Full autonomy is overrated
The obsession with fully autonomous AI is distracting enterprises from systems that already know how to get work done. The prevailing image of an AI agent is a digital worker that observes a situation, determines what to do, and executes the entire process independently. This model can be useful, but it is not the only model for agentic AI, nor is it the best fit for many enterprise processes.
An agent does not need to own every task associated with a business outcome. It can rely on a deterministic workflow that owns the overall process. Within that process, the agent interprets an unstructured document, assesses an ambiguous situation or chooses between a defined set of paths.
Massimo Pezzini makes this distinction in “To Succeed With Agentic AI, You Need an Orchestration Strategy”. Agents can act as non-deterministic orchestrators in uncertain situations, but they can also participate as individual tasks inside structured business processes that remain under deterministic control.
That second pattern deserves more attention than it currently has. The most effective enterprise AI system may not be one in which the agent does everything, but one in which the architecture knows exactly when to involve the agent and exactly when not to.
Proactivity does not have to be AI
Some argue that AI cannot be genuinely proactive, but that frames the problem too narrowly. Proactivity does not need to originate entirely from the AI component. It can also emerge from the orchestration of the complete system.
A regular Workato recipe can monitor for a business event. When that event occurs, the recipe can gather the relevant information and ask an AI agent to determine the next course of action. Once the agent has made that determination, the recipe can execute the appropriate steps across the relevant enterprise systems. The trigger is deterministic; the decision may involve AI reasoning, and the execution can revert to deterministic automation.
That is not a compromise. In many cases, it is the right division of responsibility.
A Genie workflow can follow the same pattern. It can use deterministic tools and skills to gather and preprocess information, eliminate routine cases, identify what matters, and involve a person only when judgment, accountability, policy, or genuine exception handling requires attention. The Genie can then determine what context and recommendation to present. The user experiences a proactive system even though not every individual step is powered by generative AI.
Proactivity is a property of the system. It is not a requirement that every component pretend to be intelligent.
Do not make the agent assemble the enterprise from scratch
An agent with stale or incomplete context can sound intelligent while being confidently wrong. There is a tendency to discuss agents as though the information they need will simply be available when they need it. It will not.
Before an agent can reason about a customer, supplier, invoice, transaction, or employee, the relevant data must be retrieved from the systems of record. It will need to be synchronised, reconciled, transformed, filtered, and assembled into a form the agent can use.
That could mean synchronising relevant operational information into Salesforce to provide a trusted customer context. It could mean retrieving the current status from an ERP, matching it with CRM data, checking a policy repository, and determining whether an approval is already in progress. It could also mean identifying which pieces of information are relevant to the current task so the agent does not receive an excessive amount of unnecessary context.
Much of the work that makes an agent appear intelligent happens before the model is invoked. Massimo’s orchestration framework makes this explicit: enterprise AI requires current context, event delivery, access to reusable enterprise capabilities, and mechanisms for inserting agents into deterministic processes. The quality of an agent’s reasoning cannot compensate for poor context architecture.
Stop paying models to rediscover your process
The boundary between agent reasoning and deterministic execution affects more than reliability. It also affects cost. A cheaper model does not make unnecessary reasoning economical. If a deterministic recipe can execute a routine step more reliably and at lower cost, there is little value in repeatedly asking a model to reason through it.
“Cheaper Tokens, Bigger Bills: Why AI Agent Costs Keep Rising” argues that agent economics are shaped by not just model pricing, but also execution architecture. When every system operation by an agent is exposed as a separate tool, the model must repeatedly waste tokens reinventing the wheel at every step.
The same architectural problem appeared in Part 1, albeit in another form. Exposing granular operations to every agent can recreate point-to-point integration. Worse, it can force the model to repeatedly reconstruct coordination logic that the enterprise has already encoded. The result is not only fragmented architecture, but also unnecessary reasoning, additional latency, more context, and higher cost.
That design can force the model to reconstruct the same established business process every time it runs. It may repeatedly return the same information, carry more context than the task requires, spend reasoning capacity coordinating routine application calls, and create additional opportunities for inconsistency or failure.
The more important architectural question is not simply how to reduce the cost of each model call.
Does the process need to return to the model for that step at all?
Models should be used where judgment materially improves the outcome. Routine coordination should remain outside the model loop.
Use AI at the points of uncertainty
The model should handle the part of the process that genuinely requires interpretation or judgment. It should not have to reconstruct the parts that the enterprise already knows how to execute.
That boundary will look different depending on the process. AI is best used where a business process genuinely requires judgment under uncertainty, such as assessing an unusual situation or classifying an exception. Once that uncertainty has been resolved, deterministic orchestration can take over, applying established, known execution workflows to achieve the outcome. If uncertainty appears again later, the model can be invoked again.
The architecture does not need to decide that an entire workflow is either “AI” or “non-AI.” It can deliberately switch between probabilistic reasoning and deterministic execution depending on the nature of each step.
This is the same principle described in “Cheaper Tokens, Bigger Bills: Why AI Agent Costs Keep Rising”. Models should interpret intent, resolve ambiguity, and make judgment calls where those capabilities improve the outcome. Established sequencing, business rules, approvals, retries, and recovery should not be repeatedly rediscovered through model reasoning.
The same principle applies when an agent owns the broader orchestration. In less structured work, a Genie may dynamically determine which specialist capabilities to invoke. Uncertainty can remain with the agent where it adds value, while established business operations execute through governed capabilities rather than being reconstructed from raw application calls.
The agent may interpret the goal, but it does not need to reason through every technical step required to complete it. That makes the overall process more predictable, easier to govern, easier to reuse, and easier to measure.
The best AI workflow is often only partly AI
Enterprise AI should not be evaluated by how much of a process the model controls. It should be evaluated by whether the process reaches the right outcome with the appropriate combination of judgment, reliability, cost, speed, and accountability. Using less AI can make the outcome more intelligent.
A Genie can orchestrate a broader outcome by delegating bounded work to another Genie that operates as a reusable specialist capability. It can also invoke a recipe that executes an established process. A workflow can detect an event, assemble context, ask an agent for a decision, and continue execution without requiring the agent to coordinate every application call. These patterns can be combined because they do not belong to separate worlds. AI reasoning, deterministic automation, data orchestration, application integration, specialist Genies, and human approval are all participants in the same enterprise process.
Workato has helped my team avoid asking whether we need an AI workflow or a regular one. That distinction becomes less useful when the same process can use AI where it adds value, and conventional orchestration everywhere else.
As I argued at the start of this series, the agent is not the architecture. The agent does not have to be the whole workflow. It can be one participant in a larger system designed to produce a business outcome.
But once that participant can take action, another set of questions becomes unavoidable: Who is it acting as? What is it allowed to do? Can the enterprise prove what happened?
Those questions are the difference between an impressive AI experience and an accountable enterprise system. They are the focus of Part 3.