# How does Bland Norm work? It builds and tests AI phone agents without touching live calls

Letting an AI assistant edit a phone agent that is already taking calls raises one question before any other: what stops a bad edit from reaching a caller? [Norm](/norm), Bland's AI assistant, works in development, not in the version your callers are on, and a change reaches live calls only when a saved version is published and promoted.

That separation is why a team that runs calls, and not code, can ask Norm for changes in plain English and still read exactly what changed before a caller hears it. Norm and Bland's visual builder are the two ways your team builds agents without code, and either one lets you run a [no-code AI call center](/no-code-ai-call-center) with no engineers needed on your side.

## How does Bland's Norm work?

Bland's Norm is an AI assistant built into Bland that builds, edits, tests, and debugs your AI phone agents from a plain-English conversation, and agents, tools, analytics, and triage all work from that one conversation. [Bland's Norm guide](https://docs.bland.ai/tutorials/norm) gives example requests for building agents, debugging calls, and querying analytics.

Bland says Norm works the way an experienced Bland engineer would: it "inspects the live state of your account before making changes, works in development so production keeps running, tests against real call data, and only ships changes when you're ready."

Norm can create a new agent from a description or edit an existing one, including its prompts, tools, knowledge, and settings, so the assistant that fixes your agents is also a [natural language AI agent builder](/norm/no-code-agent-builder).

Development is one of an agent's three environments, the named slots that decide what a call runs. Development holds the working copy and is where you edit, while staging and production each hold one saved version and keep serving it until you point them at another.

Outside a coding tool, your team reaches Norm in two places. Norm chat is the chat surface in Bland's dashboard, full page on the Overview and in a side panel everywhere else. Norm is also available over iMessage, so your team can text it to build, test, and monitor agents from a phone.

## How does Norm keep live calls safe?

Norm keeps live calls safe by never editing what they run: a call runs the version pinned to production, and every change moves through three stages first.

### Edit in development

Norm makes changes in an agent's development environment, which is where you edit, and production keeps serving its pinned version while you and Norm work.

### Save a version

Once a change passes validation and any tests you run, Norm saves it as a new version. Saving a version does not change live traffic.

### Publish and promote

To put a saved version on live calls, it is published to staging and then promoted to production. An agent that has never been live skips staging, so its first publish goes straight to production. After that, publishing to staging leaves production alone, and callers keep the version production already holds.

Bland puts the line between saving and shipping this way: "Approval to save a version is not approval to move live traffic. Production is untouched until the version is promoted."

### Evaluations can gate the move

An evaluation is "a judge reading a conversation and saying whether the agent did the thing you wanted," run against simulated callers. Once the Evals feature is enabled for your organization, you can attach judges on staging and mark them required. A judge passes when half or more of the conversations it scored match the verdict you wanted.

In Bland's dashboard, a required evaluation that fails blocks the publish into staging, and the version stays saved on development without a number and without a pin. A blocked promotion is stopped twice: Bland's dashboard dims **Promote**, and Bland's servers enforce the same rule again for an API caller, so production keeps serving the version it already had.

The gate has a documented bypass. **Force promote** sends development straight to production and skips the evaluation gate, and Bland records each bypass in the audit log. Like every publish, promotion, and rollback, it takes an owner, admin, or operator role.

### Every step leaves a record

<!-- screen: an agent's Environments tab showing development, staging, and production with pinned versions and the activity log, captured and dated for this page -->

An agent's Environments page keeps an activity log with a row for each event, such as **Version saved**, **Published to staging**, **Promoted to production**, and **Rolled back production**, with who made it and when. **Compare** opens two versions side by side, and its **Simple** view steps through each field that changed. Bland says you can "review diffs between the original and updated prompts" before you deploy.

If a promotion goes badly, **Rollback** "lists the versions production held before and repins to one you choose. It is a repoint, not a rebuild."

Scope is a safety control too. Norm follows surgical-edit rules, so "it changes what you asked for and nothing else."

### What Enterprise adds

Every enterprise customer runs on its own isolated infrastructure, run by Bland: dedicated compute, containers, and resources, with releases the customer chooses when to adopt and canary deployments on a share of real calls.

A staged release is a separate feature on enterprise plans. It splits production traffic between a new version and the version live now, you choose how much traffic goes to the new version and raise its share step by step, and starting one checks the same required evaluations as a promotion. Isolated infrastructure and staged releases both come with [Bland's enterprise AI voice agents](/enterprise-ai-voice-agents).

## How does Norm test and debug an agent?

Norm tests an agent by running simulated conversations, and it debugs one by tracing a failed call back to a root cause through your call logs and transcripts. Norm has direct access to both, and Bland says "When you point it at a problem, it gathers the evidence before suggesting a fix."

Start from a real call when you can. Paste a call ID and "Norm will pull the full call and reproduce the failure before suggesting a fix." Bland's own examples are "Why did call `abc-123` fail to transfer?" and "This agent is misrouting refund requests to the cancellation flow. Find out why."

Before a change ships, Norm can pit your agent against a simulated caller to stress-test it.

On Bland, a simulated caller is a test case: a title, a prompt describing who the caller is and what they want, and a maximum number of turns, 20 by default. Bland can also read up to 50 of your real calls and write one test case per call. Bland's [walkthrough for testing an agent before launch](/no-code/how-to/test-an-ai-phone-agent-before-launch) follows the clicks.

For one bad answer, Testbed re-runs one turn of a real conversation with your changes, "without redialing anyone." You choose how many times to run it, and Bland groups identical answers and marks each group **Matches original** or **Different from original**. Testbed does not replay the rest of the call, so a fix that works there still gets tried against whole conversations in evaluations.

## How does Norm fix a bad call?

<!-- screen: a Triage issue with Norm Fix and the reviewable change, captured and dated for this page -->

Norm fixes a bad call from Triage, a part of Bland where issues capture problems found on calls and bundle their evidence. Flag a call as a Triage issue and hit **Norm Fix**, and Norm "pulls in the failing call and its agent, reproduces the failure, applies a fix, and verifies it against the original call."

Every change Norm makes is reviewable on the issue, so you can see exactly what it touched before the fix reaches production. A fix is a change like any other, so a fix reaches live calls only through a saved, published, and promoted version.

A comment can start the work too: write @norm in a comment on an issue, and Bland dispatches a Norm investigation. Bland can also draft an evaluation judge from a Triage issue. Bland's [bad-call walkthrough](/no-code/how-to/fix-a-bad-call-with-norm) follows a fix step by step.

## What are Norm skills, and how do developers reach Norm?

Norm skills are reusable workflows that scope Norm to a specific task, either a built-in skill such as `/systematic-debugging` or a custom skill your team writes. Type `/` in Norm chat to open the skill menu.

Custom skills are scoped to your organization and encode work your team repeats: "brand-voice enforcement, common agent scaffolds, or your standard QA scenarios for every new agent."

Developers reach the same agents through Bland's API, Bland's hosted MCP server, and the Norm plugin for Claude Code and Claude Desktop. The plugin pairs the hosted MCP server at `api.bland.ai/v1/mcp` with skills and `/norm:*` commands to build, migrate, validate, and test agents from a coding agent.

The same rule holds there: Norm "asks before saving an agent version," and Bland says "Saving a version does not authorize changing live traffic."

Idaho Housing and Finance Association saves $750,000 a year by replacing its IVR with Bland, which handles 4,000 inbound calls a day and cut average call time by 20%.

> "We save thousands of dollars a day."

Greg Blake, Chief Information Officer, Idaho Housing and Finance Association

## Frequently asked questions

### Does Bland have a no-code option?

Yes. Norm, Bland's AI assistant, builds voice agents using natural language descriptions, and Bland's dashboard includes a visual builder for non-technical users. Either route builds a [no-code AI phone agent](/no-code) that ships through the same development, staging, and production environments.

### Can Norm break a live agent?

Norm edits a Bland agent in development, so a change reaches live calls only when a saved version is published and promoted. Where evaluations are configured on staging, the required ones must pass first once the Evals feature is enabled for your organization, and Force promote skips that gate and is recorded in the audit log.

### Who approves Norm's changes?

On Bland, "Approval to save a version is not approval to move live traffic," so a version Norm saves stays in development until it is published and promoted. Publishing, promoting, and rolling back an agent take an owner, admin, or operator role, and each one is logged on the agent's Environments page with who made it and when.

### Can Norm explain a change before making it?

Yes. Bland calls "Walk me through what you'd change before doing it" a valid prompt for Norm, and Norm follows surgical-edit rules, so it changes what you asked for and nothing else.

### Does Norm work over iMessage?

Yes. Bland's Norm is available over iMessage, so your team can text it to build, test, and monitor agents from a phone.

### What does a forward-deployed engineer add on Enterprise?

A forward-deployed engineer is a Bland engineer who helps a customer safely make changes and launch new agents. Enterprise customers also get Bland's forward-deployed engineers, and Bland says "Norm codifies what those engineers know."
