Build a small test, check the result, and know which operations need a live provider.
Start with one customer question
Choose a guide below to set up an agent, add business information or review a follow-up. Use sample data while learning.
Open the workspace to see available sign-in options. Live calls, model answers and messages each need their own provider setup.
The product, in practice
See each step. Try it in your workspace.
Short recordings show how to set up an agent, test a request and review the result. Follow the captions or read the steps below.
Build, test, and improve an agentCAPTIONED DEMO ↗
Silent recordings with captions. Sample data is used; these are setup walkthroughs, not live calls.
Read the chapter transcript
0:00
Give a sample agent one clear reception task.
0:07
Keep the greeting and boundaries explicit, then save.
0:14
Open the conversation workflow and inspect its connected steps.
0:21
Validate the graph, then publish an immutable version.
0:28
Run a local customer question without placing a call.
0:35
Read the saved transcript and identify one useful improvement.
0:42
Save feedback, then use it to guide the next revision.
Speak or type your setup
Create your clinic reception setup
Start from one clinic admin job, then review what your assistant should do.
Open the workspace and choose an available sign-in method. Sign-in options depend on how the workspace is configured.
Open Talk to set up. Apply the clinic admin starter, then replace it with real clinic services, hours, fees, address and staff callback details.
Answer or edit the short questions by voice or typing. Keep it admin-only: no medical advice, no symptoms/history collection by default, and no confirmed appointment without staff or provider proof.
Choose Review setup, check every answer, then save. Open Test playground to try the saved assistant. Edit its instructions in Agent studio whenever needed.
What to expect
The workspace saves an assistant, clinic source, workflow and labelled sample conversation. Connect a voice provider for a real browser conversation. Phone numbers and follow-up channels have their own setup.
Local simulation
Test a conversation and record feedback
Check expected paths before trying a real model or voice channel.
Open Test playground and select the agent you want to check.
Enter a sample customer message and choose Run simulation. Read the transcript and workflow events.
Try the handoff, opt-out, and instruction-attack fixtures. Run the four regression cases to check the defined paths.
Choose Review & leave feedback. Record what needs to change, then edit the agent or workflow and test again.
What to expect
Simulation results are saved locally. They do not establish real model quality, live call latency, or protection against every attack.
Local retrieval; model tests gated
Add knowledge and check source matches
Give the agent business information you can review.
Open Knowledge and choose Add a source. Paste text or import a .txt or .md file. Add a clear source title.
Save the source. Use Preview retrieval to search a likely customer question and inspect the matching passages.
In Grounded chat, choose Configure model & sources. Select only the documents that this agent should use.
Keep supporting evidence required when answers must come from those documents. For written answers, save your own text-model key in Integrations when available, then choose that provider. The provider bills your account when you send a question.
What to expect
Source search calls no model. Source URLs are labels, not crawled pages. Import text or Markdown; PDF import is not supported. Review answers against the original information.
Local steps work without a provider
Publish and run a local automation
Run a versioned process and inspect each step.
Open Automations and choose Create automation. Start with Local input example or Support routing example.
Review the graph JSON and save the draft. Choose Validate, then Publish to freeze a version.
Choose Run, select a published version, and enter the manual input JSON. Review the exact version and input before submitting.
Read the saved run trace. Model steps require live runtime access. An action draft pauses the run for separate review.
What to expect
These automations execute real local steps. They are separate from call-workflow simulations. Manual runs do not create a schedule.
Local review; execution gated
Review an email, calendar, or WhatsApp draft
Check the exact action before it can leave the workspace.
Open Action approvals and choose Draft an action.
Enter the recipient and content, or the event and time zone. Record the actual consent reference and applicable purpose.
Save the draft. Review the exact payload and choose Approve exact action only when it is correct.
Execution is a separate control and remains disabled until the operator has configured access. An uncertain outcome needs receipt review, not a replacement request.
What to expect
Saving or approving a draft sends nothing. Drafts expire. A confirmed provider receipt is needed before claiming the action succeeded.
Local permission ledger
Record consent and respect an opt-out
Keep a customer address separate from permission to contact it.
Open Contacts and add an email or phone identity. Creating it does not grant permission.
Open its permissions and record an actual grant or revocation. Include the channel, purpose, evidence reference, source, and capture time.
A revocation suppresses the channel. A later grant leaves that suppression in place.
Reinstatement requires a fresh grant after suppression and an explicit review of the current suppression record.
What to expect
The ledger stores operator-supplied evidence. It does not independently prove lawful consent. Deletion preserves hashed suppression safety records.
Customer key setup
Connect your own model or voice account
Use your provider account and see what each connection enables.
Open Integrations. Choose a text provider for written answers or the separate xAI Voice connection for browser conversations. Saving a key does not connect Gmail, WhatsApp or a phone number.
Store a key from an account you control. It is encrypted and marked unverified. Saving costs no model usage; your provider bills your account when you actually use its model or voice.
For text, open Knowledge and configure the model and business sources. For voice, open Test playground and select Browser voice, your agent and xAI. Read the account and microphone notice before starting.
If a control is unavailable, follow its setup message. A saved key still needs the right provider permissions and balance. She Said AI plans and provider usage are separate charges; review both before starting.
What to expect
After a real conversation, review its saved transcript in Conversations. Browser voice does not receive telephone calls or transfer to a staff phone. External actions need their own connections and review.
Local data controls
Export or delete workspace customer data
Understand what is removed and what safety records remain.
Open Settings, then Data controls. Download a workspace JSON export if you need a review copy.
Read the deletion scope before confirming. It includes saved conversation and action content, automation run payloads, and contact details and evidence.
Contacts are suppressed before their details are deleted. Hashed safety records prevent deletion from silently restoring contact permission.
Workspace configuration, knowledge sources, published definitions, and audit or safety receipts have separate retention. Backups and provider-held data need their own process.
What to expect
Deleting local data does not reverse a sent message, end a provider operation, or delete every copy held elsewhere.
Common questions
How do I sign in?
Open Sign in and choose Google or your workspace email and password. Use the same method as your existing account. Ask support if you need to link a different method.
Why is Send to model or live calling disabled?
Read the message beside the control. You may need to save your own provider key, choose a model or connect the required service. Demo mode blocks paid provider activity. Using a platform-funded account requires separate access.
Why did publishing not change a previous run?
Published automation versions are immutable. A run records the version it used. Publish a new version and start a separate reviewed run to test a change.
Can I retry after a timeout?
A timeout does not establish that the provider did no work. Use the retained request and receipt controls. Do not change its operation key or create a replacement until the outcome is reviewed.
Does saving a provider key verify it?
No. The key is stored encrypted and marked unverified. Saving it does not test model access or confirm the provider account.
Why is a recreated contact still suppressed?
Deletion removes identity and evidence content but keeps hashed safety records. Recreating the same address does not erase its opt-out history.