In partnership with
Ask a CEO whether AI agents are running parts of the business, and most will say yes. Ask what share of any single function actually has one deployed, and the number falls fast. McKinsey surveyed organizations and found that 88 percent now report regular AI use in at least one business function. Only 23 percent are scaling an agentic system anywhere in the enterprise, and in any single function the ceiling stays under 10 percent.
That's the real shape of the AI coworker story right now. Everyone bought in. Almost nobody scaled past the pilot. The gap between those two facts is where most of the online noise comes from.
The same McKinsey survey found that reported AI agent use skews heavily by industry, showing up most in technology, media and telecommunications, and healthcare. That doesn't mean those industries run agents at scale. It means people there talk about agents more, try them more, and show up in more of these surveys. Reported use and production use are two different numbers, and most coverage online blurs them into one.
The integrated coworker for AI native teams
Empower your team to do their best work with Adapt, the integrated coworker that works alongside your team in Slack and deeply understands your business.
Here’s how Adapt is different
Set up takes minutes: connect your tools, add to Slack, and it’s right there for anyone to tag @Adapt for help
Does real, high-ROI work: automates work on a schedule; builds internal tools with live data; and does complex, multi-tool tasks on demand
Learns your business as you work, becomes your company brain
Uses the best AI model for the task, not tied to a single provider
SOC2 Type II, RBAC, and support for personal and company-wide integrations
A familiar shape, if you've watched tech before
This isn't the first time a technology got ahead of its own deployment. Box's CEO Aaron Levie has compared this moment to 2010, when the entire tech industry agreed cloud computing was inevitable, long before most companies had actually moved anything into it. The conviction arrived years before the infrastructure did. Agents look the same way right now. Everyone agrees this is where work is headed. Almost nobody has finished building the version that works.
The gap between claim and use
MIT's NANDA group studied this same period and called it the GenAI Divide. Their number was blunt. Their research put the pilot failure rate at 95 percent, meaning most generative AI pilots never produced a measurable return. PwC's CEO survey in January 2026, drawing on 4,500 executives, found 56 percent report AI has given them neither more revenue nor lower costs so far. Gartner's CIO survey put actual deployed agent adoption at 17 percent once you strip out the pilots and the demos. These aren't fringe numbers. They come from the three research groups every enterprise buyer already reads.
Where it's actually working
Where agents do work, the gains are real. Stanford's HAI group reviewed a range of published studies this April and found productivity gains of 14 to 15 percent in customer support, about 26 percent in software development, and as much as 50 percent in marketing output. McKinsey found the heaviest agent use sitting inside IT and knowledge management, where a service desk ticket or a research summary is easy to hand off. The pattern holds across every source. Agents work best on narrow tasks with one clear right answer. Everywhere else, they still struggle, and the failure rate above is mostly built from those attempts.
The part nobody puts in the pitch deck
McKinsey ran a separate survey between December 2025 and January 2026 on trust and governance, gathering around 500 organizations whose leaders hold direct responsibility for AI risk and investment decisions. Their average governance maturity score rose to 2.3 from 2.0 the year before, which reads like progress until you see the rest. Only 33 percent of those organizations meet the bar for governing an autonomous agent, and two thirds named security as the main reason they haven't scaled further. An agent that acts on its own can make a wrong call before anyone reviews it. That's a different kind of risk than a chatbot giving a bad answer. A bad answer gets caught on the way out. A bad action is already done.
If you run a small team or a single product, none of this should scare you off agents. It should change how you pick the job you hand over. Look for work that's narrow, repeats often, and has one clear right answer. That's where the productivity numbers actually held up. Save the messy, judgment heavy work for a person, at least for now. The companies stuck in that 95 percent failure pile mostly skipped this step. They handed an agent ambiguous work and expected the judgment of a coworker who already knew the business.
Time Tools Built for This Era, Not the Last One
Most time tracking was designed before the AI that makes it work existed. Timeglass was built for it from the start - automatic capture, automatic attribution, timesheets that write themselves. Start free.
An agent isn't a coworker yet. It's closer to a fast intern who never asks for context and never says when it's confused. That's fine, as long as you remember which one you actually hired.
This newsletter won't close that gap for you either. Treat it as a standing check in, something that flags what changed and what's worth a second look before you'd find it on your own. The actual work still happens in your hours, on your own systems, after this email is closed. Read it often enough, over enough months, and it starts showing up quietly in the calls you make without having to think them through.
