AI Chat and WhatsApp Agents | Rel8 | Resolution, Not Deflection

The problem with most chat

Deflection is not resolution

Most contact centre chat exists to keep people away from the phone. It answers from a knowledge base, cannot act on an account, and hands over the moment anything real is asked. Customers learn this within two exchanges and ask for a human immediately.

A bot that reads

Matches your question to an article and pastes it back. Cannot check a balance, take a payment or change anything. Success is measured by tickets avoided, which is why it feels like being fobbed off.

An agent that acts

Authenticates the customer, reads the account, makes the change, takes the payment, and writes it to the system of record. Success is measured on whether the thing the customer wanted actually happened.

Same brain as voice and email

One agent, one set of tools, one set of compliance rules across all three channels. A customer who starts on chat and phones an hour later does not get a different answer, which is the failure mode of buying a separate chatbot.

Where chat is different

Text is not voice with a keyboard

The channel has its own constraints, and designing as though it were a transcript of a call is why so many chat deployments feel wrong.

Sessions are asynchronous. A customer replies in four seconds or forty minutes. The agent has to hold context across a gap, and know when a conversation is genuinely abandoned rather than paused.
Customers paste and attach. Screenshots, statements, reference numbers in the wrong format. The agent needs to extract from what is given rather than insist on a form.
Nothing is deniable. Every word is written down and screenshottable, so a wrong disclosure in text is a harder compliance problem than the same slip on a call. Text channels raise the bar rather than lower it.
Handover has to preserve the thread. When a human takes over, they get the conversation, the account context and what the agent already did, in the same window the customer is already using. No new chat, no repeating.
Vulnerability signals read differently in text. Hesitation and distress do not sound like anything, they look like short replies, repetition and time gaps. Detection has to be tuned for the channel rather than reused from voice.

The channel your customers already use

Delivered through the WhatsApp Business Platform on Amazon Connect, so it lands in the same routing, reporting and audit trail as every other contact rather than in a side system nobody can see.

What it is good for

Higher open and response rates than email, a persistent thread the customer can scroll back through, and a channel people already have on their phone. For arrears, service updates and appointment flows it consistently outperforms outbound calling.

The rules that catch people out

Outside a 24 hour window from the customer's last message you can only send pre approved template messages, opt in has to be captured and evidenced, and your sender identity needs approval before launch. Any plan that assumes free form outbound at will is a plan that fails review.

Under the hood

What sits inside a chat agent we ship

Identity verification before any account action, with verified status carried into a handover so nobody re-authenticates.
Payments handled inside your existing PCI compliant flow, so card data never enters the transcript or the model context. Security detail.
Mandatory disclosures evidenced per conversation, so adherence is a measured number rather than a hope.
Vulnerability detection tuned for text, with supervisor alerting on high confidence signals.
Immutable per turn audit trail carrying the message, the tool calls and the prompt and model versions. How we measure.
Escalation into your existing desktop, on Genesys, Avaya, Amazon Connect or whatever you run. The overlay route.
Containment and resolution measured against a baseline captured before go live, with repeat contacts subtracted.

The same standard on every channel

What does not change when the channel does

Voice, email and chat are one agent with one set of tools and one set of rules. The channel changes what the agent can do. It does not change what it is held to.

FCA Consumer Duty, on every channel. A customer disclosing hardship in an email carries the same duty as one disclosing it on a call. Vulnerability signals are detected, evidenced and escalated in all three, tuned to how they actually present in that channel rather than copied across from voice.
UK GDPR, and data that stays where you put it. Deployments run in your AWS account and your region by default. Personal data does not leave the UK or EEA without a written instruction. Security and certification status.
Designed against the EU AI Act's high-risk obligations. Human oversight, logging, transparency to the customer and technical documentation are built in rather than retrofitted. We will tell you plainly which obligations apply to your use case and which do not.
Every decision replayable. An immutable per turn audit trail carrying the transcript or message, the tool calls, the guardrail decisions and the prompt and model versions, retained to your policy. If a regulator asks why the agent did something in March, you can show them.
Payments inside your existing compliant flow. Card data is captured by your current payment provider and never enters the transcript, the model context or our logs.
Measured the same way. Containment, resolution and cost per contact carry the same definitions and the same exclusions on every channel, against a baseline captured before go live. The definitions in full.

Chat that resolves, not a bot that deflects