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.
Layla Haddad
AI handlingVoice callPOL-4471-B
Omar Nasser
AI handlingWhatsAppCLM-8823
The flow behind Layla’s call. Scroll it — an ops lead edits this, not an engineer.
- 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.
A real conversation
Not ticket #39218.
Rana, whose windscreen just cracked.
Heard doesn’t triage her into a queue. It listens, understands who she is and what she needs, does the work, and resolves it — then learns from it. Watch it think.
Rana
Comprehensive motor · POL-1190-A
My windscreen cracked on the Irbid road. I’m insured with you — I just don’t know what to do.
What Heard understood
- Identity
- Rana H. · policy POL-1190-A · matched from number
- Intent
- Windscreen damage → first notice of loss
- Sentiment
- Stressed, mid-commute — wants certainty, not options
- Context
- Comprehensive · windscreen covered · JOD 25 approved-fitter excess
- Action
- Open FNOL · book approved fitter · no NCD impact
Booked before she had to explain twice
Claim opened, an approved fitter booked for Saturday, the excess and no-claims impact answered in the same breath — in one WhatsApp thread, in Arabic, in under two minutes.
1m 48s
Listen to resolved
- 01
Listen
Hears the customer, in their language
- 02
Understand
Identity, intent, context, feeling
- 03
Act
Does the work in your systems
- 04
Resolve
The customer gets an answer
- 05
Learn
Gets better, because it is corrected
Windscreen FNOLs cluster on the Irbid road after winter. Heard flags the pattern and surfaces the approved-fitter booking one step earlier next time.
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.
- 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.
- 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.
- 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.
- 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.
“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.”
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 worksAction
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.
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
Automation rate
?72%
Avg. handle time
?2m 14s
Cost per resolution
$0.42
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.