Guides
Guides

Limitations

What the connector will not do, and what to expect during early access.

Service account, for now

Actions in your Salesforce orgs are performed by the nCino Communicator user, a dedicated service account, rather than as the person signed in.

  • The account is API-only. It cannot sign in to the Salesforce user interface at all.
  • It has broad access. Its permissions include View All Data and Modify All Data, through a standard Salesforce permission cascade, so the connector can reach anything in a registered sandbox. Don't register a sandbox that holds data you are not willing to share with your AI provider.
  • Its access is visible and controllable in Setup, like any other Salesforce user, so your security team can inspect exactly what it can reach.
  • Audit logs record the service account name, not the person who asked. Pair those with your AI assistant's own conversation logs to establish who initiated an action.

Per-person sign-in to your orgs is planned. Once it ships, work in your org runs as the person who asked instead of the service account.

Audit

Salesforce records the connector's changes under the nCino Communicator user.

  • Metadata changes, such as picklist values, page layouts, and permission sets, appear in the Setup Audit Trail like any other metadata change.
  • Record-based configuration, such as nCino system properties, and other records the connector creates or updates, do not appear in the Setup Audit Trail.
  • Salesforce Shield (Event Monitoring) captures the connector's API activity, including reads.

Standard Salesforce does not log record reads. That is Salesforce behavior rather than something specific to this connector, but it matters here because most of what the connector does is read. If you need an audit record of reads or of record-based changes, Shield is the mechanism.

Sandbox only

Every tool that acts on an org is checked against your registered sandbox orgs before it runs, reads included. Production orgs cannot be registered, so they are never reachable. Production support is on the roadmap and will come with its own controls.

Approval for changes is set in your AI assistant

Your AI assistant decides whether to ask you before running a change, and nCino cannot enforce that. nCino's change tools are written to summarize what they intend to do and ask first, but the gate that cannot be bypassed is the one your assistant's administrator configures.

Approval is coarse today. Your assistant sees a small number of nCino tools rather than one entry per capability, so approval settings apply to nCino as a whole. You cannot currently approve reads automatically while requiring confirmation only for changes. That leaves two practical choices:

  • Require approval for nCino actions. Your administrators confirm each step, including read-only ones. Safest, and the right default for a first rollout.
  • Allow them to run without a prompt. Faster, and appropriate only once you trust the access list, since it also allows the change tools.

Per-capability approval is a known gap and is planned. Until it ships, your access list is doing most of the work.

Where to set it:

  • Claude: per-connector tool permissions
  • Copilot Studio: Ask the end user before running, under Additional details on the tool (set to No by default)
  • ChatGPT: the app's Permissions setting, such as Always ask, which a workspace admin sets

Change tools are off by default

The change tools for feature flags, system properties, declarative metadata, and records come turned off, so the connector starts out read-only. To use them, ask your Customer Success Manager, and nCino turns them on for your institution.

Everyone with access sees the same capabilities

Access follows your identity federation. Anyone your institution federates and enables can use the connector, and you grant and revoke that through your identity provider.

A session that is already connected can keep renewing for up to about 30 days after you disable someone in your identity provider. When someone leaves, also remove them from your AI assistant.

The connector does not yet apply per-user or per-role scoping of its own, so everyone who can reach it has the same capabilities against your registered sandboxes. Role-based permissions are planned. Until they ship, treat your access list the way you would a list of sandbox system administrators.

Repeated sign-in prompts in some assistants

Some AI assistants ask you to sign in again during an active working session rather than staying connected.

Where we have seen this:

  • Microsoft Copilot Studio. Connections may drop and require a fresh browser sign-in. Reported intermittently and not consistently reproducible.
  • The mcp-remote bridge. A sign-in page may open, though tools generally keep working.

If you hit this, sign in again when prompted. Nothing in your environment is affected and no configuration is lost. Report it through submit_feedback or your Customer Success Manager, and include your assistant, its version, and roughly how often it happens. The frequency is the most useful detail.

nCino is actively working on this. Behavior may change during early access.

Other things to know

  • Early access. The connector is experimental and provided "as is." What it does and whether it is available may change.
  • Where it is available. Institutions in nCino's United States, Canada, EU West, EU East, Japan, and Oceania trust zones can use it. The Middle East trust zone is not available yet.
  • Batched changes are not all-or-nothing. If one item in a batch fails, other items may still take effect. When a change matters, ask the assistant to confirm the resulting state afterward rather than assuming the whole batch succeeded or the whole batch rolled back.
  • Large answers. Queries return large results in pages, and the assistant can ask for the next page. Other large answers may be shortened, so ask about a specific object, field, or release to get everything.
  • Feature flags change the flag only. Turning on an nFORCE feature flag does not run the Enable Feature step in Feature Management. If a feature is normally turned on there, check Feature Management after the change.

Did this page help you?