Workflow vs Agent vs Manual Process: How to Choose

Should This Task Be a Workflow, an Agent, or Manual?

Choose the least autonomous operating model that can perform the task well. Stable rules and predictable paths usually favor a deterministic workflow. Variable language inside a fixed process may justify an AI-assisted workflow. A changing goal that requires tool choice and plan adaptation may justify an agent. Low-volume, high-consequence, or context-heavy work may remain manual.

The label on the product does not make the decision. The task's uncertainty, consequence, reversibility, frequency, observability, exception rate, and owner do.

flowchart TD
    A["Define one task"] --> B{"Can stable rules express the path?"}
    B -->|Yes| C{"Is the action low-risk and observable?"}
    C -->|Yes| D["Deterministic workflow"]
    C -->|No| E["Workflow with approval"]
    B -->|No| F{"Is model judgment bounded inside a fixed path?"}
    F -->|Yes| G["AI-assisted workflow"]
    F -->|No| H{"Does the goal require adaptive planning and tool choice?"}
    H -->|Yes| I["Constrained agent with checkpoints"]
    H -->|No| J["Manual process or redesign"]

Four operating models

A manual process leaves interpretation and action with a person. Software may store information or present tools, but the person decides the path.

A deterministic workflow follows declared triggers, rules, branches, and actions. The same relevant input should follow the same path.

An AI-assisted workflow keeps the orchestration fixed while a model performs a bounded step such as classification, extraction, summarization, or drafting. The model may vary its output, but it does not decide the entire plan.

An agent receives a goal, context, tools, and some discretion over how to proceed. It may choose actions, sequence work, seek more information, revise a plan, and stop when it believes the goal is complete.

These are working definitions, not universal product categories. Many vendors call a predefined flow an agent. Google itself now uses agent language around Workspace Studio, even though the editor still exposes starters and steps.

Stable work should not become agentic by default

If a task has a fixed event, a known path, and a small set of exceptions, a deterministic workflow is easier to test and operate.

Saving attachments from three approved senders into a private folder does not need a model to decide what "important" means. Sender, attachment presence, subject text, and destination can be declared.

Google's Workspace Studio condition guide illustrates the distinction. Check if evaluates fixed conditions. Decide uses Gemini for subjective or complex criteria. A fixed check is preferable when the rule can be written reliably.

Adding a model where a rule is enough creates another failure mode without adding useful flexibility.

Use an AI-assisted workflow for bounded judgment

Some steps cannot be expressed as simple matching rules. An email may need to be classified by intent. A meeting transcript may need action items extracted. A customer message may need a draft response based on an approved knowledge source.

Those jobs can fit an AI-assisted workflow when the surrounding path remains controlled. The trigger, allowed context, output shape, destination, and review rule can all be fixed even if the model's judgment varies.

The E097 recording included that structure. A flow evaluated whether an email contained a question. Another assembled a summary and delivered it to Chat. The useful question was not whether Gemini could produce text. It was whether the result appeared in the right place, used the right context, and led to an appropriate action.

If the model output directly controls a consequential step, insert approval or convert the action into a draft.

Use an agent only when adaptation is part of the value

An agent becomes useful when the goal cannot be reduced to one stable path and the system must choose among tools, gather missing information, change sequence, or recover from partial results.

Google Cloud's agentic design-pattern guide distinguishes predictable sequential work from model-orchestrated, custom-logic, and human-in-the-loop patterns. It recommends human intervention for subjective or high-stakes tasks that require oversight or final approval.

E120 provides a useful contrast with E097. Workspace Studio required Dalton to assemble declared flows for guest operations. Gemini Spark accepted a broader responsibility involving stalled conversations, research, a tracking sheet, and draft outreach. The lower setup burden transferred more judgment into the delegated goal.

That can be valuable. It also makes the finish state, allowed tools, authority, and evidence harder to specify.

Manual is an operating choice, not a failure

Keep the task manual when volume is low, the process is still changing, exceptions dominate, the context cannot be represented safely, or the cost of an error is high.

Manual work can also be the right fallback. The NIST Generative AI Profile includes fallback technologies and manual processing in its risk-management considerations.

The problem is not that a person remains in the loop. The problem is an undefined process where nobody knows when the person must intervene or what evidence they need.

Score the task, not the demo

Do not average away a severe condition. One irreversible high-consequence action can justify approval even when the rest of the task is routine.

Four worked choices

Email attachment filing fits a deterministic workflow when approved senders, file presence, and destination are known. The first release should use a narrow trigger and a private folder.

Daily newsletter summarization fits an AI-assisted workflow. The schedule and sources can be fixed while a model produces the summary. Delivery should remain self-only until source quality and output are understood.

Customer refund resolution may justify an agent only if policy, account context, tools, limits, and escalation can be encoded. A human should approve exceptions or payments above a defined boundary. For many teams, a workflow that gathers evidence and drafts the decision will be safer than an autonomous refund agent.

Executive hiring, medical triage, credit decisions, and legal commitments should not be automated from a generic matrix. Their consequences, regulation, rights, and domain standards require specialized review. A manual or decision-support process may remain appropriate.

Put approval at the action boundary

Approval should occur before money moves, access changes, content publishes, data deletes, an external commitment is made, or a regulated decision affects someone.

The reviewer needs to see the input, proposed action, reason, uncertainty, and consequence. A prompt without that context creates review theater.

For low-risk and reversible work, excessive approvals can make the system unusable. The design goal is enough human control for the consequence, not a human click attached to every step.

Make one conditional choice

Write one sentence that chooses the model and names the reason: "Use a deterministic workflow because the trigger and destination are fixed, but require approval before any message is sent externally."

Then complete How to Evaluate an AI Workflow Before You Turn It On. The guide turns the category choice into tests, permissions, monitoring, rollback, and ownership.

For a deeper agent-management model, continue with E120's [[How to Manage an AI Coworker|How to Manage an AI Coworker]]. The operating principle is the same across both episodes: delegate the path only after the outcome, authority, and proof are clear.

This decision guide was developed from current Google Cloud and Workspace documentation, NIST materials, and the E097 and E120 episode evidence. AI assistance was used for research organization, drafting, and validation. The matrix does not replace domain-specific legal, compliance, safety, or professional review.