An agent is a set of written standards that AI check measures your policies against. Blissbook gives you four to start with, and you can rewrite any of them, add your own, or build them out of the guidelines your team already has.

This is the part of AI check worth spending time on. The built-in agents are deliberately general, and an agent that knows your terminology, jurisdictions, and your house style produces recommendations your team will actually accept.

AI check is currently in beta.

You'll see both names. Agents is what they're called when you pick them for a check, and Policy Skills is the set of rules that an agent uses to analyze your policy.

There are two ways in, and they land on the same page.

From your Policy Library, click Settings in the top right, open the Policy AI tab, and click Configure on the AI Policy Check row.

From the main navigation, click Blissbook AI and click Configure on the same row.

Which one you use depends on your role, and the two roles don't overlap:

So a Team Owner who isn't an admin can write agents but can't turn the feature on, and an admin who isn't a Team Owner can turn it on but can't open the agents. If you need to do both, ask for both.

If you don't see any of this, AI check probably isn't enabled for your organization yet. Contact us.

You can also get here from a check itself: click Edit AI Agents in the window that opens when you run an AI check. It opens the same page in a new tab, so you won't lose your place.

Every organization gets these the first time AI check is enabled, already released and ready to run:

Start with Legal Compliance and Brand Voice.

Legal Compliance has an Applicable Jurisdictions section, and it arrives set to United States (federal). Until you change it, every policy is reviewed against US federal employment law, so if you employ people anywhere else, this is the first edit to make. You could list the countries, states, counties, and cities you operate in, but...

Every policy in your library has a Jurisdiction property (Country, State, County, and City) and when a check runs, that policy's jurisdiction is handed to your agents along with the rest of its properties. If you set the Jurisdiction property on your policies, you can simply tell the Legal Compliance agent to review each policy against its own jurisdiction property, and fall back to your list when a policy doesn't have one set.

Brand Voice, meanwhile, can't match a voice it has never read. Pasting in a few paragraphs that demonstrate your brand voice as examples will help the Brand Voice you're happy with and tell it what you're going for.

Open the agent, make your edits, and click Release in the top right.

Write it the way you'd brief a new hire. Name the terms you use and the ones you avoid. Say which sections every policy needs, and what reading level you're aiming for. Specific instructions help you get more specific recommendations.

Releasing an agent doesn't re-check your policies. Run the check again on the ones you want to check against the new version.

Click Add Policy Skills and choose Create a blank Policy Skill. Give it a name, type or paste your guidance, and click Release.

One agent per subject works better than one agent for everything. A short agent covering just your policy framework gives you cleaner recommendations, and when a recommendation misses, it's easier to know exactly which agent to fix.

If your standards already live in a style guide or a compliance checklist, upload it and Blissbook drafts the agents for you.

Click Add Policy Skills, choose Import Policy Guidelines, then drop in one Word (.docx), PDF, or text (.txt) file up to 25 MB and click Analyze Guidelines.

Analyzing takes a few minutes. You can close the window and it keeps going.

When it finishes, Review Your Skills shows what it found: how many rules it pulled out of your document, which existing agents it updated, which new agents it created, and a Not Included section listing content it couldn't turn into rules.

An import isn't finished when you close that report. Brand new agents arrive released and ready to run, but changes to an agent you already had arrive as suggested edits on that agent, the same cards you work through after an AI check. Open each updated agent, apply the edits you agree with, and click Release. The count beside each updated agent in the report tells you how many edits are waiting.

Note: An import replaces open recommendations on any agent it changes, and Blissbook tells you before it does.

When your colleagues flag a recommendation as wrong while working through a check, those flags collect at the bottom of your policy library, grouped by the agent that made them. A cluster of flags under one agent is the clearest signal you'll get that the agent needs a rewrite, and you can hand that job to Blissbook rather than doing it yourself.

In that review, click Update Skill on an agent and Blissbook reads its flags and drafts edits to the agent. They land here, on the agent, exactly like the edits from an import: apply the ones you agree with, then Release. A Skills left to review count in the library footer keeps track of agents with edits still waiting.

Select the agent and use the three-dot menu to Archive it. Archived agents stop appearing when you run a check and stop influencing results.

Archiving is reversible. Unarchive puts the agent back with its content and version history intact. Delete Permanently appears in the same menu once an agent is archived, and that one can't be undone, so archive unless you're certain.

The same menu has Clone Policy Skill, which is the easy way to try a rewrite: clone the agent, edit the copy, release it, and archive whichever version you like less.

Your agents are the part of AI check only you can write. Spend twenty minutes on Legal Compliance and Brand Voice now and every check you run afterwards measures your policies against your standards instead of generic ones.

Note: The Legal Compliance check is based on underlying models and training data from Anthropic, with specific instructions and training data from Blissbook (and each customers' agent skill) layered on top to increase recommendation quality. We're confident it provides solid advice, but just like any AI agent (or human for that matter), it can make mistakes. It will improve over time; you can help by flagging any recommendations you think are off.