The language of enterprise AI has moved quickly from copilots to agents. The architectural difference is important. A workflow follows a defined sequence; an agent decides which steps and tools to use as the task unfolds. The second can handle more variation, but it also introduces more ways to fail.
The 2026 Stanford AI Index captures the gap between interest and deployment: organisational use of AI is widespread, while agent use remains in the single digits across nearly all business functions. That caution is rational. Autonomy should be earned through evidence, not assumed from a compelling demonstration.
Use the least autonomy that solves the problem
Anthropic's engineering guide to building effective agents distinguishes workflows, in which models and tools follow predefined paths, from agents, in which the model directs its own process. It recommends starting with the simplest solution and adding complexity only when it improves outcomes.
That principle is particularly useful in enterprise settings. If a task can be expressed as retrieve, classify, draft and route, a workflow offers predictable control points. The model still adds judgement inside each step, but the organisation retains a legible process.
Agents become valuable when the path cannot be known in advance: investigating a changing incident, reconciling evidence across systems or iterating until a result meets a defined standard. Even then, the objective and boundaries should remain explicit.
Build a progression, not a leap
A responsible path to autonomy has four stages.
First, provide assistance without action. The system gathers evidence, proposes an answer and leaves execution to a person. This reveals the quality of retrieval and reasoning without granting operational authority.
Second, automate a deterministic workflow. The system can write to tools, but only through approved steps with clear validation.
Third, introduce bounded choice. The agent may select among a limited set of tools or recovery paths, subject to cost, time and permission limits.
Finally, allow broader planning where the evidence shows it is justified. High-impact actions remain separately controlled, and the agent produces a receipt of what it did.
Each stage should have its own acceptance criteria. Promotion depends on observed reliability, not elapsed time.
Design the boundary before the prompt
Prompt quality matters, but it cannot carry the full security and governance burden. An enterprise agent needs a separate identity, narrowly scoped credentials, explicit tool permissions and a transaction boundary that distinguishes reversible preparation from consequential action.
The system should know which actions can run automatically, which require contextual approval and which are prohibited. It should also cap resource use. A looping research agent can waste time and money without ever performing a dramatic action.
Observability is equally important. Record the sources consulted, tool calls made, policy decisions applied and final state of the task. A natural-language summary is helpful; an immutable execution record is better.
Make stopping a first-class capability
An agent needs conditions for completion, escalation and refusal. It should stop when evidence is missing, sources conflict, permissions are insufficient or expected state has changed.
This is not a weakness. A system that recognises uncertainty is more useful than one that always produces a fluent answer. Human intervention should resolve a meaningful decision, not compensate for missing basic controls.
Autonomy is an operational property
An impressive model does not make an autonomous system safe. Autonomy emerges from the combination of model behaviour, tools, identity, data, memory, policy and feedback.
The practical enterprise question is therefore not, “Can this model act as an agent?” It is, “Within which bounded workflow has this system demonstrated that it can act reliably, visibly and reversibly?”
Start with that boundary. Prove the workflow. Learn from real outcomes. Then widen autonomy only where the organisation can explain why the next step is justified.