Models.
The first wave of enterprise AI was about which model to buy.
Enterprises do not have an agent-building problem.
They have an agent operating-system problem.
Most organisations are now past the question of whether AI agents can be useful. The real question is whether they can be trusted, governed, observed, escalated, improved, and operated inside the enterprise. That is where the market is moving.
Copilots.
The second wave put an assistant next to every seat — and left the work itself unchanged.
The agent operating system.
Not one product. Not one vendor. Not another chatbot layer.
An agent operating system is the set of capabilities that allows AI agents to work safely and usefully across a real organisation.
Where it fails
The gap between prototype and production.
This matters because most enterprise agent initiatives are currently failing in exactly that gap. The demo works. The pilot impresses. The business case is plausible. Then the hard questions arrive.
Q·01
Who owns the agent?
Q·02
What data can it access?
Q·03
Which tools can it use?
Q·04
What happens when it is unsure?
Q·05
When does a human approve the action?
Q·06
How do we monitor what it did?
Q·07
Can we replay the decision?
Q·08
Can we prove compliance?
Q·09
Can we stop it safely?
Q·10
Can we improve it without breaking trust?
These are not model-selection questions. They are operating-system questions.
The convergence
The major enterprise platforms are now converging around this reality. Microsoft, Google, ServiceNow, AWS, Salesforce and others are all moving toward governed agent infrastructure: registries, gateways, identity, policy enforcement, orchestration, memory, observability, evaluation, and human-in-the-loop controls. That tells us something important.
The future enterprise AI battleground is not "who has the best model?" It is "who controls the agent estate?"
Agent sprawl.
If each team builds its own agents in isolation, the organisation quickly inherits a fragmented, opaque and risky agent estate:
- Different tools
- Different permissions
- Different memory stores
- Different evaluation methods
- Different escalation rules
- No single view of operational risk
That is not transformation. That is another uncontrolled technology estate.
An operating model, from the beginning.
A serious enterprise agent strategy needs five layers — designed together, owned deliberately, and measured continuously.
- Every agent known, registered, accountable
- Controlled access to tools and systems
- Clarity on what agents know and remember
- Human authority designed in, not bolted on
- Evidence of what agents did, and whether it worked
The stack
Five layers of a serious agent strategy.
Agent inventory and identity
Every agent should be known, registered, owned, permissioned, and accountable.
Execution and orchestration
Agents need controlled access to tools, workflows, data, systems, and other agents.
Shared context, memory and state
Enterprises need clarity on what agents know, remember, retrieve, update, and pass between workflows.
Governance, approvals and escalation
Human authority must be designed into the system, not added after something goes wrong.
Observability, evaluation and assurance
Organisations need to see what agents did, why they did it, what evidence they used, what it cost, and whether it worked.
Where Praxis stands
This is where the consulting opportunity is shifting. Clients do not only need help building agents. They need help designing the control plane around agents — practical architecture, governance, delivery methods, operating roles, risk models, evaluation loops, and adoption patterns. They need someone who can connect the board-level ambition to the engineering reality.
That is the space Praxis AI Partners is built for. Our view is simple: enterprise AI value will not come from scattered experiments. It will come from governed, observable, human-accountable agent systems that can move real work through real organisations.
The winners will not be the companies with the most agents. They will be the companies with the clearest control.
Point of View
← All essays
Previous
AI dies in the operating gap →
Begin
Ready to govern the estate?
One conversation, no pitch deck. Bring the hard questions — they're the ones we design for.