Bland
Other people are letting Grok Bot dial. Bland is the service they named. The useful question is which of our calls should ever leave a human mouth.
← AI Agents · Tech · GrokBot TOC · Require Approval · Granola · Keychain
Why this page
The note that started this: people are using Grok Bot to make outbound phone calls, and the name they gave was Bland (bland.ai). A bot can already draft the email. A phone call is louder. A bad one to a sponsor costs more than the minute rate.
This file is the sitting, not a vendor review. What Bland is, how it actually attaches to Grok Bot, which of our jobs fit, and the first call — which is to your own phone. Not a paid placement. Pricing and caps move; the links are the source.
One-sentence TL;DR
Bland places a real phone call from a task and a voice, then hands back a recording, a transcript, and whatever you asked it to collect. Grok Bot can drive that if the connection is real. Use it on yourself first, then on US admin calls you would have made anyway. Do not point it at Chiang Mai, a cold list, or a VIP.
What Bland actually is
You give it a number, a task (the goal, not a word-for-word
monologue), and a voice. An agent dials, talks, and can be
interrupted. When the call ends you get a call_id,
a transcript, a recording if you asked for one, and a summary.
Docs:
quickstart
·
agent lifecycle
·
docs.bland.ai.
It also answers inbound on a number you rent, and on the Agent
Phone Plan it can text. Texting is a conversation: you pass an
objective, and Bland keeps replying toward that
goal. That is a different product surface. This page is about
calls.
Opposite tool on this site: Granola listens to a call you are already on and enhances your notes. Bland is the thing that picks up the phone when you are not in the room.
Grok Bot does not have a Bland button
There is still no connector in a plugin list you can toggle. What exists now is Bland’s own guide: docs.bland.ai/integrations/mcp/grok-bot. Read that before anyone pastes a key. Two paths, and a caveat they print on the page.
| Path | What you do | Catch |
|---|---|---|
| Let the bot sign up | It reads their Grok Bot doc, texts you a code, you relay the code. Paid path uses Stripe Link. | A Link virtual card is refused. It needs a shared payment token from the Link CLI, or you open the link they show you. The bot must not open that page itself. |
| You already have a key | Custom MCP server https://api.bland.ai/v1/mcp with header Authorization: Bearer … | Bland says the documented Plugins flow does not prove Grok Bot can add an arbitrary MCP URL plus a header. Confirm on your build before you depend on it. |
If the header controls are not there, don’t invent them. Their fallback is a client that documents the setup (they name Poke and OpenClaw), or the REST API from the bot’s computer with the key stored like any other vendor key — never pasted into the chat. Revoke from Bland → Settings → API Keys.
Their MCP marks create_call as destructive: a
well-behaved client asks before it spends money or dials. That
is the same fence as
Require Approval. A routine that
“just calls the list at 9am” is the failure mode, not a feature.
Tool list:
MCP tools.
Without a paid plan the org starts on a couple of dollars of credits and dials from a shared Bland number. No texting. Fine for the test call to yourself. Not fine as the caller ID a sponsor sees.
Where I would actually use it
Ranked by how little it can embarrass us. The Agent Phone Plan
(~$15 the first month, then $30) only calls and texts the
US and Canada, one call at a time. A Thailand
number comes back blocked
(CALL_PLAN_DESTINATION_BLOCKED). Build or Scale can
dial further, at a per-minute rate plus telephony. That is a
different buying decision, not a toggle.
| Job | Use Bland? | Why |
|---|---|---|
| Call your own phone | Yes. This is call one. | You hear the voice, the pause, the “I’m an assistant” line. If you would hang up, nobody else gets it. |
| US admin you would have dialed anyway | Yes, after the self-call | Restaurant hours, “did the shipment leave,” a venue’s Tuesday availability. You are the customer. Still say it’s an assistant. |
| CBS attendance or dietary needs | Not the Thailand dinners | Chiang Mai and Bangkok numbers are outside the phone plan. Those people already have email. Email Quarterback stays the desk. A US attendee who already agreed to a call is a maybe, one name at a time. |
| Sponsor who went quiet | Later, maybe never | Warm is not the same as “fine if a bot fumbles the name.” Draft the script. You place it, or you approve one call after you have heard the voice on a stranger-free test. |
| Cold outbound | No | US rules treat an AI voice as an artificial voice. Consent first. Reputation either way. This is not a lead dialer. |
CBS GTM still wants a human on sponsor calls. An on-stage co-host (CBS avatar) is a different job from a phone agent. Don’t merge them.
The rule we already use for messages
The bot drafts two things. You approve both. Then one call goes out. Not a batch.
- The task: who it is, why it’s calling, what “done” means, when to hang up.
- The list: one number, with a reason that number agreed to hear from us — or it is your phone.
A task that matches the first test. Point it at yourself:
You are an assistant calling for Mike Michelini. Say that in the first sentence.
Ask if this is a good time. If not, apologize and hang up.
Ask one question: does Thursday at 7 still work, and is there anything you don't eat.
Do not book anything. Do not mention other guests.
If they say stop, or ask for a person, apologize and end the call.
Do not invent a confirmation. After it hangs up, read the transcript before anyone else is on a list. Gray-area answers (“I think they said yes”) are a no. Same skim as a Granola note: 60 seconds, numbers and names only.
Two fences
The rule, not a legal memo. In February 2024 the FCC said AI-generated voices count as an “artificial or prerecorded voice” under the TCPA, so those calls need the called party’s prior express consent unless an exemption applies (FCC 24-17). The headline “robocalls are illegal” is about scams and missing consent, not “a business may never use an AI voice.” Other countries have their own rule. Thailand is not this ruling. This page is not advice — if a list is large or cold, ask a lawyer, don’t ask the bot.
Bland advertises TCPA rails and disclosures on outbound. That is their guardrail, not yours. A platform setting does not create consent you never collected.
Your name. The first sentence is “I’m an assistant calling for Mike.” Not a cloned voice of you. Not silence about what it is. A clumsy call to someone who writes checks is more expensive than dialing it yourself. Same lesson as the gym hack: write access with no fence is how an agent books the wrong thing.
What it costs, as of this writing
Check bland.ai/pricing and billing before you put a card down. Snapshot while writing this:
| Plan | Rough price | Fit for us |
|---|---|---|
| Start | $0 platform, about $0.14 per connected minute. ~100 calls/day, 10 at once. | The experiment. Shared or existing caller ID. Credits are a couple of dollars until you pay. |
| Agent Phone | $14.99 first month, then $29.99. “Unlimited” US/Canada minutes inside hard caps. | One dedicated US number. One call at a time. Docs say 20 calls/hour and 50/day; the marketing page says 60/hour and 500/day. Believe the number on your billing page. Not a blast dialer either way. |
| Build / Scale | $299 or $499 a month, then about $0.12 or $0.11 a minute, telephony extra. | Only if we truly need parallel calls or a number outside the US and Canada. Not week one. |
Per-minute plans bill connected time to the second. Transfers to a human cost extra on Bland’s numbers. Bring-your-own Twilio is a later argument, and Twilio is the other rail if Bland’s Grok Bot header never shows up — you still write the task, you still approve the list. ElevenLabs and voice notes on this site are speech and desktop voice, not a phone agent. Don’t shop them as a substitute until someone has actually placed a call.
Week-one sitting
- Read their Grok Bot page. Confirm this bot can hold an MCP URL and an Authorization header. If it can’t, stop and use the dashboard “send a call” button for the test. Don’t debug a plugin that isn’t there.
- Account in the browser, or their headless flow where you only relay an SMS code. Card stays in your hand. Key stays out of the chat.
- Call your own phone with the task above. Listen all the way through.
- Read the transcript. If it talked over you, invented a yes, or forgot to say it was an assistant, fix the task and call yourself again. Nobody else.
- One US business call you were going to make anyway. Then decide if a second name exists. Sponsors wait.
What we keep vs skip
- Keep: task + one number + your approval; assistant disclosed in sentence one; self-call before any other human; transcript before the next dial; US/Canada on the cheap plan; email for Thailand logistics.
- Verify: whether this Grok Bot build can add the MCP header; which cap your account actually has (20/hour vs the marketing page); international rates before anyone talks about Bangkok.
- Skip: cold lists; a cloned voice of you; batching the dinner list; a 9am routine; pasting the API key in chat; treating Bland’s TCPA toggle as consent.
Related on this site
- Grok Bot · table of contents — the desk this plugs into
- Require Approval — same fence as their destructive
create_calltool - Grok Bot course — the workshop said calls were “maybe.” This is the maybe, narrowed.
- Keychain — if the key has to live on this Mac
- Granola — capture a call you’re on; Bland places one you aren’t
- Email Quarterback — default for attendance and dietary, especially outside the US
- CBS GTM · MCP · gym hack
Primary: bland.ai · Grok Bot setup · Agent Phone Plan · FCC 24-17
Comments
Approved comments appear below. Log in once with GFAVIP — it applies across the whole site. GFAVIP login
View comments archive