How to use AI

Planted 02026-05-29

A new design material for building more productive organizations

Technology is neither necessary nor sufficient for productivity improvement.

The durable source of productivity is better organization design; a change in how work is divided and coordinated, how behavior is governed and controlled, and how information is processed into decisions and learning.

The new organization is an organization where some work is delegated to agents, some work is supervised by humans, and some work is restructured entirely because agents change the cost of coordination, analysis, documentation, monitoring, and reuse.

• Who does what?
• Who decides?
• Who coordinates with whom?
• What information is needed?
• How is work controlled?

With AI agents, the same questions remain, but the answer changes as agents can become part of the organization’s coordination and information-processing architecture.

Coordination: fewer handoffs, smaller queues, faster routing

In many organizations, productivity is lost because work sits in queues, gets handed off too often, waits for clarification, or requires excessive meetings.

AI agents can improve coordination by acting as routing, preparation, follow-up, and escalation layers.

Instead of this:

Person A asks Person B for information.
Person B asks Person C.
Person C sends a spreadsheet.
Person A makes a deck.
A meeting is scheduled.
The meeting identifies missing information.
The cycle repeats.

You get this:

An agent gathers the needed information, checks prior work, identifies gaps, drafts a recommendation, routes exceptions to the right owner, and prepares the decision meeting only if needed.

Move routine coordination work out of meetings, email, chat, and status updates into agent-mediated workflows.

Control: from supervision of activity to governance of outcomes and exceptions

The benefit is shifting management attention away from monitoring low-value activity and toward exceptions, judgment, and system improvement.

In a traditional organization, control often happens through:

• hierarchy
• approvals
• meetings
• reports
• budget reviews
• performance metrics
• process compliance
• managerial supervision

With agents, some of this control can become continuous, embedded, and exception-based.

Instead of managers asking:

“What did everyone do this week?”

Agents can answer:

“Here are the deliverables completed, decisions pending, risks emerging, commitments missed, customer issues unresolved, and places where human intervention is needed.”

Managers stop supervising activity and start governing decision quality, exceptions, priorities, and learning.

An improved control system could have three layers:

1. Layer 1: Agent-executed controls
   • Agents handle routine checks:
     • Is the request complete?
     • Is it within policy?
     • Does it match prior precedent?
     • Is there a known risk?
     • Is the data internally consistent?
     • Does this require legal, finance, security, or executive review?
2. Layer 2: Human exception review
   • Humans intervene where there is:
     • ambiguity
     • high risk
     • ethical judgment
     • customer sensitivity
     • strategic tradeoff
     • irreversible commitment
     • reputational exposure
3. Layer 3: Governance of the agent system itself
   • The organization also needs to govern the agents:
     • What can agents decide?
     • What must they escalate?
     • What data can they access?
     • Who owns their outputs?
     • How are errors detected?
     • How are models, prompts, tools, and permissions audited?
     • How is performance measured?

Without clear decision rights, accountability, and governance, agents can amplify confusion.

From individual expertise to institutionalized intelligence

Much knowledge work today is wasteful because organizations repeatedly recreate context:

• writing similar memos
• rebuilding analyses
• rediscovering prior decisions
• asking the same experts the same questions
• searching across scattered documents
• making slides instead of decisions
• producing analysis that is not reused

Instead of knowledge living mainly in people’s heads, old decks, Slack threads, emails, and shared drives, agents can help create an active organizational memory.

The future knowledge-work organization may have agents that:

• retrieve prior work before new work begins
• generate first drafts using company context
• compare new problems to past cases
• maintain decision logs
• convert meetings into actions and reusable knowledge
• identify contradictions across documents
• summarize expert judgment
• update playbooks after projects
• create templates from repeated work
• make best practices easier to follow

Knowledge management moves from passive repositories to active agents embedded in work.

The new division of labor

What work should humans do, what work should agents do, and what work should be redesigned away entirely?

Agents should not merely be added to existing jobs. Jobs should be decomposed.

An “analyst” role may become:

Decision architectures

What decisions are made in this process, what information do they require, who has authority, and which parts can be prepared or executed by agents?

For each workflow, you can map:

The productivity gain comes from:

• fewer handoffs
• faster triage
• clearer ownership
• better reuse of prior cases
• less meeting time
• faster escalation
• better organizational memory
• improved learning loop

New roles

Before and after

Organizations often rely on:

• meetings for coordination
• managers for status tracking
• analysts for information synthesis
• employees for repetitive documentation
• shared drives for knowledge storage
• approvals for control
• committees for cross-functional decisions
• memory residing in individuals

The AI-agent-enabled organization relies more on:

• agents for routine coordination
• exception-based management
• human judgment for high-value decisions
• active organizational memory
• clearer decision rights
• automated monitoring and escalation
• reusable knowledge assets
• smaller teams with more leverage
• continuous learning loops

The goal is not fewer people doing the same work. The goal is less organizational friction per unit of output.

A vendor can improve model alignment, but you still own deployment alignment

Fake productivity looks like:

• faster reports with subtle inaccuracies,
• fewer staff but more unresolved exceptions,
• higher utilization but lower resilience,
• lower service cost but worse customer trust,
• faster maintenance planning but more safety risk,
• smoother dashboards but less ground truth,
• apparent compliance but weaker accountability.

AI is not a vendor promise, it is an operating discipline.

Define autonomy levels

Use least-privilege tool access

Use approval gates

Test against failure modes before deployment

For an inventory agent, test: stockout, bad vendor data, duplicate SKU, promotion spike, weather event, supplier delay, return surge.

For a maintenance agent, test: noisy sensor, conflicting readings, stale data, safety-critical anomaly, known failure pattern, unknown pattern.

For a customer agent, test: angry customer, refund abuse, legal threat, discrimination complaint, safety issue, ambiguous policy.

Use separation of duties

Monitor behavior after deployment

Install tripwires and kill switches

Keep humans meaningfully in the loop

Bad human-in-the-loop design:

Agent makes 500 recommendations. Human clicks approve because there is no time.

Good design:

Agent handles routine cases automatically within narrow bounds; ambiguous or high-impact cases are escalated with evidence and alternatives.

Own the data context

Stage deployment