Why AI Review Beats AI Approval

DAVE's Workflow Editor gives you ten node types. Two of them are specifically about review: Agent Review and Human Review. They look similar in the Node Palette and they both sit after an Agent Interaction node in most workflow designs. But they do fundamentally different things, and confusing them produces workflows with accountability gaps that are invisible until something goes wrong.

Agent Review: automated quality check

The Agent Review node (pink, Bot icon) invokes a second AI agent to evaluate content against defined criteria and return a pass/fail result. It runs automatically. No human is involved in the evaluation. The agent processes the workflow context, applies its configured quality criteria, and the workflow branches on the result. When it works well, it is fast, consistent, and scalable: the same criteria applied to every item, every time, with no reviewer fatigue.

What it is not: an approval. The agent has no authority. It has no accountability. It cannot be held responsible for a wrong decision. It can only report what it found against the criteria it was given. If those criteria are incomplete, or if the content falls outside the range the agent was designed to evaluate, the result may be confidently wrong.

Human Review: a decision with a named owner

The Human Review node (amber, Eye icon) pauses the workflow and creates a task in the Task Inbox of the assigned reviewer. The reviewer reads the payload, which contains the content and any upstream context the workflow has accumulated, and makes a decision: Approve, Reject, or Request Changes. The workflow branches on that decision. The task is recorded with the reviewer's identity, their decision, and any comment they added.

That record is what makes Human Review different from Agent Review in governance terms. A human made a named, logged decision. If that decision turns out to be wrong, there is accountability. If it turns out to be right, there is evidence. Neither is possible when an AI agent approves its own output.

Agent Review is not a second-rate substitute for Human Review. It is the right tool for a specific job: high-volume, criteria-driven quality checks where the criteria are well-defined, the content is within the agent's competence, and the cost of an occasional error is low enough to accept in exchange for speed and scale.

Good uses for Agent Review:

• Format and completeness checks. Does the output contain all required sections? Is it within the required word count? Does it follow the specified structure? These are mechanical criteria that an agent can evaluate reliably and consistently.
• Style and tone consistency. Does the output match the brand voice guidelines? Does it avoid prohibited terms? Is the reading level appropriate? With a well-written review prompt and a locked agent configuration, these checks are more consistent than human reviewers who vary in their interpretation of style guides.
• Pre-filtering before human review. Use an Agent Review node to catch obvious failures before the content reaches a human reviewer. The human sees only items that passed the automated check, which means their attention is focused on the harder judgment calls rather than mechanical errors.
• High-frequency intermediate checks. In a multi-step content pipeline, Agent Review nodes can check quality at each stage without creating a bottleneck. A Human Review node at every stage would make the workflow too slow for practical use.

The common thread: Agent Review is appropriate when you can write explicit, testable criteria and when the consequences of a wrong result are recoverable. When either of those conditions does not hold, you need a Human Review node.

One important nuance from the platform: Agent Review tasks do appear in the Task Inbox with type agent_review. This means a workflow can be designed so that an Agent Review result is surfaced to a human for override before the workflow continues. This is a useful hybrid pattern: the agent does the evaluation, but a human can intervene if the agent's result is flagged or disputed. The design choice is yours to make in the Workflow Editor.

There are workflow contexts where replacing Human Review with Agent Review is not a design trade-off: it is a governance failure. The Human Review node is not optional when any of the following conditions apply.

The output will be acted on externally

If the workflow's output will be published, sent to a customer, submitted to a regulator, or used to make a consequential decision, a human must approve it before it leaves the system. An AI agent cannot be accountable for that approval. The Human Review node creates the named, logged decision that makes accountability possible.

The criteria cannot be fully specified in advance

Agent Review works when the quality criteria are explicit and complete. Many real-world quality judgments are not: they involve context, nuance, and domain knowledge that cannot be fully captured in a prompt. When the criteria for "good enough" include judgment calls that a domain expert would make differently depending on context, that judgment belongs to a human reviewer, not an agent.

The content involves sensitive categories

Legal, medical, financial, and compliance content carries risk that scales with the consequence of error. The same is true for content involving personal data, content that could affect safety, and content that represents the organization's position on a contested matter. These categories warrant human review regardless of how well the automated check performs on average.

Regulatory or contractual requirements specify human sign-off

Some frameworks require documented human approval for specific categories of decision. In those cases, the Human Review node is not a design choice: it is a compliance requirement. The task completion record (reviewer identity, decision, timestamp, comment) is the evidence that the requirement was met. DAVE's audit trail supports SOC 2 CC6.1 and GDPR Article 30 when those standards are active.

The workflow is new or the agent is untested

A new workflow or a newly configured agent has an unknown error rate. Until you have enough production data to characterize the agent's performance on your specific content, a Human Review node after every Agent Review node is a sensible default. As confidence in the agent's performance grows, you can adjust the routing to send only low-confidence or flagged results to human review, and let high-confidence passes through automatically.

The most robust pattern for AI workflow quality control is not a choice between Agent Review and Human Review. It is a sequence: Agent Review first, Human Review for the cases that need it. Here is how to design it in DAVE's Workflow Editor.

The basic pattern

Place an Agent Interaction node to generate the content. Connect it to an Agent Review node configured with your quality criteria. Connect the Agent Review node's output to a Routing node that reads the review result. Route passing results to the next workflow step. Route failing or low-confidence results to a Human Review node. The Human Review node's Approve, Reject, and Request Changes decisions each route to their own downstream paths.

This pattern means humans spend their review time on the items that actually need judgment, not on mechanical quality checks that an agent handles reliably. It also means every item that reaches a human reviewer has already passed an automated baseline check, so the reviewer's attention is focused on the harder calls.

Configuring the Human Review node

In the Workflow Editor, select the Human Review node and open the Properties Panel on the right. Set the Label to something specific (for example: "Editorial Director Approval" rather than the default "Human Review"). Under User Assignment, specify the assigned user IDs or roles for this review step. Under Notifications, choose whether this step follows the workspace notification rule, suppresses notifications, or notifies a custom audience. For a high-stakes final approval, a custom audience that includes the instance owner and the assigned reviewer is a sensible default.

What the reviewer sees

When the workflow reaches the Human Review node, a task is created in the Task Inbox with type human_review. The reviewer navigates to Tasks in the sidebar to find it. The task detail page shows the payload (the content and context passed from the workflow), the current status, the due date if set, and the decision panel with Approve, Reject, and Request Changes buttons. The reviewer adds an optional comment and clicks a decision. The workflow resumes on the corresponding outgoing edge from the Human Review node.

Task access requires the Admin or Use role. DAVE ships with six default tenant roles (Admin, Create, Curate, Use, Reporting, and Financial), all editable by your administrator. Reviewers who need Task Inbox access but should not have administrative access should be assigned the Use role.

Saving and versioning

Every time you click Save Version in the Workflow Editor, a new immutable version of the workflow graph is created. Previous versions are preserved. If you change the routing between Agent Review and Human Review nodes, the old version remains available. Only Admin and Curate roles can open the Workflow Editor and save new versions.

The argument for using Agent Review as an approval mechanism usually sounds like efficiency: faster throughput, no reviewer bottleneck, consistent application of criteria. These are real benefits. The argument fails when you ask a harder question: when the agent approves something it should not have, who is responsible?

The answer, with a pure Agent Review approval design, is: no one in particular. The agent has no identity in the accountability sense. It has a configuration, a prompt, a model, and a version. You can audit all of those. But none of them constitute responsibility for a decision in the way that a named human reviewer's logged Approve action does.

This matters in three practical situations. First, when a customer or regulator asks who approved a piece of content, "the agent" is not an answer that satisfies the question. Second, when something goes wrong and you need to understand whether the failure was in the criteria, the agent's interpretation of them, or the content itself, the absence of a human decision in the record makes the investigation harder. Third, when the organization needs to demonstrate that a human was in the loop for a specific category of decision, Agent Review alone cannot provide that evidence.

None of this means Agent Review is wrong. It means Agent Review is a quality check, and Human Review is an approval. Designing workflows that treat them as interchangeable is the mistake. Designing workflows that use each for what it is actually good for produces systems that are both efficient and accountable: fast where speed is appropriate, careful where accountability is required.