Trending Topics
Trending
AI reality check: Why infrastructure, not intelligence, will decide the winners
This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.
Ask three executives what “agentic workflows” means, and you’ll get three different answers. Ask about “AI governance” and the gap only widens. Years into the AI craze, the industry still argues about its own vocabulary, and that’s not confusion for its own sake. It’s roughly where things actually stand.
Something has shifted in the last six months, though. The market is sobering up. The conversation used to circle around what AI could theoretically do. Now it’s about what it’s actually doing, and whether the bill matches the result.
Everyone is converging in the same direction: agentic systems, software that stops waiting for a prompt and starts acting on its own. We’ve lived through a version of this shift before, decades ago, when cloud migration forced every company to rebuild itself around cloud infrastructure it barely understood. Everyone knew they had to move. Almost nobody knew how to do it without breaking the systems still keeping the business alive.
That collision between ambition and everything already in place is where AI sits today. It’s surfacing five problems, and none of them shows up in a product demo.
The adoption wall
Most AI transformation projects rarely die in engineering. They die the week someone actually has to use them for real. A company spends millions building a sophisticated agent, then watches employees quietly route around it and go back to the spreadsheet.
The reason is almost never that the code is bad. It’s that the agent asks people to give up a workflow they already trust for one they don’t trust yet. If a team doesn’t trust what you built, or the rollout feels like another mandate handed down from above, they’ll ignore it politely and move on, and that’s not to mention other factors like fear of losing a job or a sense of losing one’s own value. Deploy the best model on the market. If nobody adopts it, the return is zero.
And the effect isn’t evenly distributed. The bigger the enterprise, the bigger the pushback, and the more visible the botched rollout. Large, established companies need to face something uncomfortable: startups and newer tech companies are already miles ahead, mostly because they have nothing legacy to unlearn. Stay behind long enough, and there’s a real chance you never catch up.
The governance gap
A year ago, an AI inventing a statistic was a screenshot people laughed at over lunch. Today, that same hallucination showing up in a financial report or a customer contract has your company’s name on it.
Which is why governance can’t be a policy document nobody reads. It has to answer one specific question before any system goes live: exactly where does a human step in? AI can produce an answer or decision fast. It can’t be held responsible for one. Every serious deployment defines, in writing, the point where a person reviews, signs off, and owns what happens next. Skip that step, and you haven’t deployed AI. You’ve deployed a liability with a good interface.
The access nobody wants to give
Security is where enthusiasm runs into the wall fastest. Enterprises are right to be careful about proprietary data flowing into public models, which is why private and on-premises deployment has become the default request in every serious RFP, not the exception.
But be honest about where the industry actually stands: still uneasy. Running a model behind your own firewall takes real infrastructure investment, and just defining what a semi-autonomous agent is allowed to read, write, or touch inside your database can eat up months on its own. And the controls over how data provided for training such models is used and interpreted are still non-existent. Everyone wants the efficiency AI promises. Far fewer are equally enthusiastic about the access it needs to deliver it.
The pricing model everyone outgrew
The economics are shifting faster than most budgets can follow. Pricing built purely around tokens is losing relevance to the value a task actually produces, and vendors are already testing pricing per task instead. Good. It’s overdue.
The harder problem is routing. Companies are realising they can’t justify firing an expensive, reasoning-heavy model at a question a cheaper one could answer just as well. Engineering teams now have to build that judgment into the system itself: send the simple query to the cheap model, save the expensive one for the problem that actually needs it. Skip that architecture and costs don’t creep. They compound.
The question with no clean answer
Last, the one nobody has resolved yet: if an AI writes a piece of critical software, who owns it? And when a semi-autonomous agent executes a task that costs a client money or exposes a flaw, who answers for it? The model provider? The vendor who built the agent? The company that deployed it?
Nobody has a tidy answer yet. What is clear: “the algorithm did it” has never once worked as a defence, and it isn’t about to start.
The companies that actually win this
Here’s the part that never makes it into a keynote: the model isn’t what’s holding companies back anymore. It’s everything around it: the governance, the process, the access controls, the cost discipline that most organisations still haven’t built.
Models will keep getting better whether anyone is ready or not. Without the infrastructure to run them responsibly, that improvement doesn’t turn into anything usable. It just sits there, expensive and impressive, doing nothing for the P&L.
The AI bubble isn’t bursting. It’s maturing, and maturing is slower and far less glamorous than either the hype cycle or the backlash would like it to be. The companies that win the next five years won’t be the ones who moved fastest or talked loudest about transformation. They’ll be the ones who treat AI like what it actually is: serious infrastructure that demands real management, real oversight, and a relentless focus on whether it’s actually making the business better, not just busier.
Related articles
About The Author
