Agentic AI didn’t happen in 2025 and won’t happen in 2026…so how far away is ROI?

Last week, The Information ran a long, unusually candid piece about Anthropic’s new “agentic” tool, Cowork. It’s one of the rare stories that doesn’t just repeat the promise — it documents the experience.

And the experience matters.

The reporter tried to use Cowork for two very normal tasks:

automate a small “word-a-day” app he’d wanted for years
offload routine weekly research

What followed was not automation. It was work.

He wrote a 352-word specification explaining the task and files.
The system stalled, timed out, and reset itself.
It silently switched models mid-session, forcing him to start over.
It asked configuration questions he didn’t understand.
It required Terminal access, debugging, copy-pasting error messages back and forth.

Setup alone took over an hour.

After two hours, the app finally ran — but:

definitions were missing or wrong
requested example sentences never appeared
outputs were incomplete

In his words: “I sure was in no mood to smother Cowork in kisses.”

Other users reported the same pattern:

task lists that quietly ignored half the work
“comprehensive” outputs that weren’t
systems that had to be corrected, pressed, supervised

One researcher described using it as “like onboarding a new employee.”
Another said he’d need to “budget a lot more time” to make it worthwhile.

And that’s the line that matters most:

“It’s hard to imagine the product taking off with a general audience until it works more smoothly and offers a gentler learning curve.”
— The Information, Jan 31, 2026

That sentence isn’t about UX.
It’s about economics.

If a tool requires:

patience
technical literacy
ongoing supervision
debugging and correction

…it isn’t saving time for most people. It’s consuming it.

Yes, the reporter eventually shaved some time off a weekly task — after hours of setup, iteration, and learning. That’s not labor replacement. That’s deferred payoff for expert users willing to invest upfront time.

Which leads to the uncomfortable conclusion the article never quite states:

Agentic AI doesn’t fail because it isn’t smart enough.
It fails because it still requires too much human labor.

Until these systems work out of the box — without specs, terminals, babysitting, or cleanup — ROI stays narrow, slow, and expert-only.

The question isn’t when agentic AI arrives.
It’s how long companies can carry the cost before it actually pays back.

Right now, based on the most honest reporting we’ve seen, that answer is: not soon.