A few dozen automations are easy to watch. Govern a thousand.
Caddi's AI agent governance gives you one firm‑wide view of every Caddi agent and the work it hands to Claude and Copilot, and replays any Caddi run, decision by decision.
Replayable runs·SOC 2 Type II·From the Team planRunning, being built, discovered. One list.
The question is not whether one loop worked. It is what is running in my firm right now, in which tool, who owns it, what is about to join it, and what is still done by hand.
- RunningEvery agent Caddi runs or hands to Claude and Copilot
Name, owner, where it runs, and what it did this week. The skills and agents Caddi hands to Claude and Copilot report each decision and API call back, next to the loops Caddi runs. An agent with no owner is the thing this view exists to prevent.
- Being builtEvery loop in Automate, and who is building it
What is coming, before it arrives. The pipeline, without asking anyone for a status update.
- DiscoveredEvery finding, and what is still manual
What Discover found across the firm, which team it belongs to and which tool it should run in, so the work still done by hand is a list you can assign.
Every loop reports its week in plain English.
Open any loop and the first thing on the screen is what it did in the last seven days: what it picked up, what it changed, where it put it. The run log is still there. A program lead does not have to read it.
- Written from the record
The sentences are a template. The numbers in them are the run record's own counts for the last seven days, so the paragraph cannot say more than the runs did.
- Work counted by system
Documents read, updated and created. Fields extracted. Matters listed. One total a leader can quote, and the rows behind it.
- The people it worked for
Who the loop saved time for this week, so the count has a team behind it.
Caddi picked up 64 files from the daily SharePoint mail folder. AI extraction identified names and matter numbers in 64 documents, and 64 files were renamed to the standardized format. 53 documents were uploaded to their matched Clio matters.
Every run on the record. No black box.
Deterministic code does the work, so the same input takes the same path every time and the path can be written down.
- Every run, replayable
- Every decision, with its permission
- Every stop, immediate
“Thankfully, Caddi has a mature audit logs program, so I'm not concerned.”
One report across every tool.
“The fact that you have detailed, transparent information, and can provide it reasonably easily in a data extract. That is a really big selling point for me.”
- Run records with counts by system
- Action-level audit logging
- The record is yours to pull
What a program lead asks
Can a loop decide something on its own?
No. Deterministic code runs the work, AI is used inside specific steps that need judgment, and changing a process takes a human approval. The plan is the boundary, and the plan is written down.
Can we prove what it did, and how fast can we stop it?
Every run replays decision by decision, and every decision names the system and the permission it used. That record is what an auditor reads and what the security review asks for first. To stop a loop, pause it or pull its connection to a system. Either takes effect mid-queue.
Does Govern monitor Claude and Copilot?
Yes, for the work Caddi hands them. When a Discover finding is exported, Claude gets it as a skill and Copilot as an agent, and that skill reports each decision and API call back to Govern, next to the runs in Caddi. It does not see a person's ad hoc chats or automations built in Claude or Copilot without Caddi.
Which work should run in Caddi, and which in Claude or Copilot?
Caddi is built for high-volume, high-accuracy work: deterministic code runs it, so the same input takes the same path every run. Claude and Copilot are great for the judgment work in a person's day. Discover recommends a tool for each finding, and Govern monitors it either way.
How do teams, access and plans work?
Loops and findings roll up by team, and the firm-wide view is on the Team plan. Action-level audit logging, SAML SSO, centralized MFA, SCIM provisioning and centralized integration governance are on Enterprise. Each connected system carries one scope, under the access the person already had. See pricing.
Is our data used to train models?
No. The contents of your connected systems are never used to train models. OpenAI and Anthropic calls run under zero-retention terms, and accounts under a Data Processing Addendum exclude their content from model development entirely.