The agentic CX loop

Most CX AI answers.

Clarity learns.

Clarity listens, reasons, acts, verifies and learns. Every conversation makes the next one better.

Today, in the loop
Tickets
0
Resolved
0
Handed over
0
Scored
0
Knowledge fixes
0
IDConversationChannelScoreOutcome
Every one of these was scored on the same rubric, and every failure has a reason code.

The problem

An open loop forgets.

Signal goes in. An answer comes out. Nothing returns. The same gap reappears next quarter, and nobody knows it was there.

Open loop

#48120 · refund window

Customer asks

09:14

AI answers

Sent

Conversation closed

09:16

Nothing checked it

no score

Nothing returned to the source

no fix

Same question, next quarter

Same question, next quarter

Wrong again

Closed loop

#48120 · refund window

Customer asks

09:14

AI answers

Sent

Scored against the rubric

Unverified

Knowledge base updated

+1 article

Answer now grounded in an approved source

Verified

Same question, next quarter

Same question, next quarter

Answered from source

What it costs

An open loop leaks at every stage.

Not one failure. Five, compounding. These are figures customers gave us about their own operations.

Listen
0%of negative public posts never answered.

A national service operator, measured across its own channels.

Reason
0weeks to compile insight by hand.

And fifteen reports rebuilt every day to keep up in the meantime.

Act
0%of agent time on repetitive queries.

80% of inbound calls still open a ticket.

Verify
0%of conversations reviewed.

Enterprise organisations name a single digit figure. Evaluating eight calls per agent per month.

Learn
0%of failures were never the agent.

They trace to knowledge gaps. 10% were scored against the wrong article entirely.

And when it fails, it fails in public.

−47

NPS, voice channel

A telco virtual advisor answered live customers for months before anyone measured it. Then it was switched off.

20

Calls a day, peak

Two vendors, six months. The pilot never handled more than twenty calls a day, and failed the moment real customers used it.

26%

Of a scored sample

Answers that were wrong or had no source to check them against. Found only once someone scored the conversations, not before.

$600k

Per month

Recruitment and onboarding spend to replace agents who leave, in a team where the same gaps resurface every quarter.

How it works

Five stages. One data layer.

Follow one customer conversation all the way round. Each agent does its part on the same data layer, and what the loop learns is written back, which is why the next conversation is better grounded than the last.

Listen
Voice of Customer
Reason
Voice of Customer
Act
AI Support Agent
Verify
Agent QA
Learn
Brain Guard
01 · knowledgeUpdated
Brain Guard
Idle
Refund policy v4.2 · damaged goods
Order #48213 · delivered 12 Jul
Refund window 14 days · 3 remaining
New · damaged-goods refunds under €100 auto-approve
knowledge base3 articles
Refund policy, damaged goodsv4.2
Courier claim SLAv2.0
Refund ceilings, auto-approvenew
rule written
Damaged-goods refunds under €100 auto-approve, so no agent re-checks the ceiling again
ticket
ID#48213
StatusOpen
ChannelChat
CategoryRefunds
CustomerJosh Brennan
Handled byAI Support Agent
AI Support Agentlive
Hey Josh, how can we help today?
My order #48213 turned up damaged. Can I get a refund?
Reading order history and refund policy…
Damaged-goods clause applies · within the 14 day window
Sorry about that, Josh. Order #48213 qualifies under our damaged-goods policy, so I have issued a full refund of €82.40, back with you in 3 to 5 business days. No need to return the item.
Perfect, thank you!
Resolved · first contact
Message…
02 · verifyAgent QA
Waiting for a handled conversation
Policy accuracyv4.2
Refund amount€82.40
Tone and empathypass
Disclosure includedpass
Resolution confirmedpass
94/ 100 · passed
03 · outcomeResolution
Handle time
Escalated
CSAT
04 · learnWhat the loop found
Answer was correct, but 40s went on re-checking the refund ceiling
The same clause was looked up three times this week by human agents
Writing back to Brain Guard…
verified findings return to the knowledge base

One conversation, all the way round: Voice of Customer ingests it, Brain Guard grounds it, AI Support Agent answers, Agent QA scores it against your rubric, and what the loop found is written back. Synthetic conversation, real verdict taxonomy.

The loop runs wherever your data has to live: in-region cloud, sovereign private cloud, or fully on-premise. Same topology, same stages.

The loop runs wherever your data has to live: in-region cloud, sovereign private cloud, or fully on-premise. Same topology, same stages.

In production

The loop, running today.

Five stages, live in banking, telecom, marketplaces and delivery.

55%

less time spent on voice of customer analysis.

“Clarity helps us replace guesswork with real data from our feedback. I can go into product meetings with clear evidence.”

Jen Barwick

Head of Customer Experience

Chosen by OpenAI to surface insights from customer conversations securely.

700 million

weekly active customers

One AI layer across a national contact centre operation.

126 million

global subscribers

“Clarity treats all of our unstructured customer feedback like true data. Their NLP models pick up on sentiment and sort it into categories. We can search everything instantly, see the numbers, and even build our own customizable feedback dashboards.”

Alexandra Motto

Director of Product and Design Operations

20 million

monthly customers

Turning rider and captain feedback into smarter everyday experiences.

70 million

weekly active customers

Listen · Voice of Customer

Every channel. One stream.

Calls, chat, WhatsApp, social, app reviews, tickets and surveys land on the same data layer, in the same shape, in real time.

Channels and systems of record
Voice and IVR
recordings, transcripts
0
Chat and email
web, in-app
0
WhatsApp
business API
0
Social and reviews
public sources
0
Tickets and CRM
accounts, entitlements
0
Surveys
CSAT, NPS
0
One data layer
0
conversations ingested today
Same shapePII removed at ingestionNo lag
Two or three days becomes the same afternoon, because nothing waits for a report to be written.

Signal · Voice of Customer

The complaint outside and the ticket inside.

Public signal joined to contact centre signal on one data layer. A spike that means nothing alone means something next to the other one.

Public signal · reviews, social
Contact centre · tickets, calls
Anomaly detected11:42:06
Building 4, water. Tickets up 340% against a 7 day baseline.
Correlated with 11 public mentions in the same window. Routed to facilities and to the contact centre lead.

Act - AI SUPPORT AGENT and AGENT ASSIST

It resolves what it can. It hands over what it cannot.

AI agents work across chat and voice. Agent Assist sits beside your people with a drafted answer and the source it came from. Every response carries a confidence score, and the score decides who finishes the conversation.

Resolves end to end.

Chat and voice, in your customer’s dialect, inside your existing contact centre.

Copilots your agents.

A drafted reply, the source it came from, and a confidence score, in the agent’s own console.

Knows when to stop.

Below your confidence threshold the conversation goes to an agent, with everything so far attached.

Queue
Your threshold
70
#48312Drafting
Where is my refund?
confidence0
#48317Drafting
Change my delivery address
confidence0
#48324Drafting
Why was I charged twice?
confidence0
#48331Drafting
Cancel my subscription
confidence0
Handed to an agentbelow threshold
Nothing waiting

Verify · Brain Guard and Agent QA

Three ways an answer can be wrong.

Only one of them is the agent’s fault. Unverified means no source existed to check the answer against. That gap belongs to the knowledge base, not the person.

Reviewed by hand
0
Scored by Clarity
0
VerifiedGrounded in an approved source, and it passed.
UnverifiedNo content existed to check it against.
FailedWrong information reached the customer.
Over a third of failures trace to knowledge base gaps, not agent error.You scored the agent. The agent did not have the answer. Nobody scored the thing that failed her.

Verdict taxonomy is the real one: verified, unverified, incomplete, failed. Distribution illustrative.

100%

of conversations scored

3 to 8%

is what quality teams sample today, on their own numbers

One rubric

yours, applied to human and AI alike

Reason codes

on every failure, so the fix has an owner

Learn · Brain Guard

It finds the gaps, then closes them.

Brain Guard reads every verified conversation. Where the knowledge base has no answer, it identifies the gap and writes the article. Where two articles contradict each other, it finds the conflict and fixes them. Most platforms cannot tell you the gap existed.

Gaps in the knowledge base3 articles
Roaming packages
asked 41 times
No article
Plan expiry rules
asked 18 times
No article
Dispute timelines
asked 12 times
No article
Duplicate top-up charges
asked 9 times
No article
Articles that contradict each other
Refund policysays 14 days
contradicts
Returns FAQsays 30 days
Conflict
Roaming chargessays included
contradicts
Tariff sheetsays billed
Conflict
fixed
One source of truth per question, and the agents that quoted the wrong one are told which article replaced it.

Deployment

The loop doesn’t stop at your border.

In-region cloud, sovereign private cloud, or fully on-premise with GPU inference in your own datacenter. Model and cloud agnostic: bring your own models, or run open weights on your hardware. Same topology in every mode.

Reference architecture

Channels

voice · chat · social · tickets

Your boundary

Data layer

Inference

Verify

Knowledge

Deployment

in-region · sovereign · on-premise

How to start

Thirty days. Your data. Your rubric. A number you didn’t have before.

Step 1 · scope

A working session, not a slide deck. We agree the use cases, the success criteria and the rubric. Yours, not ours.

Step 2 · one file transfer

No integration. Historical call transcripts, ticket exports and your knowledge base, moved once.

Step 3 · the readout

Every conversation scored. Verified, unverified and failed, with reason codes. Plus the quality baseline that sets your opening bar.

Fixed price, fixed scope, credited in full against a first-year contract.

The thirty days start when your data arrives, not when you sign.

For security and data governance

You’re being asked to release call recordings. Here’s everything you need to say yes or no.

We know this decision sits with you, and we know it is the step where these projects actually stall. So rather than ask your colleague to relay answers, here is the pack: what we need, what we do with it, where it is processed, what gets stripped before storage, and what happens to it afterwards.

What data we need to start, and what we don’t

PDF

PII removal at ingestion, technical note

PDF

Processing locations and data residency options

PDF

DPA template and sub-processor list

PDF

Close the loop on every conversation.

And somewhere at the end of the loop, a customer got a straight answer the first time they asked, and an agent had the right one to give.

Clarity. Secure customer service that builds the product.