The Firm, Rewired: Part two
Management
How do we move fast without burning out our people, or losing control?
Bolted onto today's org chart, agents create more review work than they remove. The gains come when the workflow is redesigned around your people.
Humans in Mind
Agents never sleep. They generate drafts, fixes, experiments and recommendations around the clock, which sounds like pure upside, until you meet the humans who have to check it all. We have felt this directly in our own years of working with agents: the productivity gain is real, and so is the management burden.
Our stance is AI with humans in mind. Agents should absorb effort, not transfer it, lifting load off your reviewers rather than piling it on. Design the workflow around what people do best: judgement, exceptions, relationships, accountability.
None of this is a reason to avoid agents. It is a reason to adopt them with a method and an architecture, not just enthusiasm.
Burnout is a real risk
Not because the AI is struggling, but because the humans are drowning in output. As generation becomes abundant, review becomes the constraint.
Vigilance decays
When output is continuous, fluent, and well-formatted, reviewers habituate. They scan rather than read. They approve rather than judge, the same failure mode seen in semi-autonomous driving.
Relationships thin
As employees collaborate more with agents and less with each other, research measures a rise in workplace loneliness. Fewer social exchanges, less informal coordination, more solitary execution.
"Workslop" compounds
Over 40% of workers in larger firms have encountered low-quality AI output costing nearly two hours of rework per instance, productivity gains from one person instantly becoming overhead for another.
Building the Infrastructure
Context as Infrastructure
For an agent to do real work in your business, it needs what a good new hire needs: the priorities, the rules, worked examples, access to the right systems, and someone to ask when it is unsure. That operating knowledge is the infrastructure, and in most firms today it lives in scattered prompts and one person's chat history.
The early web went through the same messy phase before it found its standards, and the firms that built the plumbing first set the pace for everyone else. The same is happening now around agents. Being mid-transition is no failure of ambition. Staying there, once the standards are visible, is a choice.
Two Kinds of Agent
Three-quarters of executives now say they see agents as co-workers rather than tools. That instinct is half right. There are two fundamentally different kinds of agent, they need different management, and confusing them is one of the fastest ways to waste money.
Most firms need both. The agentic company manages both, and knows which is which. Below are the two patterns we build against every day.
Factory Agents
Run the same workflow hundreds or thousands of times a day against a defined policy, triage, quality reviews, invoice matching, claims intake. Frameworks like Google ADK or LangGraph orchestrate steps, decision points, and escalation rules.
The learning loop is institutional: escalations and errors feed back into tighter policy, sharper boundaries, new rules.
Staff Assistants
Each interaction is different, drafting a proposal, analysing data, debugging code. Tools like Claude Code or Cowork operate here. You build skills, connect systems, and teach preferences through use.
The learning loop is personal: accumulated history, learned preferences, expanding access. The assistant improves because it has worked with you.
What Changes, Department by Department
The pattern is the same everywhere: agents absorb repetitive cognitive load, humans move to judgement, exceptions, and relationships. We have documented five concrete workflows where agents are already delivering measurable results. The gains only materialise when you redesign the workflow, not just add an agent to it. We cover that in detail in Workflow Redesign for the Agentic Era.
Operations
Scheduling, allocation, compliance checks, exception flagging, the team moves from executing routine decisions to managing the system that executes them.
Customer Service
Becomes escalation-only. Agents triage, match patterns, draft responses. Humans handle cases needing empathy or authority.
Sales
Gains leverage on preparation, prospect research, meeting briefs, pipeline tracking, buying back time to actually think about the client.
Finance
AI triages or prioritises work, not executes it alone, invoice matching, reconciliation, exception handling.
Marketing
Agents generate drafts and variants. Adapt to a world where AI is increasingly the audience, not just the tool.
Regulatory
Agents monitor and flag for quality control; accountability stays human. The eval framework matters most here.
The Accountability Trifecta
When an agent acts and something goes wrong, who answers for it? That question decides whether your governance works. As agents become more capable, you will hand them real rights: to retrieve information, draft messages, run workflows, even trigger actions within defined limits.
Every right you delegate creates a responsibility around it. Someone has to own the rules, the thresholds, the tests and the monitoring that keep each agent inside its lane.
The strategic work is precision: which rights move to agents, what responsibilities get built around them, and which named person stays accountable when the system fails. We have explored what this looks like in practice.
Agents should absorb the load, not pile it on. We design the workflow around your people.