Enterprise AI call center operations
AI answers the call. Your team gets the ones that matter.
Zumu is the AI call center platform built for operators: agents that answer every inbound call in seconds, brief your people before any handoff, and keep every minute recorded, measured, and honestly billed.
- Warm handoff with an AI briefing
- Listen, whisper, barge in
- English and Spanish
- One config for voice, chat, email
- Greeting starts, cached
- ~30ms
- Prompt-cache hit ratio
- 0.98
- Interruption-detection precision
- 86%
- Minutes a week, one operator
- 49,000
Measured in production for a single flagship operator. Your results depend on call mix.
The recording bar does not stop when the human joins. Same room, same session.
Somebody still answers the 2 AM call. It just is not a person anymore.
The cost of the queue
What the queue costs you
Nobody budgets for hold time. It gets paid anyway, in bookings that never happen and callers who tell the story twice.
Today
The bill you are already paying
- Hold times bleed callers before anyone says hello
- After-hours voicemail is a booking that never happens
- Every transfer restarts the story from the top
- QA reviews two percent of calls and guesses at the rest
With Zumu
What changes on the first day
- Answered in seconds, every hour of the day
- Handoffs arrive briefed, so the caller repeats nothing
- Recording and supervision survive the transfer
- Every call scored, not a sample
The queue-hold experience belongs to you, not to chance.
How it's built
Plug in anything. Deploy everywhere.
Bring your own model. Pick your own voice. Connect your tools and your knowledge. One agent brain reasons through the case, then answers on whatever channel the caller used.
Inputs
Any AI model
OpenAI, Anthropic, Google, and more. Your own keys.
Any voice
Pick the voice per agent. Swap it like a setting.
Your tools
MCP, your APIs, your own integrations.
Your knowledge
Your documents. The knowledge pipeline builds the rest.
One agent brain
Outputs
Phone calls
Inbound and outbound.
Website widget
Chat and voice, right on your site.
Email
Drafted, ready for your team to send.
One definition. One brain. Every channel at once.
In-room escalation
The handoff is the product
Most voice AI treats a transfer as an exit. It dials a second call, drops the recording, and hands your person a stranger. Zumu treats it as a participant joining a room that is already open.
- 01
The caller goes to hold music
Nothing about the call ends. The room stays open, the recording keeps running, and the caller hears music instead of a dial tone.
- 02
The AI writes the briefing
Four to six sentences: who is calling, what they asked for, what has already been tried, and what the agent thinks should happen next.
- 03
Your person accepts or declines
They hear the summary before they hear the caller. Declining is a real option, and it falls back to voicemail or a retry instead of dumping the caller.
- 04
They are bridged into the same room
Not a new call. The same session, the same recording, the same supervisor still able to listen. The transfer is a participant change, not a restart.
- Warm transfer
- Cold transfer
- Fallback to voicemail or retry
Frustration escalation is deterministic: more than thirty cues can hand a call to a person, and the model still gets a say in whether it should.
Live operations
Your supervisors keep their hands on the floor
The AI answering the calls does not make the room opaque. Every live call is one click from a supervisor, in three degrees of involvement.
Listen in
Hidden subscriberA supervisor joins the room as a hidden subscriber. The caller hears nothing change, because nothing did.
Whisper
Agent-only audioCoach the person on the call without the caller hearing a word of it. Private audio, same room.
Barge in
Consent-gatedTake the call over when it needs a supervisor. Gated behind an explicit consent setting, because some states require it.
Caller sentiment, last 60 seconds
Whisper is open. The caller hears the agent, not the coach.
Built for two-party-consent states.
One agent, three channels
Configure the agent once. It shows up everywhere.
The same prompt, the same tools, the same knowledge graph. Changing how the agent handles a refund does not mean changing it in three places and hoping they agree.
- PhoneThe protagonist
Inbound and outbound, your numbers or ours, with the full supervision surface behind it.
- Caller: I need to move tomorrow to Thursday.
- Agent: Let me pull that trip up for you.
- ChatOn your own site
One script tag, isolated from your page styles, with voice available in the browser.
- Visitor: Do you cover the north county?
- Agent: We do. Which address should I check?
- EmailDrafted, not sent blind
The same agent, the same knowledge, writing a reply your team can send or edit.
- Subject: Re: Thursday pickup window
- Draft ready for review
Voice is the protagonist. Chat and email are companions, running the same config.
Knowledge
Point it at your documents. It argues with itself before it answers your callers.
Ingestion runs an eight-stage adversarial pipeline: one pass drafts, another attacks the draft, and only what survives lands in the graph the agent queries mid-call.
What you point it at
- Documents
- Call audio
- Your website
Onboarding crawls your site and builds a first knowledge base before you finish signing up.
The adversarial pipeline
- 01Extract
- 02Normalize
- 03Chunk
- 04Link entities
- 05Draft answers
- 06Challenge
- 07Reconcile
- 08Publish
Retrieval is hybrid and reciprocal-rank fused, so a caller question hits the store that actually knows the answer instead of whichever one scored first.
Where it lands
- RelationalFacts and records
- GraphHow things relate
- VectorWhat sounds similar
Three stores, one query path. The agent does not know or care which one answered.
Built for operators
The bill has to survive a finance review
Every call carries a full session report: cost by component, which provider served it, how long it waited in queue, and the trace of every tool the agent used. Nothing is rolled into a line item you cannot open.
The human leg is on the recording and in the QA review. It is not on the invoice.
On transferred calls you pay only for the minutes the AI handled. For one operator that excluded about a third of the weekly bill.
- Five metered components, each on its own line
- Rate cards per organization, effective-dated so a price change never rewrites last month
- Footer totals to the cent, matching what the ledger above adds up to
Running in production
49,000
Minutes a week, thousands of calls, for a single operator. One number, one customer, measured rather than modelled.
- 11
- Named latency optimizations between a caller’s last word and the reply
- 2
- Languages live today, English and Spanish, on a stack that reaches further
These are the figures we can stand behind for a named production deployment. We are not going to publish a percentage we cannot show you the query for.
The floor at 6 PM, when the queue never got away from you.
Developers
Everything the console does, your code can do
The console is a client of the same API you get. There is no private surface we kept for ourselves.
- REST
POST /v1/callsPlatform API
An OpenAPI surface over agents, calls, numbers, knowledge, and campaigns. Everything the console does, you can do.
Read the docs - MCP
zumu.calls.searchMCP server
Run your call center from Claude Code or Cursor. OAuth in both directions, so your agent can call ours and ours can call yours.
Read the docs - Events
call.transferredWebhooks
Signed with rotation overlap, retried, dead-lettered, and replayable. Flag rules are graded by a model, not a regex.
Read the docs
Questions
What operators ask first
Still have questions?
Bring a call you wish had gone better. We will run it past the agent with you watching.
Book a demo
See it answer your calls
Bring a recording of a call you wish had gone better. We will point the agent at your own documents and you can hear what it does with the next one.
- No credit card for the demo
- Your own knowledge base in the first session