AI Copilot modes, guardrails, and rollout

Settings → AI Copilot controls workspace-aware policies for AI-assisted replies, qualification and routing, follow-up, scoring, summaries, and call intelligence. Some settings can cause customer-facing actions or change CRM records, so agree on an owner and test with internal records before enabling autonomous behavior.

The same settings page can also appear from CRM → AI Copilot. Confirm the active workspace or sub-account before making a change.

Availability and access

The current AI Copilot screen identifies the feature as available on Starter and above. Access can also depend on the active workspace, member permissions, and an agency-controlled sub-account capability. If the navigation item is missing, the page is locked, or a direct link is unavailable, ask a workspace or agency owner to verify the plan and capability rather than trying to bypass the restriction. See feature availability.

When the screen shows its plan-lock banner, the controls are disabled. The banner states that existing settings are preserved and resume after an upgrade. Review those saved modes and guardrails before restoring availability so an old Auto policy does not resume unexpectedly.

Only change these shared policies when your role authorizes you to do so. Users who do not own workspace configuration should stay within the draft, summary, and conversation-review controls their role exposes.

Understand the two kinds of controls

Three action flows have Off, Draft, and Auto modes:

• Instant replies handles inbound lead messages.
• Auto-qualify & route qualifies leads and can enroll them in active sequences.
• Follow up cold leads responds to leads that have been silent after the last message.

The modes mean:

• Off: the flow does nothing.
• Draft: Copilot suggests an action and a person approves it.
• Auto: Copilot acts when the saved guardrails pass and escalates otherwise.

Choosing Auto opens a confirmation dialog. Read it, confirm the intended flow, and select Turn on only after checking its saved guardrails. Canceling the dialog leaves Auto off.

Three additional capabilities use Off and On, not Draft and Auto:

• Lead scoring: computes AI intent and qualification during enrichment when enabled.
• Conversation summaries: makes cross-channel, on-demand thread summaries available when enabled.
• Call intelligence: after a voice call, can use the transcript to update the lead, save a summary note, and create suggested follow-up tasks when enabled.

Treat generated scores, summaries, transcript-derived updates, and tasks as material that still needs human review. A poor transcript or incomplete conversation can produce incomplete context.

Know exactly when a change saves

There is no page-level Save button.

This distinction also applies when a panel already contains unsaved edits:

1. Changing the row-level mode saves the new mode with the flow’s last-saved guardrails.
2. The unsaved panel edits remain staged; the mode click does not include them.
3. Select the panel’s Save button to persist those edits separately.

Wait for the Saved confirmation after each operation. If the page reports Failed to save, do not assume the visible selector is the persisted value. Refresh, re-check the flow, and retry only after resolving the error.

Top-level Off/On selectors for lead scoring, summaries, and call intelligence also save on click. Toggles inside the expanded Instant replies, Auto-qualify, or Follow-up panels are staged guardrails and require that panel’s Save.

Configure the common guardrails

Expand an action flow to configure its limits. Instant replies, Auto-qualify, and Follow-up expose these common controls:

• Min. confidence: the minimum confidence required for autonomous handling.
• Max auto-actions / day: a daily cap for the flow. Leaving the field blank represents Unlimited in the current screen.
• Business hours only: restricts the flow to the business-hours rule used by the policy.

For a flow without a previously saved policy, the screen currently initializes the common guardrails at 80% minimum confidence and 25 auto-actions per day. Existing workspaces may have different saved values, so use the displayed, saved policy as the source of truth.

Start with a high confidence threshold and an explicit low daily cap. Do not clear the cap during an initial rollout. Save the panel, refresh, and confirm the values before changing its mode to Auto.

Configure Instant replies

The Instant replies panel adds two controls beyond the common guardrails:

• Send replies autonomously (WhatsApp): permits eligible WhatsApp replies to send autonomously when the flow is in Auto and its saved checks pass.
• Booking link: a link Copilot can send when it detects demo intent.

Autonomous reply sending currently applies to WhatsApp. On email, SMS, Messenger, and Instagram, Copilot drafts a reply for a person to review and send. The screen also states that pricing and commitment questions are escalated to a human rather than answered autonomously.

Verify the booking URL in a private browser and confirm it opens the intended calendar, event type, duration, and owner. A syntactically valid link can still lead to the wrong booking experience.

Configure qualification and routing

The Auto-qualify & route panel can map confidence thresholds to sequences. Only active sequences are loaded into the sequence picker. If none are active, the panel prompts you to create one before adding an enrollment tier.

For each tier:

1. Set the minimum confidence percentage.
2. Choose the intended active sequence.
3. Check that thresholds do not overlap in a confusing way.
4. Review the selected sequence’s audience, sender, steps, and stop conditions.
5. Save the Copilot panel.

Re-enrollment can restart a sequence at step 1, so inspect existing enrollments before enabling automatic enrollment. Follow the safeguards in Sequences.

The same panel can Escalate hot leads to a human. Its controls include an escalation confidence threshold, the Purchase, Demo, and Pricing intents, and a notification toggle. The screen describes assignment as round-robin when a sufficiently confident, high-intent lead qualifies. Confirm that the relevant team can receive and act on the resulting handoff before enabling it.

Configure cold-lead follow-up

The Follow up cold leads panel adds:

• Days silent before follow-up
• Max nudges per lead
• Smart send-time

It also uses the common confidence, daily-cap, and business-hours controls. In Auto, eligible nudges are currently sent automatically on WhatsApp. On other channels, Copilot creates a follow-up task for a person instead of sending autonomously.

Set a small nudge limit, account for any active sequence or manual follow-up, and test with records that represent replies, opt-outs, bookings, and disqualification. Check both the conversation and task list so multiple systems do not follow up with the same person.

Preview before enabling Auto

Each expandable action flow provides Preview what would auto-send. The preview is read-only: it samples recent inbound conversations, classifies each as Would auto-handle or Would escalate, and can show the reason an item would escalate. Opening the preview sends nothing.

Preview evaluates the saved policy, not uncommitted fields in the panel. Save any guardrail changes first, then refresh the page before using the preview as evidence for the new configuration. If there is no recent inbound activity, the preview reports that there is nothing to sample; it is not proof that the policy is safe.

Review every available sample for:

• Incorrect identity, intent, or channel assumptions.
• Pricing, commitment, consent, or policy language that needs a person.
• Missing conversation history or attachment context.
• An unsafe enrollment, follow-up, or escalation outcome.

Read the forecast and approval history

When enough history exists, an expanded action flow shows a forecast based on the last 30 days. It includes:

• Total AI suggestions.
• An estimated number that would auto-send per month under the evaluated policy.
• The percentage of drafts you approved.

The approval percentage is an aggregate trust signal derived from prior approve/dismiss behavior; the settings screen does not present it as a row-by-row approval log. If there is no approval history, the percentage is blank and the screen recommends reviewing several drafts before enabling Auto. If there are no suggestions, the forecast box is not shown.

Save changes and refresh before interpreting the forecast for a revised policy. A strong historical approval percentage is useful evidence, but it does not replace reviewing high-impact edge cases.

The strip at the top of the page summarizes the last 24 hours: drafts, qualified leads, enrollments, escalations, and follow-ups. Use it for a quick volume check, not as a complete audit trail. Inspect the underlying Inbox conversations, CRM records, sequence enrollments, and tasks when validating a result.

Use a controlled rollout

1. Confirm the active workspace, authorized owner, connected channel, and downstream sequence or task dependencies.
2. Leave the flow Off while configuring and saving its initial guardrails.
3. Use Draft long enough to build real review history and find failure cases.
4. Save conservative confidence, business-hours, and daily-cap settings.
5. Refresh, then inspect both the preview and 30-day forecast where data exists.
6. Turn only one flow to Auto, accept the confirmation, and verify the Saved result.
7. Monitor the 24-hour summary and underlying customer records after a small amount of live activity.
8. Increase scope or limits only after the observed results meet the workspace’s written acceptance criteria.

Do not enable Instant replies, qualification enrollment, and cold-lead follow-up together on the first test. Changing one behavior at a time makes an unexpected result easier to identify and contain.

Pause or roll back

For a workspace-wide rollback, select Draft or Off on the affected flow and wait for the Saved confirmation. This mode change saves immediately with the last-saved guardrails. Turning a separate capability Off also saves immediately.

For an individual messaging conversation where reply Auto is active, the Inbox composer can show Take over. This pauses autonomous replies for that thread so a person can handle it; Resume AI returns that conversation to autonomous handling. Use the thread-level pause when one conversation needs intervention without changing the whole workspace policy.

Changing a mode does not undo actions already taken. After rollback, review outbound messages, sequence enrollments, CRM updates, notes, and created tasks. Correct any inappropriate state, document what happened, and leave the flow in Draft until its guardrails or dependencies have been revalidated.

For operating and recovery guidance, continue with Inbox conversations, Sequences, custom fields and lead scoring, and communication troubleshooting.