Agents pay off when they work with the whole team An agent that serves only one person builds a new silo
What we've seenA supplier invoice that doesn't match its order touches six people, and it spends most of its time waiting between them.
What we think it meansAn agent that works for one person can't remove waiting it can't see.
What we're testingWhich tasks an agent can move from “ask first” to “do alone”, one task at a time.
Most AI agents today work for one person. You open a chat, hand it a task, and it plans the steps for you. That's useful, and it's also how a company ends up with a new silo.
Work in a business rarely belongs to one person. It moves between people, and most of its time is spent waiting.
We see three kinds of AI work
- In automation, code decides every step. It's fast and predictable, and it stops at anything it didn't expect.
- A single-player agent plans its own steps, for one person.
- A multiplayer agent works toward one goal with a whole team, under its own identity.
On our projects, most of the value has come from the third kind, because that's where the waiting is.
One invoice touches six people
Take a supplier invoice that doesn't match its purchase order. It touches payables, purchasing, the warehouse that received the goods, the supplier, an approver, and treasury.
Give each of them an assistant, and six assistants hold six partial pictures while the invoice waits between them.
Michael Hammer reported one insurer's estimate that an application spent 22 days in process and was worked on for 17 minutes. The waiting was the cost, and a single-player agent can't remove waiting it can't see.
Source: Damco's agent practice, 2026
A multiplayer agent needs five things
- It signs in as itself, with a named owner and written limits, so you always know whose account did what.
- It knows who's on the case and who may approve what, so it asks purchasing about the order and payables about the payment.
- It works from the same view as the people. The case, its history, and the next decision sit where both can read them.
- Its rules live in the tools, not in its prompt. A prompt is a request, and a check inside the tool runs the same way every time.
- Every step lands on the log, with who took it, a person or the agent.
Autonomy is earned one task at a time
Every new task starts on “ask first”: the agent drafts, and a person decides. When the log shows the agent gets a task right, that task can move to “do alone”. Some tasks stay with people for good.
We run one test before anything goes live. If every agent were switched off tonight, the business should lose only speed. No knowledge, decision, or record should leave with them.
Every agent needs a safe way into your systems
We build that once per client as Damco Switchboard. It has tools named in your business terms, sign-in for people and agents, and rules checked in every tool.
I wrote this up at more length in a guide on putting agents into production. It walks through all seven steps, from the first sticky note to the first live case.
- Michael Hammer, “Reengineering Work: Don't Automate, Obliterate,” Harvard Business Review, July–August 1990.
Tell us what you're seeing on your own programs
A strategist reads every reply and writes back.