Voice AI for high-stakes calls.
Bland is fully self-hosted and built for the hardest calls a business takes. Its agents sound like people, go live almost instantly, and get better on their own.
Bland owns every layer of the call.
Most voice AI rents its parts from different companies, and every call hops between them. Bland runs all of them itself, on its own GPUs.
- 1Telephony
- 2Speech to text
- 3Language model
- 4Voice
- 5Agent
Your own instance. Nothing shared.
Enterprise customers run on dedicated infrastructure. Another company’s traffic can never slow your calls.
- 1GPUs
- Reserved for your calls. Nothing else runs on them.
- 2Releases
- Pinned to you. When you promote one, it rolls out with zero downtime, starting at 1% of calls.
- 3Network
- Static IPs you can allow in your firewall.
Sounds like a person. Answers like one.
Powered by Bland Speech, the first Human Speech Engine (HSE). It’s trained for phone calls, not podcasts or social media, and it replies in about 200 milliseconds.
Lock the steps that can’t go wrong.
Everything between them stays a natural conversation.
One click to ship. Tested first.
A version can’t go live until its checks pass. Then it takes more calls in steps.
Share of live calls on the new version. Staged releases are on enterprise plans.
See every call, measured.
Where calls come from, how many the AI handled, and how long they took, for every agent and every day.
Every bad call makes it better.
Flag a call. Norm finds the cause, fixes it, and tests the fix on the call that failed.
- 1A call goes wrong
- 2Norm finds the cause
- 3Norm drafts a fix
- 4The fix is tested on that call
The phone network is built in.
Numbers, SIP trunks, and transfers run on Bland’s own stack. No separate carrier to wire up.
Text, call, or chat. Same agent.
A customer texts, switches to a call, then picks it up in chat on their laptop. The agent remembers the whole thread, so nothing gets repeated.
Example conversation. Texting is available on enterprise plans.
Your dispute for the $84 charge is open. Case 4417.
Thanks. Can you call me? I have a question.
Calling you now.
Hi Dana, it’s about case 4417. What’s your question?
- VoiceInbound and outbound calls on your numbers.
- SMSTexts, plus RCS and iMessage, with the same memory.
- ChatA chat widget on your website, with handoff to your team.
Built to pass your security review.
SSO, roles, signed webhooks, and audit logs, with the paperwork your legal team will ask for. Trust Center
Everything in the platform.
42 features across six areas. Read the docs for the rest.
01Build10
- Agent builderStart from one prompt and add sub-flows only where the call needs them.
- PathwaysSteps that always run in order, with edge labels, loop conditions, and global steps.
- Authentication zonesIdentity checks that gate every flow nested beneath them.
- Flex ModeLet the agent skip ahead or backtrack between steps when the caller does.
- ToolsWebhooks, custom code, and integrations, with secrets and per-tool failure rates.
- Custom code and scheduling stepsEnterpriseRun JavaScript on Bland, or book times against your calendar.
- KnowledgeFiles, synced websites, Notion, and Google Docs, with query logs and unanswered questions.
- MemoryFacts, open items, and a running summary for every contact, across voice and text.
- NormDescribe an agent in plain English; Norm builds and edits it in a draft.
- API, CLI, and MCPBranches, environments, and experiments by API; agents as YAML in your repo.
02Voice7
- Bland SpeechRealtime voice that streams as the model writes and stops cleanly when interrupted.
- Performance tagsAbout forty cues like warmly, laughs, and sighs, plus pause and pace control.
- Voice cloningAn instant clone from about ten seconds of audio.
- Voice rotationUp to twenty voices per agent, one chosen per call.
- Babel transcriptionBland’s own multilingual speech recognition, with keyword boosting. In beta.
- Turn-takingFive interruption levels; a brief “uh-huh” doesn’t stop the agent mid-sentence.
- Live translationSpeech to speech across 23 languages, including during transfers.
03Test and ship6
- EvaluationsJudges that grade transcripts or audio, run against simulated callers.
- Tests from real callsTurn up to fifty production calls into test cases at once.
- TestbedRerun any turn from a real call with a changed input.
- Required checksEvaluations that must pass before staging or production. Overrides are audited.
- Environments and versionsDevelopment, staging, and production, with rollback and side-by-side compare.
- Staged releasesEnterpriseFive, twenty-five, fifty, then all traffic, gated at every step.
04Channels and telephony7
- Phone numbersBuy numbers, port your own, or bring your own Twilio account.
- SIP trunksInbound and outbound, with failover, live SIP traces, and TLS and SRTP only.
- Batch calls and triggersCall a whole list, or start calls from Bland, Salesforce, and HubSpot events.
- Voicemail and phone menusLeave a message and follow up by text, or wait through a phone menu without talking over it.
- Warm transferEnterpriseThe agent briefs your team on a second line, then brings the caller in.
- SMS, RCS, and iMessageEnterpriseThe same agent and memory over text, with fallback between channels.
- Web chat widgetVoice and chat on your site, with handoff to Zendesk, Intercom, or Freshdesk.
05Observe and improve7
- ConversationsEvery call and thread in one place, with listen-in, transcripts, and each decision the agent made.
- IssuesFlagged calls filed by severity and category, from transcription to telephony.
- Norm FixNorm reproduces a bad call, fixes a draft, and checks it against the original.
- DispositionsEnterpriseStanding questions answered for every call, with quotes from the transcript as evidence.
- GuardrailsFour TCPA checks on every plan; custom rules, backtested on past calls.
- AlertsInfo, warning, and critical alerts by email, phone, or your own tool.
- Analytics in plain EnglishAsk Norm a question about your calls and get a number, table, or chart.
06Enterprise5
- Dedicated infrastructureEnterpriseIsolated per customer, with releases pinned to you.
- Zero-downtime releasesEnterpriseRolling deploys that start with 1% of traffic.
- SSO and rolesEnterpriseOIDC or SAML, with five roles mapped from your identity provider.
- Audit logsEnterpriseAn audit log API for your SIEM.
- Signed webhooks and static IPsHMAC-signed webhooks, JWT signing on enterprise, and fixed outbound addresses.
Start with your hardest calls.
Our engineers build the first agent with you, end to end. Kin Insurance was in production in 3 to 4 weeks.