Operator's Lens
What is actually happening in enterprise AI. Five years running product and product marketing inside a global software company, paired with six months building, selling, and delivering agentic solutions with my own money at risk.Both vantage points led me to the same conclusion.
Inside the Machine
What five years at enterprise scale taught me
1
Better tools never made the enterprise faster.
We had modern collaboration tools and access to internal enterprise tech. Neither moved the operating rhythm. The bottleneck was people: how long it took someone to read something, decide what it meant, and loop in the right person across time zones and calendars.
2
Projects die in three places: alignment, perfection, and hearing every voice.
It was never the technology or the talent. I watched people with real execution ability check out or get quietly reassigned before the org finished deliberating.
3
Speed is manufactured.
I treated visible progress as the only metric that mattered. Ship something small, let people see it move, and don't let a project sit in one phase for weeks. By the time a room full of stakeholders sat down, I had already mapped the real options and pushed toward a decision.
4
First and imperfect beats finished and late.
Shipping at 80 percent put us in front of accounts the polished version never reached, and being early meant better data on what buyers wanted while competitors were still refining. Cultures that wait for perfect give up ground they never get back.
5
The executive job is judgment. Everything else is mechanics.
Who to call, what channel to use, how to deliver news people do not want to hear, when to push and when to sit still. The actual chasing, summarizing, and reconciling belonged to my staff, and that is the layer AI is eating right now.
Outside the Machine
What six months of building with my own money proved
6
Agents compress production. Coordination stays human.
Human coordination is not going away in the near term. That is why engineers are being deployed into the field: enterprise time is consumed by activities agents do not touch.
7
Demo speed is not delivery speed.
An audit runs in seconds. A website spins up in minutes. Anything deterministic and repeatable comes out just as fast. Even with a swarm of agents, a content platform or a real application still takes months or longer to build because human intervention is required. The work that compresses is the work that was already predictable, and most of what actually happens inside a company is not.
8
Personal grade and production grade agents are different species.
Getting an agent to work for me proved little about running it in production, where maintenance, model updates, and real workflows take over. A personal agent runs on my permissions, whereas a production agent needs its own role, privileges, and accountability. Something that worked well a month ago can be obsolete today.
9
Enterprises are buying AI and not absorbing it.
You cannot take an organization from almost no automation to full automation in one jump. The data, the processes, and the people all need a ramp, and almost nobody is selling one. In my own pipeline, the security and compliance objections were usually an organization admitting, in the only language it is allowed to use, that it was not ready.
10
Autonomy is an open secret.
Everyone says autonomous on stage, because that is what decks and multiples reward. Everyone in the room also knows agents are barely used outside engineering, and underused even there. The budgets prove it: agents draw roughly 25 percent of AI spend. That gap between the pitch and the reality is exactly where the money is quietly moving.
Convergence
The same discovery, made twice, from both sides
The binding constraint on enterprise AI is absorption, not capability.
I lived it from multiple fronts. First as the operator drowning in coordination, then as the vendor watching my own deployments stall. Builders ship at five times the old pace. Organizations absorb at one. Nearly everyone is selling capability into a bottleneck that has little to do with capability.
The gap gets monetized by whoever can work in it without saying it out loud.
No vendor can admit their agent needs a human babysitter, so the money moves quietly instead. Platforms turn the ramp into a feature. Trusted people sell it as a service. The forward deployed engineering trend and the expanding footprint of IT departments are both this same force, just wearing different clothes.
Conclusion
Shipping product is getting easier. The real work is getting it absorbed.
Every failed AI rollout I have watched, from inside a big company and from outside building my own, failed for the same reason. Capability showed up faster than the organization could take it in. The winners of this next stretch will not be the teams with the best model. They will be the ones who treat absorption as something you engineer on purpose: staged trust, visible progress, work that gets inserted into how people already operate, and outcomes as the metric that matters.
Absorption is an operating discipline, and it is the one I already know how to run.
Max Skalatsky | max@skalatsky.com