Every organization runs on approvals, and each one of them takes time. Stakeholders define the approval logic in a policy, a developer translates it into workflow logic, and still every approval request lands in a human's queue. It's time to change!

 

@Riya Kapoor, Senior Product Manager on the Workflow Automation team, walked through how Intelligent Approvals closes that gap. The pitch is simple: stop making developers translate policy into logic, and stop asking humans to review every decisions the policy already made. Let the document itself drive the approval.

 

https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FQJ63KN2TF6w%3Ffeature%3Doembed&display_name=YouTube&url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DQJ63KN2TF6w&image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FQJ63KN2TF6w%2Fhqdefault.jpg&type=text%2Fhtml&schema=youtube

 

What's new with Intelligent Approvals

 

Intelligent Approvals reads a policy document (a PDF, or a knowledge base article you already have) and turns it into approval logic on its own. From there it evaluates incoming requests at runtime and lands on one of three outcomes:

 

• Auto-approve: the request matches conditions the policy already answered.
• Auto-reject: the policy is just as clear about what shouldn't go through.
• Escalate: the request is genuinely ambiguous, so it falls back to your existing approval workflow.

It sits on top of the approval workflows you already have as a filter, catching the clear-cut cases, so only the requests that genuinely need a review lands in the human's queue.

 

Seeing it in action

 

The flow starts from the Intelligent Approvals Home application: upload a policy PDF, or search for a knowledge base article and preview it before you commit. You can also kick it off with a Now Assist prompt like "approve my leave policy" instead of clicking through the Create New button.

 

Once you upload, the AI extracts the governance rules and proposes a trigger, meaning a table and a condition for when the approval should fire. There is a human-in-the-loop edit option too, if the AI's trigger guess is wrong, you can fix the table or condition yourself before moving on.

 

That takes you to the policy's home page, which is really the whole product in one view:

 

• Overview: a plain-language summary of what the AI understood from your document.
• Scope and coverage: the trigger you just configured, with a pie chart showing what percentage of requests it expects to decide on its own versus send to a human.
• Interpreted policy areas: which parts of the policy map to an automatic decision, and which map to a manual one.
• Source document: the file you uploaded, kept as the system's source of truth.
• Test results: AI-generated test cases run against your policy, each with a decision and the reasoning behind it.
• Suggested improvements: flags for ambiguous wording (For example, a policy that said a request "maybe" approved) with suggestions for tightening the language so the AI stops defaulting to manual review.

 

Once you're happy with the test results, you can publish. Building and testing a policy costs zero Now Assist credits; each automated decision after publish costs 25.

 

In the demo, Riya's change request policy said a normal, low-risk change should auto-approve and anything emergency and high-risk should always go to a person, no AI decision at all. Submitting a normal, low-risk change auto-approved it in seconds, with a comment showing the reasoning and the related approval records marked as no longer required. Submitting an emergency, high-risk change did the opposite: the approval records were created as requested, and a work note explained that Intelligent Approvals couldn't decide and why.

 

Working with existing workflows

 

A few questions came up during the session worth calling out directly:

 

Do the suggested improvements actually make the policy easier for the AI to read? - Yes. Ambiguous language, like a policy that says a request "maybe" qualifies, makes the AI more conservative and pushes more requests to manual review by default. Tightening that language is what moves the pie chart above 50%.

 

Can someone override a decision after Intelligent Approvals makes it?  - No. Once a decision is made, or the record reaches a completed state, it's final. Intelligent Approvals takes precedence over your existing ask-for-approval step, and it only falls back to that step when it genuinely can't decide.

 

How does Intelligent Approval coexist with the approval logic already built into flows? - Today's ask-for-approval step is hardcoded into the workflow and always routes to a human, with no intelligence behind it. Intelligent Approvals separates the approval logic from the workflow logic entirely, reads the policy in natural language, and only hands a request to a person when the policy is genuinely ambiguous.

 

Availability and licensing

 

Intelligent Approvals is running on Australia Patch 4 or Zurich Patch 11. Full functionality, meaning the AI making live decisions on published policies, needs Now Assist for Platform (Pro Plus), AI Platform Advanced, or Prime. PDF and knowledge base articles work as document sources today; IRM policies and external cloud documents are on the roadmap.

 

What's next

 

Riya shared two things her team is actively building.

The first uses historical approval data: right now the AI only reads your policy document, not the thousands of past requests already sitting in approved or rejected states, so a request that lands in "can't decide" has no memory of similar past decisions to lean on.

The second is a recommendation mode for the cases where teams aren't ready to hand over the decision entirely, especially emergency or high-risk requests. Instead of auto-approving or auto-rejecting, the AI would surface its recommendation and let a person make the final call.

 

Which approval queue would you hand over to Intelligent Approval?