Our riders include people with speech and hearing differences. What happens then?

The agent is tuned not to talk over a caller. Interruption detection measures at 86% precision in production, and turn timing is configurable per agent. When a conversation is not working, the frustration detector counts more than 30 explicit cues underneath the model and forces a handoff to your dispatcher rather than letting the call go in circles. We would rather escalate early than be clever.

Does this come with a compliance certification for our access-program contract?

No, and we will not imply one. We publish the controls we actually operate: per-organization data isolation enforced at the row level, encryption at rest, audit logs, consent gating on supervisor barge-in for two-party-consent states, and a privacy-shield stage that strips personal information before anything reaches a knowledge base. The current list is on our security page, and where your contract needs something that is not on it, we will tell you plainly instead of hedging.

What does the record of a call look like afterwards?

One session report per call: the diarized transcript with speakers attributed, the recording, every tool the agent called and what it returned, the outcome, the queue timing and the cost. Transferred calls also carry a reason code, a resolution grade and a coaching note about the human leg. You can filter and export the lot, so answering the program office is a query rather than an excavation.

Paratransit and access programs

Peak hour arrives and the dispatch queue arrives with it. Nobody has to hold.

Zumu is paratransit dispatch AI that answers riders and drivers on the same phone system without either one waiting behind the other. It confirms a pickup, says where the vehicle is, and when a call needs your dispatcher it brings her into the conversation already in progress with the whole story told first.

Riders and drivers answered at once
Every call recorded end to end
Escalates to a dispatcher, not to voicemail
English and Spanish today

Book a demo

A transfer bridges your dispatcher into the call the rider is already on, and the recording continues straight through the handoff.: One room
Explicit phrases counted by a rule-based detector sitting underneath the model, so a rider who asks for a person gets one.: 30+ cues
Every transfer is graded afterwards, including the leg your own dispatcher handled.: Reason coded

These are properties of how the system is built, not outcomes we are promising you. The measured figures are further down the page, with their caveats attached.

A pickup that has to happen on time

The pressure

Where a paratransit day actually breaks

Three moments from a normal Tuesday on an access-program dispatch line. Each one is a queue problem that started somewhere else.

07:15Tuesday
Peak hour fills the queue faster than anyone can empty it.
Between seven and nine, every rider with a dialysis appointment, a shift start or a school run calls at once, and so does every driver who needs a reroute. The queue is not long because your people are slow. It is long because demand on an access program has a spike in it and headcount is flat all day.
08:40Tuesday
A driver is stuck behind a rider who is stuck on hold.
The driver at the curb needs thirty seconds to report a no-show so the run can move. He is somewhere down a queue of riders asking where their vehicle is. Every minute he waits, the vehicle waits, and the next pickups on that run slide later, which puts more riders back on the phone.
16:30Tuesday
The program office wants the call record and you have a spreadsheet.
On-time performance, complaint handling, how long callers waited and what they were told. Answering means stitching together a phone report, a dispatch log and somebody’s memory of the conversation. What you cannot produce quickly starts to look like what you cannot defend.

The queue is a scheduling problem wearing a phone system’s clothes. It gets shorter the moment the calls that never needed a dispatcher stop reaching one.

How Zumu answers it

Four capabilities, pointed at a peak-hour dispatch line

None of this is a roadmap. Each block is something the platform does today, and each one links through to the page that owns the mechanism.

In-room escalation
Your dispatcher joins the call rather than inheriting it cold
When a rider needs a person, the transfer bridges your dispatcher into the same call instead of handing the caller back to the carrier and dropping out. The recording keeps running, a supervisor can still listen to the human leg, and the rider is not asked to explain the trip again to a second voice.
Warm transfer with hold music while the agent writes a briefing of 4 to 6 sentences on caller intent, facts collected, pending actions and what was already verified. Your dispatcher accepts, declines or reports voicemail, with cold transfer, voicemail, retry and hang up as the fallbacks.
See how a warm transfer runs
Call flow and extensions
Drivers get their own way in, so they stop queueing behind riders
A driver reporting a no-show and a rider asking where the vehicle is are not the same call and should not share a line. Split them at the front door: a direct number or extension for drivers, the rider path for everyone else, each landing on an agent configured for that conversation.
A visual call-flow builder with trigger, receptionist, agent, transfer, voicemail and end-call nodes, keypad and intent edges, five inbound routing paths, and real SIP extensions for your own staff with roles and voicemail.
Open the call-flow builder
Live supervision
Watch peak hour happen instead of reading about it at five
Every active call on one board, with sentiment moving as the call does. Your supervisor listens in without the caller knowing, sends the agent an instruction the rider never hears, and takes the call herself when that is the right thing to do.
A live wallboard over the active-call table with sentiment rebroadcast every 10 seconds, hidden listen-in that also works on post-transfer rooms, whisper limited to one message every 5 seconds, and barge-in behind an organization-level consent acknowledgement.
Look at the supervision console
Knowledge graph
Your program’s own rules, not a generic answer
Eligibility conditions, no-show policy, how far ahead a trip can be booked, what a rider is told when a vehicle is running late. Those rules already live in documents your team wrote. Point Zumu at them and the agent answers from them mid-call.
An eight-stage adversarial pipeline including a privacy-shield stage and a multi-model debate stage, written across a relational store, a graph store and a vector store, retrieved with hybrid dense and keyword ranking and a graph-aware path tried first.
See how the graph gets built

Your dispatchers still make every decision about a rider’s day. They just stop spending peak hour on the calls a system could have finished.

What we can stand behind

In production for operators in medical transportation and paratransit dispatch, handling thousands of real inbound calls a week.

One of those deployments is a paratransit broker serving a major-metro access program. We publish no call volume for that account, because we have not measured one we could defend, and we will not borrow somebody else’s number to fill the gap.

In the room: A transfer bridges your dispatcher into the same call, and recording and live listen both continue through the handoff.
30+ cues: Counted by a deterministic frustration detector underneath the model, so a rider asking for a human is not argued with.
86%: Interruption-detection precision, measured in production. A rider who starts speaking is not spoken over.
English and Spanish: Live today, on a multilingual speech stack that supports 10+ languages.

There is no on-time-performance figure here, no complaint-reduction percentage and no certification badge. A program’s numbers belong to the program, and we have not been given permission to publish anybody’s. What is above is either a property of the product or a figure from a real invoice.

Questions

What access-program operators ask first

Still have questions?

Bring a peak-hour recording with riders and drivers in the same queue. We will run it with you listening to both legs.

Book a demo See our security controls

Yes. An agent is configured per line, so the number a command-center agent dials can reach an agent with different rules, different tools and a different escalation path than the one riders reach. That separation is configuration rather than a code change, and you can adjust it between calls.

Book a demo

Show us your worst peak hour

Pick a morning where the queue got away from you. We will run the same mix of rider and driver calls through a Zumu agent while you listen, and you can judge the handoff on one question: would your dispatcher have wanted that briefing?

Book a demo Talk to sales

No credit card for the demo
Rider and driver calls, side by side
Your dispatchers keep every rider decision

Peak hour arrives and the dispatch queue arrives with it. Nobody has to hold.

Where a paratransit day actually breaks

Peak hour fills the queue faster than anyone can empty it.

A driver is stuck behind a rider who is stuck on hold.

The program office wants the call record and you have a spreadsheet.

Four capabilities, pointed at a peak-hour dispatch line

Your dispatcher joins the call rather than inheriting it cold

Drivers get their own way in, so they stop queueing behind riders

Watch peak hour happen instead of reading about it at five

Your program’s own rules, not a generic answer

What access-program operators ask first

Still have questions?

Show us your worst peak hour