US WORKING PAPER No. 1 · THE PROBLEM
When agents scale and people try to catch up
Agents and autopilots are arriving faster than organisations can see them, pay for them sensibly, or teach people to work with them.
Three kinds of “them”
| Word | What it does | Who decides each step |
|---|---|---|
| Assistant | Answers and drafts when a person asks | A person, every time |
| Agent | Takes a goal, plans steps, uses tools, may hand off to other agents | A person sets the goal and approves some steps |
| Autopilot | Runs a whole process on a trigger or schedule | A person owns the process and checks outcomes |
The problem isn’t any one of these. It’s the move along that line, from “a person asks” to “it runs on its own”, happening across thousands of people at once without anyone deciding it should.
The catch-up gap
Picture two lines. One is the number of agents and autopilots at work; it grows in steps. The other is people’s ability to work with them: knowing what exists, who owns it, what it costs and how to use it well. That grows slowly. The space between is where things go wrong.
- 75% of knowledge workers surveyed in 2024 used generative AI at work, and 78% of them brought their own tools. MICROSOFT & LINKEDIN 2024
- 46% of leaders surveyed say their companies already use agents to fully automate workflows or processes. MICROSOFT WTI 2025
- Gartner predicts at least 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028, up from 0% in 2024. GARTNER 2025
Risk
- Shadow AI. One in five organisations in IBM’s 2025 study reported a breach involving shadow AI; high shadow-AI use added about USD 670,000 to the average breach cost. IBM 2025
- No access controls. 97% of organisations that reported an AI-related breach lacked proper AI access controls. IBM 2025
- No policy. 63% of breached organisations had no AI governance policy or were still developing one. IBM 2025
- Agent-specific attacks. Goal hijack, tool misuse, memory poisoning, cascading failures and rogue agents. OWASP AGENTIC TOP 10
- Literacy duties. The EU AI Act’s AI-literacy obligation has applied since 2 February 2025. EUROPEAN COMMISSION
Cost
- Projects that don’t pay back. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027. GARTNER 2025
- Pilot purgatory. MIT NANDA reported about 95% of organisations it studied saw no measurable P&L return from GenAI. MIT NANDA VIA FORTUNE
- Seat sprawl, consumption surprises, duplicate agents, rework and agent washing — the person who builds an agent rarely sees its bill. OUR VIEW
Enablement
- The silicon ceiling. Frontline regular use has stalled at 51%. BCG 2025
- Training that doesn’t stick. Only 36% feel adequately trained; five or more hours, in person and with coaching, lift regular use. BCG 2025
- No cover from leaders. Only 25% of frontline employees say they get enough leadership support. BCG 2025
- Work-arounds. 54% would use AI tools even if not authorised. BCG 2025
Set up without…
01
An owner
Nobody answers for it.
02
An inventory
Nobody can list it.
03
A boundary
Nobody wrote down what it may do.
04
A meter
Nobody sees the bill.
05
A teacher
Nobody learns to use it well.
06
A shared place
Nobody sees the mistakes or the good ideas.
The fix isn’t to slow the technology down. It’s to make the work visible and shared. Start with the solution in plain words, then the framework.