Skip to content

One agent. Every channel. Every stage.

Heard listens.
Then it thinks.
Then it does the work.

One customer agent across voice, WhatsApp, web and email — not four bots with four knowledge bases. It understands, retrieves, answers and validates before it replies, then executes the work in your systems. Arabic-native. Every answer cited.

Abdullah Saadoun

Saudi · Male

Abdullah Saadoun, Saudi male voice

of conversations resolved without a human
70%+
voice response latency
<1.5s
coverage, no overflow queue
24/7
from signup to a live agent
<1 day
Built by an insurer
Heard comes out of Yasmina, on live insurer traffic — not adapted from a horizontal bot.
Arabic and English
Gulf and Levantine dialects handled natively on voice, not translated at the edge.
Auditable by design
Every answer cites its source document; every action is logged on the conversation timeline.

The problem

Answering is the easy part

Most of this category is a search box with better manners. It tells the customer what the policy says, what the refund rules are, where the parcel might be — then hands them to a person to actually do something about it. That handoff is where the cost always was.

  • The customer waits days for something that could have been done during the call.
  • Teams drown in repeat contacts asking where things got to.
  • Every process change needs an engineering ticket and a release.
  • Nobody can prove what the AI said, or what it was reading at the time.

Four bots is the other failure. A voice bot, a WhatsApp bot, a web widget and a helpdesk macro — four knowledge bases, three different answers to the same question, and no single place where a customer’s history lives.

Heard is one agent. One brain, one set of logic, one memory of the customer, reachable on every channel they already use.

How it works

Sense and Mind

Two halves, and the product is only honest when both have run. Sense is how the customer reaches us. Mind is everything that happens before a single word goes back.

Sense

Listen

One agent, reachable everywhere the customer already is.

Voice, WhatsApp, web, email and SMS arrive at the same agent — not five bots with five knowledge bases that answer the same question three different ways. Arabic is native: Gulf, Levantine and Egyptian dialects on voice, not English translated at the edge.

  • One agent across every channel
  • Arabic and English, in dialect
  • Sub-1.5s to first word on voice
  • Full conversation history follows the customer

Mind

Think

Four stages, and an answer that failed one of them never ships.

Most products in this category call a model and return whatever comes back. Heard runs a pipeline, and the last stage is the one that matters: the answer is checked against what was actually retrieved before the customer ever sees it. When it cannot be supported, the agent says so and escalates rather than inventing something plausible.

  • Intent understood before anything is retrieved
  • Grounded in your documents and live systems
  • Generated with the citation attached
  • Validated against the source, or escalated

Heard and Understood

The customer got an answer that was true, and the work behind it was done before they hung up.

Architecture

Four stages. The fourth is the one nobody else runs.

Calling a model and returning what comes back is a demo. Heard runs the answer through a pipeline, and an answer that fails the last stage never reaches the customer.

  1. 01

    Understand

    Work out what the customer actually wants, in their dialect, including the half-sentences and the corrections mid-call.

    Intent, entities and sentiment — before a single document is fetched.

  2. 02

    Retrieve

    Pull the specific clause, policy record or order status the answer depends on, from your documents and your live systems.

    Retrieval is scoped to the tenant and the customer. Nothing else is visible.

  3. 03

    Generate

    Compose the reply in the customer’s language and register, with the source it came from attached to it.

    The citation is produced with the answer, not reconstructed afterwards.

  4. 04

    Validate

    Check the answer against what was actually retrieved. If it is not supported, it does not go out — the agent says so and escalates.

    This is the stage most of the category skips. It is the reason the others hallucinate.

AI answeredValidated

“Your windscreen excess is JOD 25, and repairs through an approved fitter don’t affect your no-claims discount. A full replacement carries the standard JOD 150 excess.”

Motor Policy Wording v4.2 §7.3POL-1190-A schedule

Every answer carries the document it came from.

Not a footnote added afterwards — the citation is produced with the answer and checked against it. When the knowledge base cannot support a reply, the agent says so and escalates instead of inventing something plausible.

How grounding works

Action

A true answer is half the job

The value is in the work behind the answer, so Heard does it — inside your systems, during the conversation, with the ops lead owning the logic rather than an engineering backlog.

  • Inspect the exact data going into and out of every step.
  • Branch on anything the agent has gathered.
  • Test against a real payload before it touches a customer.
  • A failed branch never takes the others down with it.
See the flow editor
Inbound Call+962 6 500 1180
Succeeded
Claims AgentCollect FNOL details
Running
IFTotal loss suspected
Queuedtruefalse
Open ClaimPOST /claims
Human HandoffClaims desk

Industries

One agent. Four industries that cannot afford a wrong answer.

Not a horizontal chatbot with industry templates. These four share the shape that matters — high-consequence, transaction-heavy, and Arabic-speaking — which is exactly where citations, an audit trail and real actions are worth more than a cheaper bot.

FNOL at 2am, claim status, renewals and endorsements.

A collision generates a phone call, not a support ticket. Heard takes the first notice of loss properly, opens the claim and books the assessor before your team opens in the morning.

What the agent does

  • Open a claim
  • Book an assessor
  • Amend a policy
  • Confirm a renewal

Scope

Support is where this starts, not where it stops

The agent already holds the order, the policy and the whole conversation history. That is the thing best placed to onboard the customer, answer the objection and chase the renewal — not a second product bolted on beside it.

Support

Resolve the inbound conversation end to end, including the work behind it.

Onboarding

Walk a new customer through setup, documents and first use, in their language.

Sales

Qualify, answer the objection with real product facts, and close the simple cases.

Retention

Chase the renewal, recover the cart, and catch the churn signal in the transcript.

Compounding

It gets better because it is corrected, not because it is magic

Every product in this category claims to learn. The credible version names what is fed back and who signs it off.

Every conversation is scored

Resolved or not, and if not, exactly where it broke — retrieval, generation or validation.

Gaps become suggestions

A question the knowledge base could not answer arrives as a proposed article, drafted from what the human said next.

A human approves

Nothing edits your knowledge base on its own. Suggestions are reviewed, then applied — which is the difference between improving and drifting.

Measurable

The numbers your CFO will ask about

Every conversation is scored on whether it actually resolved — because that is what Heard charges for.

Conversations handled

1,481

12.6%vs. last 30 days

Automation rate

?

72%

8.4%vs. last 30 days
Target70% by September

Avg. handle time

?

2m 14s

31.2%vs. last 30 days

Cost per resolution

$0.42

18.9%vs. last 30 days
How outcome pricing works

Why us

Built by an operator, in the language the customer actually speaks

Heard comes out of Yasmina — real transactions, real users, real regulatory review. The global platforms are English-first and treat Arabic as a locale. In this region that is not a feature gap; it is the difference between an agent people use and one they hang up on.

Arabic-native

Gulf, Levantine and Egyptian on voice, in dialect — not English translated at the edge.

Executes, not answers

The work behind the reply happens in your systems, during the conversation.

Auditable by design

Every answer cited, every action logged, on the same timeline as the conversation.

Outcome-aligned

Paid when a conversation resolves. Not per seat, not per minute.

Questions

Frequently asked

What is Heard?

Heard is an AI operating layer for insurance operations. It runs voice and chat agents that answer customer conversations from your own documents and then execute the work behind them — opening claims, booking assessors, amending policies — inside your core systems.

How is this different from a chatbot?

A chatbot answers questions. Heard takes action. The agent is connected to your policy admin, claims and booking systems, so a conversation ends with the work done rather than with a promise to follow up.

Why insurance specifically?

Because the hard parts are industry-specific. Knowing what a total loss looks like at first notice, which clause governs an exception, when a question needs clinical judgement — none of that comes free from a general-purpose platform. Heard is built by an insurance company, on real insurer traffic.

How long does it take to go live?

The target is under a day: upload your knowledge base, build a flow, connect a number, go live. No engineering required for the core setup.

What happens when the AI cannot handle something?

It hands off to a human with the full transcript, the documents it read and every action it took already attached. The customer never repeats themselves, and your agent never starts from zero.

Is our data used to train models?

No. Your documents and conversations are used to serve your agents and nothing else.

See Heard on your own calls

A 30-minute call. Bring a recording of a real customer conversation and we will show you what the agent does with it.