What happens between a customer hitting send, and an answer landing back.
No flow builder, no intent tree. Splice AI reads your documents, writes the answer, scores how well it can actually back that answer up, and only sends when the score clears the bar you set.
- One knowledge base behind all five channels
- A confidence score decides send or hold, not a vibe
- Anything it cannot ground waits for a person
The pipeline
Six steps, every message, every channel
This is the whole path. It does not change between WhatsApp, Instagram, email, phone and your website.
- 01
The message arrives
A WhatsApp message, an Instagram DM, an email, a phone call or a website chat all land in the same pipeline. Nothing is channel-specific from here on, which is why an answer you fix once is fixed everywhere.
Carrier and platform webhooks are signature-checked before anything is read.
- 02
A burst becomes one question
People send three short messages, not one tidy paragraph. The agent waits a beat, reads the burst together, and answers once instead of firing back a reply per fragment.
Chat channels collapse a burst into a single queued reply. Instagram waits slightly longer than WhatsApp, because DMs arrive faster.
- 03
It searches your knowledge
Vector similarity and full-text search run against your documents at the same time and their results are fused, so a customer asking in their own words still reaches the right paragraph.
If the embedding service is slow, it falls back to text search rather than failing the reply. Marketing material is tagged at ingest so testimonials are not quoted as fact.
- 04
It writes, then scores itself
The reply is drafted with a confidence score covering how well the intent was understood and how well the answer is actually supported by what was retrieved.
The score is then capped by the product. No grounding caps it at 0.45, and a weak retrieval match caps it at 0.55, whatever the model claims.
- 05
It sends, or it waits for you
Above your threshold it goes out immediately. Below it, the reply lands in your inbox as a draft with the reason attached, ready to approve or edit.
Refunds, legal, data-deletion, an angry customer, or anyone asking for a person force a draft first, before the confidence score is even consulted.
- 06
It follows up, then knows to stop
If the customer goes quiet the agent nudges on a schedule you set, in hours for chat and days for email. It stops the moment they reply.
Follow-ups also cancel when the ticket is solved and when a teammate takes over, so nobody gets nudged after they have already been helped.
Your controls
Where you decide, not the model
Autonomy is a dial you set, and the product enforces limits the model cannot argue its way past.
You choose the bar, per channel
Run WhatsApp on autopilot and keep email fully manual, or the reverse. Auto-send is a switch per channel, not one global setting.
Any conversation can go manual
Switch a single thread to human-only without changing anything else, and the agent stops replying on that thread alone.
Taking over really stops it
One takeover pauses replies, scheduled follow-ups and background resolution checks together, rather than just hiding the bot in the interface.
The bar rises for at-risk customers
Leads marked lost need a higher confidence before anything auto-sends, and active clients are excluded from auto-send entirely.
Every draft explains itself
A held reply says why it was held, whether that was low confidence, a sensitive topic, or auto-send being off for that channel.
Gaps come back as suggestions
Questions it could not answer are grouped by how often they come up, drafted into suggested answers, and only enter your knowledge base once you approve them.
After the reply
The conversation does not stop at the answer
Answering is the visible part. What happens to the lead afterwards is what makes it worth having.
The lead is updated
Contact details, budget, timeline and requirements are extracted from the conversation and written to the record.
It routes to the right person
Rules on keyword, stage, channel or lead score assign the conversation, including round-robin across a team.
The next step gets scheduled
Meetings booked from the conversation, follow-ups queued, and tasks created against the lead rather than the thread.
Questions people ask about how it works
Does it need training, or a flow builder?
Neither. You upload documents or crawl your website, and it answers from that material directly. There is no intent tree to maintain and no retraining step. Add a document and it knows the contents on the very next message.
How long does it take to answer?
Fast enough that customers do not notice a queue, but we deliberately do not publish a number, because we have not measured one we would stand behind. Chat channels also wait a couple of seconds on purpose, to collapse a burst of messages into one good reply rather than three rushed ones.
What if the answer is not in my documents?
Its confidence is capped in code below any threshold you are permitted to set, so it physically cannot send itself. It becomes a draft, and the agent's own wording says it will check and come back rather than inventing a status or a timeline.
Can I see why it made a decision?
Yes. Each reply shows its confidence and the reason it was sent or held, and every conversation keeps its full history in the inbox alongside the extracted intent, sentiment and stage.
Do all the channels really share one brain?
Yes. One agent, one knowledge base, one inbox, one set of confidence rules. A customer who asks on Instagram and then emails gets consistent answers, and their history sits on a single lead.
What happens when a human steps in?
The agent stops on that conversation completely, including its queued follow-ups and background checks. Handing back is one click, and the customer never sees a bot message layered on top of a human one.
Which languages does it handle?
It detects the language and script of each message and replies in the same one, covering English plus twelve Indian languages, in native script or romanised. It mirrors how the customer writes rather than switching them into Hinglish uninvited.
How is it priced?
Credits are deducted as the agent works, so a quiet month bills quiet. There are no per-seat fees, which means putting your whole team in the inbox costs nothing extra.
Want the setup detail instead? Read the technical walkthrough, or see it per channel.
Talk to us
See it answer your own questions.
Same business day, usually within a few hours. We are on India hours.
Mon to Sat