WHITEPAPER
How Wappari's Conversation-Native Architecture Extends Amazon Connect to WhatsApp, SMS, Chat, and Document Channels
Wappari
Free, no signup required.
Wappari's architecture began on the WhatsApp Business API and grew outward from a single decision: every channel a business uses, voice calls, SMS, meeting recordings, uploaded documents, chat messages, is treated as one continuous conversation rather than a collection of separate systems, each with its own reader and its own disconnected record. That decision is the lesson this paper sets out to apply. Not to replace what a business has already built in Amazon Connect, but to show what the same principle yields when it is pointed at a contact center and folded together with the WhatsApp, SMS, chat, and document channels that business already runs.
Amazon Connect businesses that turn on Contact Lens are not starting from nothing. Contact Lens genuinely does real work on a call: it separates agent and customer onto their own audio channels, produces a diarized, speaker-labeled transcript from that separation, tracks sentiment turn by turn, matches the call against configured categories, and can redact sensitive data automatically once the call ends. None of that is a strawman to knock down. It is a legitimately capable analytics layer, and it solves a harder problem than it gets credit for.
The detail worth noticing is what shape that output already takes. Contact Lens does not hand back a single blended audio track and leave someone to guess who said what. It hands back agent and customer, already separated, already labeled, turn by turn. That is precisely the speaker-attributed shape a broad spectrum memory needs, the same customer, agent, customer, agent pattern a chat thread already has natively and a raw single-track recording does not. Amazon Connect has, without necessarily framing it this way, already done the hardest part of the work.
Two boundaries sit around every Contact Lens transcript, and neither is a defect in Amazon Connect, they are the natural consequence of being one channel among several.
The first boundary is what the transcript is for. Contact Lens was built to help a supervisor understand how a call went: whether the customer sounded frustrated, whether the agent gave the required disclosure, which category the call falls into, where the compliance risk sits. That is a quality-and-coaching function. It tells you how the conversation was conducted. It does not tell you whether the business process the customer called about actually got completed: a refund issued, a claim opened, an appointment booked, a dispute resolved.
The second boundary is what the transcript is connected to. Contact Lens has no way of knowing that the phone number on this call is the same person who messaged the business on WhatsApp last week, or emailed support the week before that. Every call starts the relationship from zero unless a human agent happens to remember, or happens to go looking in a CRM screen before picking up. The call is a complete, well-produced, entirely isolated slice.
Enterprises trying to close that second gap tend to reach for one of two approaches, and both are recognizable from anywhere else broad spectrum memory has been discussed.
A backend agent that queries across systems. It reaches into the contact center's search, the CRM, the email server, and tries to assemble a picture. What comes back is metadata, not conversation: a call summary, a category tag, a subject line, a note someone typed after the fact. The fidelity is low because pulling a full call, a full email thread, and a full chat log out of three different systems and reconciling them is expensive, and most integrations settle for the cheap version of that work.
A specialist for each shape. Contact Lens is, on its own terms, exactly this kind of specialist, a genuinely sophisticated reader of the voice channel. But sophistication inside one channel does not reach the others. It is excellent inside Amazon Connect and invisible to WhatsApp, invisible to email, invisible to whatever channel the same customer used yesterday. Making the voice specialist better does not fix the fact that it is still a specialist working alone.
Either path leaves the same gap in place: a business that talks to its customers on several channels, with a genuinely good reader on one of them, and no continuous record of the relationship anywhere.
Closing that gap does not mean replacing Contact Lens or rebuilding what Amazon Connect already does well. It means treating its output as one more channel to fold into the same shape every other channel already produces, and resolving it to the same customer record every other channel already updates.
Same shape. Because Contact Lens has already separated and labeled agent and customer turns, folding a call in does not require re-diarizing anything. The work is translation, not extraction, taking a transcript that is already speaker-attributed and landing it in the same structure a WhatsApp thread, an SMS conversation, a web chat, an uploaded document, or an email already lands in. The sentiment trend and category tags travel with it as context, not as the end product.
Same customer. The call is linked to the same customer identity as every other channel that customer has used with the business, using the one signal durable enough to survive the boundary between a phone call and a WhatsApp message: the phone number itself. Done carelessly, phone-number matching is exactly how two unrelated people end up merged into one record. Done properly, a cross-channel match is treated as a candidate to confirm, not a fact to assume, the same discipline that keeps an identity system trustworthy applies here without exception.
The result is a call that no longer sits alone. It sits inside the same continuous file as everything else that customer has ever said to the business, on any channel.
A folded-in call is not just easier to read, it becomes eligible for a different question entirely. Contact Lens can tell a supervisor that a call was about a billing dispute, that the customer's sentiment dropped in minute six, and that the agent's tone stayed compliant throughout. What it does not attempt to answer is whether the dispute actually got resolved by the end of the call.
That is a completion question, not an analytics question, and it depends on reading the whole conversation as one continuous thing rather than scoring it in isolation. Once an Amazon Connect call is folded into a broad spectrum record, it becomes just another conversation eligible for exactly that kind of pass, the same one already applied to every other channel, and what comes out the other side is not a category label but a structured, deterministic account of what happened as a result of the call. The category says what kind of conversation it was. The completion record says what it produced.
Every approach to giving an agent more than one channel's worth of context runs into the same tradeoff. Twilio's Conversation Memory and Zoho's Zia Conversation Summary reach across channels genuinely, pulling observations, traits, deadlines, and friction points from every place a customer has shown up. But what they hand back is a distillation, a note, a summary, a sentiment score, not the conversation itself. They are broad, but shallow. LangChain's memory types and Oracle's hybrid layered memory pattern take the opposite trade: real depth, a full and carefully retained history, but built for one channel at a time. They are rich, but narrow.
Contact Lens's transcript already sits on the rich side of that line: full text, not a summary, already speaker-attributed. The gap this paper has described so far is that it stays narrow, locked to the voice channel alone. Folding it into a broad spectrum record does not trade that richness away to gain breadth, the way reaching for a Twilio or Zoho style layer would. It keeps the full conversation and adds every other channel at the same fidelity, because every channel is reduced to the same cheap-to-read shape rather than compressed into a note about itself.
That combination, rich and broad at once, is what makes a different kind of agent possible: not one that reads on demand when a human opens a thread, but one that runs continuously across the entire record looking for signals no single channel, and no summary of a channel, would ever surface. A sentiment dip on one call is a data point a supervisor might catch live. The same dip recurring across three calls, from a customer who also stopped responding on WhatsApp and never finished uploading a document, is a signal that only exists once all three channels sit in one place at full fidelity, not summarized down to a note in a CRM field. Twilio's and Zoho's agents cannot find that signal because they never had the actual conversation to look at, only what got extracted from it. LangChain's and Oracle's methods cannot find it because they were never watching more than one channel at a time. A watcher built on a broad spectrum record can, because it is no longer choosing between depth and reach the way every method before it had to.
None of this asks a business that has already invested in Amazon Connect to give it up. It asks what else that investment could feed. A contact center that produces a genuinely good, already speaker-attributed transcript is closer to broad spectrum memory than a business realizes, it is missing one connection, not a rebuild.
For a business built on Amazon Connect, broad spectrum memory does not mean replacing the contact center. It means making sure the contact center is no longer the one channel standing apart from every other place that business talks to its customers. One conversation, every channel, the phone included.