame, and we treat it as a design requirement, not an after-the-fact audit.
Where the field is moving
From chat to agents that do work
The past year's shift is not better answers in a chat window. It is software that performs tasks: reads an inbox, looks things up in three systems, fills in the fourth and reports back when something needs a human. Standardised protocols for tool access have made it practical to give a model hands, not just a mouth.
Value moves accordingly, from the conversation to the workflow. The question is no longer what the model knows, but which tasks it may perform, with what data access, and how the result is checked. An agent is an employee with perfect memory, high speed and no judgement beyond what has been built in.
So we start with the task, not the technology: bounded, measurable, and with a human who owns the outcome. The organisations getting the most from agents are not the ones with the most pilots, but the ones that chose few tasks and finished them.
Evaluation is the new quality assurance
A language model cannot be unit-tested. It can be evaluated: against a set of real tasks with known answers, rerun at every change of model, prompt or data. Without that foundation every update is a gamble, and an improvement in one place can quietly break another.
This changes the way of working more than most expect. The most important deliverable in an AI project is often not the solution but the measurement basis: what was the starting point, what improved, what got worse, and is the difference worth the money. Without a baseline there are only anecdotes.
Meanwhile the price per task keeps falling, and smaller models now solve tasks that required the largest a year ago. That opens the door to running sensitive workloads on your own infrastructure and to choosing the model by the task rather than by the advertising. Here too, the answer is arithmetic, not creed.
The AI Act has moved from notice to reality
Obligations for general-purpose AI models have applied since August 2025, and from August 2026 most of the requirements for high-risk applications take effect. Now is when the classification has to be in place: which applications do you have, which category do they fall into, and what follows in documentation, oversight and control.
For most companies the biggest task is organisational rather than legal: a register of applications, an owner per application, logging that can show what the system did, and human oversight that is real rather than a name in a document.
We treat the regulation as a design requirement on a par with security: built in from the start, not retrofitted before an inspection. It is cheaper, and it produces better systems. Governance that works is not the brake on adoption. It is the precondition for scaling it.
Four steps, from starting point to operations
Optimism, not credulity. Large opportunities require strong execution and an understanding of the risk the technology also creates.
Niels Reinau
Niels founded iCEO after a career spent at IBM, at eBay, and running platforms at DanDomain and Zitcom, two of the largest hosting companies in Denmark. Zitcom is today team.blue Denmark, and DanDomain is one of the brands it still trades under. He works alongside consultants who have been in this industry for twenty-five years. You get that experience directly, not a partner at the pitch and a graduate on the work.