Bland AI

Executive Summary

Bland is a phone-agent platform for connecting conversations to business workflows. Its current first-agent guide covers a new dashboard rolled out September 10, 2026: create an agent, attach a phone number, then add tools and knowledge as needed.^[1] It belongs in the agentic voice APIs comparison as an operational platform, beyond a standalone speech-synthesis endpoint.

Bland says it operates its own language and speech models and offers managed-cloud, customer-VPC, on-premises and air-gapped deployment options.^[2] That can simplify supplier coordination, but it does not establish universal correctness, latency or security. Evaluate the actual contracted deployment and workflow.

Building and Connecting an Agent

The new dashboard's starter flow uses a restaurant-reservation demo or a blank agent. Attaching a number determines which agent answers; tools and knowledge are optional additions. The guide directs builders to call records and configuration when a call connects but the agent does not speak.^[1] That separates a routing problem from an agent-behavior problem.

Bland also documents Conversational Pathways: nodes contain instructions or fixed speech, and edges express routing conditions. Draft edits do not affect live calls; production uses a published version, with staging and version-specific calls available for testing.^[3] Existing pathway integrations and the new agent dashboard should not be assumed to have identical editor or API shapes.

The Tools guide distinguishes its current visual builder and managed connections from legacy Custom Tools.^[4] The MCP server supports agent creation, versioning, branches, staging and production promotion; its API key scopes access to an organization and mutating tools require confirmation arguments.^[5] Webhook signatures use HMAC-SHA256 and a separately configured signing secret.^[6] These mechanisms do not replace authorization inside the business API being called.

Testing and Production Control

Scenario tests simulate a caller and evaluate assertions such as extracted values, visited nodes, triggered webhooks and transcript matches. Some assertions use model judges. Scenarios can persist across pathway versions, and required scenarios can gate promotion to production.^[7] A passing test means the configured assertions passed—not that every possible caller interaction is correct.

Guard Rails provide required-event and prohibited-content checks with actions such as ending a call, transferring it or moving to another node. The guide also recommends backtesting and warns that ambiguous custom rules can produce false positives.^[8] Plan access matters: the pricing table places protected-call guardrails and other advanced monitoring in Enterprise.^[9]

For enterprise runtime upgrades, Bland documents dedicated infrastructure, named version-pinned releases and separate canary containers. Teams can route selected numbers or a percentage of calls to a candidate release before promotion.^[10] This is distinct from editing a conversation's prompt: both the flow version and runtime version can change behavior and should be recorded during evaluation.

Human Handoff Has Several Outcomes

Enterprise warm transfer starts a second, proxy call that briefs a person before joining the caller. Configuration covers hold behavior, IVR navigation, timeout and voicemail handling. The final states distinguish MERGED, CANCELLED, NO_ANSWER and TIMED_OUT; individual proxy attempts have their own state records.^[11]

For operations, “transfer attempted” and “customer reached a human” are different outcomes. Use the final result when measuring containment or resolution. A caller stuck in a queue should not be counted as successfully handed off merely because the destination answered with an automated menu.

A worked reservation test

Start with an agent that checks availability before booking. Ask for four seats, correct the date mid-call, and make the first availability request return slowly. Verify the saved reservation against the corrected request and check that a retry does not create a duplicate.

Then test a full restaurant, a failed booking API and an unavailable human destination. Assertions should distinguish an actual booking, an honest rejection, a callback request and a completed handoff. Repeat the scenarios before promotion and sample real call records after rollout. This is a proposed acceptance test, not a claim that this review placed calls or measured Bland's success rate.

Pricing and Limits

Published standard rates checked September 16, 2026:

The AI-minute price includes LLM, STT and TTS. The pricing FAQ explicitly separates carrier charges, whether paid directly or passed through Bland. Bring-your-own-Twilio customers do not pay Bland transfer fees.^[9] This is more specific than broad all-inclusive wording elsewhere in the product material.

The billing guide additionally lists Scale at $499 monthly, $0.11 per connected minute and $0.03 per transfer minute, although the main pricing page does not display that tier. Confirm its availability rather than assuming every signup sees it.^[12][9]

Actual call time is prorated to seconds. The billing guide lists a $0.015 minimum outbound attempt charge using Bland telephony, including failed calls; voicemail consumes ordinary call time. Warm transfer can add a separately active proxy agent before calls merge. Norm, the build assistant, is billed separately by tokens.^[12] A complete estimate needs successful calls, failed attempts, handoff behavior and carrier charges.

As simple arithmetic, 1,000 AI connected minutes would be $140 on Start or $419 including Build's subscription, before other charges. Build's higher fee also purchases capacity and features; a lower minute rate alone is not sufficient reason to upgrade.

The separate Agent Phone plan

Agent Phone is $29.99 monthly, with a first-month $2.99 promotion advertised through September 17, 2026 at 11:59 p.m. Pacific. It provides one US number and covered calling/texting to US and Canadian destinations without per-minute or per-message billing.^[13]

Its “unlimited” wording has hard limits: one concurrent call account-wide, twenty calls per hour, fifty calls and 1,000 combined inbound/outbound minutes per day, plus twenty outbound texts per hour and one hundred per day. Outbound calls are capped at sixty minutes. When relevant limits are reached, new calls can be rejected or receive a busy response.^[13] This is a single-agent subscription, not interchangeable pricing for a concurrent contact center.

Deployment and Data

Bland's trust page describes evidence available for security review and distinguishes independent reports from self-attested HIPAA/GDPR statements backed by agreements. It advertises configurable regional deployment and zero-retention processing for sensitive workflows.^[2] This review did not inspect private audit reports or an enterprise contract.

The published data agreement permits deriving deidentified data for product improvement and other business purposes. It also distinguishes customer personal data from service data, and sets conditions on end-of-service deletion and backups.^[14] Those terms deserve a precise account-level review; “own models” does not by itself mean no retention or no secondary processing.

External Evidence and Fit

OpenBenchmarks Labs published an August 1, 2026 snapshot of phone-call latency using saved carrier recordings and a specified analyzer. Its Bland row reports 429 usable turns out of 432, with 1,520 ms median and 2,248 ms 95th-percentile time from caller speech ending to agent audio beginning.^[15]

These are that harness's results, not measurements performed for this report. The study includes carrier-path overhead, uses different platform defaults and some overrides, and discards turns that fail its timing-quality checks. It does not score answer quality, interruption handling or task success. Its 38 ms Bland–Vapi median difference is inside the stated error budget, so it does not support a meaningful ranking between those rows.^[15]

Bland merits evaluation when a team wants phone operations, structured workflows, testing and enterprise rollout controls in one service. Compare Vapi and Retell AI with the same calls and full bills. Teams primarily seeking interchangeable speech components or unrestricted model selection should verify that preference against the architecture before choosing a platform. The useful decision is which system completes, explains and recovers the intended work—not which demo sounds most convincing.

---

Research by Ry Walker Research • methodology