Consequence
What actually happens in the business when the action lands.
The future of enterprise AI is not full autonomy.
It is governed autonomy.
Most AI conversations still start in the wrong place. They begin with the model, or the agent, or the tool, or the impressive demo. But enterprise value does not come from an agent doing something clever in isolation.
It comes from AI being placed into the operating layer of the business, where real work happens, real decisions are made, and real consequences follow.
The first wave of enterprise AI helped people produce content, summarise information and answer questions. Useful, but largely assistive. The next wave is different.
AI is moving into workflows. Into approvals. Into operational monitoring. Into case handling. Into client service. Into investment diligence. Into finance, risk, compliance, sales, marketing and delivery.
The question is no longer can it do the task. It is what authority it should hold.
That is an operating-model question, not a technology question. A serious enterprise cannot treat autonomy as a switch. It is not "human in the loop" versus "fully autonomous". That framing is too crude for real organisations.
The practical model
A ladder of authority.
Level 01Observes and reports
Level 02Recommends
Level 03 · HumanPrepares for approval
Level 04Executes within limits
Level 05Notifies after action
Level 06Acts autonomously
Above level three, the system acts before a human sees it. That boundary is the whole design problem. Only in narrow, well-governed domains should anything reach level six.
This is what we call governed autonomy.
Calibration
What sets the level.
Reversibility
Whether the decision can be undone, and at what cost.
Quality of evidence
What the system knows, and how well it knows it.
System confidence
How certain the agent is, and whether that certainty is earned.
Value at risk
The money, the client, the licence, the reputation.
Human accountability
Whether it is clear, in advance, who carries the decision.
In practice
Same technology. Very different operating rights.
| The action | The consequence | The operating right |
|---|---|---|
| Marketing draft | Low, easily reversed | Light review |
| Customer refund | Financial, bounded | Threshold-based approval |
| Compliance exception | Regulatory | Escalation |
| Financial transaction | High, hard to reverse | Authority, audit and control |
| Infrastructure remediation | Operational, wide blast radius | Prepared by AI, approved by a human |
Where programmes stall
They prove the agent can do the work, then never define the authority around it.
Who holds the authority?
- Who owns the decision?
- What can the AI change?
- When must it pause?
- Who approves the action?
Can the business prove it?
- What happens when evidence is weak?
- How is the action logged?
- Can the decision be replayed?
- Can you prove who was accountable?
Without those answers, AI remains stuck at the edge of the organisation: useful, impressive, but not trusted with meaningful work.
The design
Five things every AI-enabled workflow needs.
A clear business outcome
Not vague productivity. A named workflow, a measurable baseline and a defined improvement target.
Defined agent roles
The AI must have a job, a boundary and a scope. Not a general-purpose assistant wandering around the business.
Explicit edge authority
Humans should not review everything. That defeats the point. But humans must approve the actions that carry consequence, judgement, risk or irreversible impact. We named this in essay 001.
Telemetry
The business needs to see what happened: what the AI saw, what it recommended, what it changed, what it escalated, what it cost, and whether the outcome improved.
Continuous optimisation
The system should learn from traces, approvals, exceptions and outcomes. Not by becoming uncontrolled, but by becoming more reliable, more measurable and more useful over time.
A tool is used when someone remembers to use it. An operating capability is built into how the business runs.
Most organisations do not need more disconnected AI experiments. They need a controlled path from pilot to production. They need AI that can sit inside real work, with the right permissions, evidence, approvals, escalation and accountability.
The winners will not be the companies that give AI the most freedom. They will be the companies that give AI the right freedom.
Authority
Enough to improve the work.
Control
Enough to protect the business.
Evidence
Enough to build trust.
Judgement
Enough to keep responsibility where it belongs.
Where Praxis stands
This is the future of enterprise AI: not artificial intelligence floating above the organisation, but governed intelligence embedded into the operating layer. It is the same argument we have been making since essay 001, now with the authority question answered.
That is where AI stops being a demo. That is where it becomes capability.
Point of View
← All essays
Previous
The agent operating-system problem →
Begin
Where should the line sit?
One conversation, no pitch deck. Bring the workflow you cannot yet trust to AI, and we will map the authority around it.