For AI agents: a documentation index is available at https://docs.coverbase.com/llms.txt. This page is also available in markdown by appending .md to the URL.
This page is part of the User Guides collection and the Coverbase AI agent (Beta) set. It sits beside The Coverbase AI agent (Beta) and Controls, approvals and governance, which hold the detail behind each answer.
Coverbase AI is in Beta. Your Coverbase team turns it on for your organization.
Scope and purpose
What is the Coverbase AI agent?
It is an agent inside Coverbase that carries out third-party risk work for you. You describe a job, it makes a plan, and it carries out each step with the product’s own tools as you, stopping for your approval before changes that matter. See The Coverbase AI agent (Beta).How is it different from Ask Coverbase AI?
The Ask Coverbase AI chat answers one question about one record and can propose a few changes you apply yourself. Coverbase AI carries out a multi-step job across the product and makes changes as you, within checkpoints and policies. The chat keeps working beside it. See How the agent differs from earlier AI features.How is it different from the AI we already use in assessments and intake?
The AI inside a feature produces one output, such as a control evaluation or a drafted answer, for a person to review. Coverbase AI moves between features: it can open the intake, launch the assessment, read its results and raise findings in one run. It calls the same features, so an assessment it launches is evaluated the same way as one you launch.What does Beta mean here?
Your Coverbase team turns Coverbase AI on for your organization, and it is off until they do. Some modules are not available to it yet: RFPs, Inspect, risk scenarios, sending a contract for signature and allowlisting a screening match. A run that touches one of them says so and does the rest of the job.What can it do?
It can onboard vendors, research inherent risk, launch and review assessments, triage Zero Touch results, build control sets, questionnaires, workflows and dashboards, work findings and commitments, work contracts, obligations and the risk register, triage Radar, and change configuration as an admin would. What the agent can do lists more than 150 example instructions.Changes and approvals
Can it change records without approval?
Yes, for changes you can undo in the product that reach nobody outside your team, such as tagging a vendor, drafting follow-ups or changing a finding’s owner. It stops for approval at checkpoints, such as launching an assessment or turning a detector on, and always asks before archives and actions that cannot be undone. An admin can make any change ask first by setting it to Always asks, or forbid it by setting it to Off. See What Coverbase AI changes without asking.What is a checkpoint?
A checkpoint is a change that stops the run until the person who started it approves, declines, or says what to do instead. Checkpoints cover launches, turning monitors and automations on or off, changes to an assessment’s result, findings, obligations, commitments, control set and questionnaire changes, admin configuration and Radar triage. See Checkpoints.What does “Don’t stop to ask” do, and what are its limits?
It applies every checkpoint that can be undone without stopping, for one run. It never skips an archive, an action that cannot be undone, or a change your admins set to Always asks, and it never makes a change set to Off. Each change it applies is recorded as Applied under org policy. See Don’t stop to ask.What does “Approve all like this” do?
It approves the current checkpoint and every later call of the same change in that run. It is offered only for a change that can be undone and that your admins have not set to Always asks. It lasts for that run only.Can our admins restrict what it does?
Yes. Under Configuration > Coverbase AI Memory, What Coverbase AI May Do on Its Own lists each change Coverbase AI can make, with Applies on its own, Always asks or Off. The policy applies to every run in the organization. Changing it needs a role that can change organization settings. See Organization policies for each change.Can it delete things?
No. Coverbase AI never deletes a record. It archives, which removes the record from every list and report, and an admin can restore it from the archived view. Every archive stops for approval under every setting, and the card says what goes with the record. It archives only the record you named, never a related one.Can it contact vendors?
Only when you ask, and every send stops for your approval under every setting. The sends are: a questionnaire or portal invitation, a commitment reminder, a round of follow-ups, an email drafted on a vendor or assessment, an email from a Radar case, a supplier portal invitation, reopening a submitted request, and launching an assessment plan that contacts the vendor. The card shows the recipients, the vendor and the subject. A send is never part of a batch, and the report lists it under What cannot be undone.Can it contact our own people?
Yes, in the ways the product already does. A note it posts notifies the colleagues it mentions (at most five), a review request reaches the reviewer’s work queue, a finding owner gets the usual notifications, and sending an intake request back notifies or emails the requester. Sending an intake request back always asks.Can it approve a vendor or accept risk?
Not on its own. It approves an intake request, completes an assessment with an outcome, or records a Zero Touch decision only when you ask, and each is a checkpoint (approving an intake request always asks). It never closes a finding or accepts its risk, never sets a register entry’s treatment to accept, and records a control as risk accepted only when you or the evidence says your organization already accepted it. See Changes Coverbase AI never makes.Can it change who has access to Coverbase?
No. It never changes roles, permissions, members, sign-in, SSO, IP allowlists, API keys or billing.Can it create a vendor outside intake?
Only when you ask it to add a vendor directly. By default it onboards through intake. A vendor added directly starts in the Created status, which is not approval, and a vendor by the same name already on file is returned instead of a duplicate. An admin can set Adding the vendor to Always asks or Off.Access and visibility
Who can use it?
Active members with the Admin, Member or Siloed Member role, once your Coverbase team turns Coverbase AI on. What a run can do is limited to what that person’s role allows: a Siloed Member’s run reaches only their own book of work. A run started by someone with another role, such as a custom role, Auxiliary Member or Guest, ends at its first step without doing anything. API keys cannot start runs, because a run must act as a person. See Permissions and roles.What can it see?
Exactly what the person who started the run can see, checked again on every step. That includes the record they started from, the records and files they give it, your organization’s Coverbase AI memory, and public web results when a question reaches outside Coverbase. When your organization enforces roles and permissions and a person reaches only records assigned to them, a run reaches only those records.Can it see other organizations’ data?
No. Runs, their logs and their memory belong to your organization and are isolated like every other record. See Tenant isolation.Who can see a run?
Only the person who started it. A run page is private. Anyone with a role that can read organization settings can see the changes it made in the Agent Ledger, and the ledger links to the run only for its owner.Can a file or a web page tell it to do something?
No. Text inside an attached file or a web result is treated as information, not as instructions. A file that says “archive this vendor” does not make the run archive anything.Turning it on and off
How is it turned on?
Your Coverbase team turns it on for your organization during Beta. Until then, Coverbase AI does not appear in the product and no run can start.How is it turned off?
Ask your Coverbase team to turn it off for your organization. Short of that:- An admin sets any change to Off, and no run makes it.
- Pause every agent under Configuration > Coverbase AI Agents stops every standing agent at once. Nothing starts and live agent runs stop before their next step.
- Each person stops their own runs with Stop.
Can we limit how much it runs?
For standing agents, yes: Runs per Day, Spend Cap per Day, quiet hours and the playbooks each agent may use. A run a person starts has no daily cap, and it is limited by that person’s permissions and your organization’s policies.How do we stop a run that is going wrong?
Click Stop on the run page or the top bar card. A run waiting on you stops at once, and a working run stops within a few seconds. Nothing more is written. You can also type a message under the run to change its course. A message while it waits on a checkpoint withdraws the checkpoint.Review and audit
How do we review everything it did?
Three places:- The report of each run lists what it did with a link to each record, the to-dos it left, what it could not do, and What cannot be undone.
- The Agent Ledger under Configuration lists each change to a record under the agent Coverbase AI, with the record, the run, who asked and who decided. Filter it to one run or one standing agent, and export up to 50,000 rows as a CSV for an audit.
- The run’s log, every step, question, answer and change, can be downloaded with Export on the run page.
How does the ledger show who approved a change?
A change approved at a checkpoint, in a batch or with Approve all like this reads Accepted with the person’s name. A change applied under Don’t stop to ask reads Applied under org policy. A change that had no checkpoint reads Awaiting review until a reviewer accepts or rejects it. A declined checkpoint is a Rejected row with what the person typed.If we delete a run, do we lose the audit trail?
No. Deleting a run removes the run and its log, and only an archived run can be deleted. The records it created stay, and its Agent Ledger rows stay.Mistakes and reliability
What happens if it makes a mistake?
You correct the record the way you correct your own edit, from the record’s page. The report and the Agent Ledger link every record the run changed. An archive is restored by an admin. When you decline a checkpoint, untick a proposed change or move a risk level and say why, Coverbase AI keeps a preference, so later runs do not repeat the mistake. There is no single button that reverses a whole run.Does it make things up?
It is instructed never to invent a vendor, a score, an answer or a result. When research does not support an intake answer, it leaves it blank and says what evidence would answer it. Every record it cites is a link, and every claim from the web names its source. It can still be wrong, which is why the composer says “It can make mistakes, so check the sources it cites”, and why changes that matter stop for approval.What if a run is interrupted?
It picks up where it stopped and never repeats a change it already made. When nothing can prove whether a change landed, it refuses to make it again and says in its report what to check by hand. A run interrupted twice ends with a report saying so.Can it run for days?
Yes. A run waiting on an assessment records what it waits on, frees its place on the server, and resumes when the assessment finishes. You are notified when it needs you and when it is done.Models and data
Which model runs it, and where?
Anthropic Claude Opus 5.5 through AWS Bedrock in Coverbase’s AWS account. When Bedrock reports no capacity for it, a run falls back to Claude Sonnet 5.5. Bedrock’s cross-region inference can serve a request from an AWS region other than the one Coverbase runs in. See AI governance.Is our data used for training?
No. Customer data sent to the models Coverbase uses is not used to train those models. See AI governance.Is personal data sent to the model?
The records and files a step needs are sent to the model. Turn on Redact personal data before model calls on the Agent Ledger page to replace email addresses, phone numbers, account and tax numbers, and the full names of your users and your vendors’ contacts with placeholders before a model call, including Coverbase AI’s. See Redacting personal data.Does it send our data to the web?
Its web searches go to an external web search provider (Tavily). A search is one specific question the run writes, such as a vendor’s recent incidents, with the vendor’s name in it. Other features that research vendors on the web work the same way. See AI governance.What happens to the files we attach?
Each file is uploaded to your organization as a document you uploaded, then read by the run. A run takes up to 100 files a message, 50 MB each, and 200 a run. See Files and records.What does it remember?
Preferences your team taught it, in plain words built from corrections: a risk level moved with a reason, a declined checkpoint, a change kept as it was, or an instruction that starts with “always”, “never” or “from now on”. They belong to your organization only. Admins can reword, rescope, turn off or delete each one under Configuration > Coverbase AI Memory.Troubleshooting
Related
Controls, approvals and governance
The detail behind each answer.
AI governance
How Coverbase selects, tests and constrains its models.
Agent Ledger
Read, decide on and export every agent action.