‹ Legal & Compliance recipesLegal & Compliance · Agent Blueprint

Policy gap review

Where practice has drifted from written policy. What people actually do vs. what the handbook says, cited to the evidence.

Best forCompliance reviews, policy refresh
PrimitivesPoliciesPracticeDiff

Why this is memory-only. No one announces that they are deviating from policy. Practice drifts quietly - an approval skipped, a process shortened, a step done differently - and it never appears in any document anyone would search for. The only way to see the drift is to compare what people actually say and do, captured continuously, against the written policy.

Side by side: token usage, with and without Paperbase

Memory only

Without Paperbase

Not recoverable from policy documents alone

  1. Read the policy documents.
  2. Ask managers whether practice matches policy.
  3. Skim process logs for exceptions.
  4. Audit by sampling whatever happens to be recorded.
  5. Approve a policy that everyone quietly stopped following.

With Paperbase

Persistent memory layer

  1. Pull the written Policies.
  2. Ask Paperbase for how the corresponding process is actually described and performed across interactions.
  3. Diff policy against practice, finding each discrepancy.
  4. Cite the interaction where the deviation happened.

Agent prompt

You are checking {{policy_area}} policy against actual practice.

Using Paperbase:

1. Pull the written policies for {{policy_area}}.
2. Pull how the related process is actually described and carried out across recent interactions.
3. Diff policy against practice.
4. For each gap, cite the interaction that shows the deviation.

Output:

- Every gap between written policy and practice.
- For each: the policy, the practice, the quote showing it.
- A prioritized remediation list, by risk.

Placeholders in {double_braces} are inputs the agent will ask for at runtime. Give the prompt to any agent connected to Paperbase, and the rest grounds in your own memory graph.

Want a deeper recipe for your team?

Send us the prompt, the source systems, and the workflow. We will draft the recipe and sign you into the sandbox to run it.