From Workflow Automation to Agentic AI

Organizations with mature workflow automation often treat agentic AI as the next release of the same capability. It is not an upgrade path. It changes the unit of specification, and everything downstream of that — testing, monitoring, incident response, cost modelling — changes with it.

The unit of specification changes

In workflow automation you specify steps. The artefact under review is a sequence: fetch this, transform that, write there, notify them. Correctness means the sequence executed as written.

In an agentic system you specify an objective, a set of tools the system may use, and the boundaries on their use. The artefact under review is a goal and a permission set. Correctness means the outcome was appropriate — a judgement that cannot be made by inspecting a log line.

This is why teams that port their existing acceptance tests to an agentic system find the tests no longer answer the question. "Did step four run?" is replaced by "was this a reasonable way to handle this case?"

What has to be built that did not exist before

An evaluation set. A body of representative cases with agreed correct outcomes, used to measure the system before release and to detect drift afterwards. Without it there is no basis for saying a change improved anything.

Action boundaries. An explicit list of what the system may do, with irreversible actions separated from reversible ones. Sending an email, issuing a refund and deleting a record each deserve different treatment.

Decision logging. Enough recorded context to reconstruct why the system did what it did: input, retrieved context, tools called, output, and which version of the model and instructions produced it. Retrofitting this after an incident is not possible.

A stop mechanism. A defined way to pause the system that does not require a deployment, and a documented owner who is permitted to use it.

Cost behaves differently

Scripted automation has near-zero marginal cost per execution. Agentic systems are metered: every run consumes inference, often several times over as the system reasons, retries and calls tools.

The practical consequence is that cost per case becomes a design variable rather than an afterthought. A system that handles an exception by looping through additional attempts can be correct and uneconomic simultaneously. Volume forecasts belong in the design review.

A sensible migration order

Start where existing automation already halts. Every workflow has an exception queue — the cases the rules could not handle, currently routed to people. That queue is the natural first target: the volume is known, the cases are genuinely difficult to specify, and the comparison baseline already exists because someone is handling them today.

Keep the deterministic path for everything it already handles. Replacing working rules with inference adds cost and variance for no operational gain, and it is the most common way these programmes lose credibility early.

Then widen the boundary gradually, with review on consequential actions, measuring against the evaluation set at each step. The related discussion of human oversight covers how to keep that review meaningful as volume grows.

What does not change

The process still has to be worth running. Agentic systems inherit the quality of the process they execute, and a goal-directed system pursuing a badly specified objective in a contested process will pursue it efficiently. Clarity about what the work is for remains the prerequisite it always was.

References

1. National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1.
2. Amershi, S., Weld, D., Vorvoreanu, M., et al. (2019). Guidelines for human-AI interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems.
3. ISO/IEC 42001:2023. Information technology — Artificial intelligence — Management system. International Organization for Standardization.