AI & Data
Chatbots judged by
what they resolve
Deflection is the wrong metric — it counts the customers who gave up. We build assistants measured on resolved contacts, escalation quality and repeat contact rate, with a handover to a human that carries the whole conversation rather than starting again.
Overview
Support assistants measured on resolved conversations, not on deflection statistics.
A support assistant lives or dies on three things: whether it can find the correct answer in your knowledge base, whether it can act on the customer's account, and whether it knows when to stop and hand over. Most deployments get the first right and neglect the other two. That is why containment looks good for a month and complaint volume rises in the second.
We build for customer-facing and internal use alike: order status and returns, policy and HR questions, IT service desk, branch and clinic enquiries. Channels usually mean web, WhatsApp, in-app and voice, and the same intent has to behave consistently across all of them, including when the caller interrupts mid-sentence.
Escalation is designed first, not last. The assistant hands over with a summary, the full transcript, what it already verified and what it could not, so the agent does not restart the conversation. Agents who trust the handover stop routing around the assistant, and that is what actually moves the resolution rate.
- 43.6%
- Contacts fully resolved without an agent at Tinsmith Retail
- 2m 12s
- Average agent handling time saved on escalated conversations
- 6.1%
- Repeat contact rate within 72 hours, against a 14% baseline
Capabilities
What this covers
Six areas we staff properly. If your problem sits outside them, the honest note at the foot of this page says so.
Knowledge grounding and answer sourcing
Answers drawn from your published policies and help centre with the source shown, plus a weekly report of questions the assistant could not ground. Stale content becomes visible rather than quietly wrong.
Account actions and authentication
Once the customer is verified, the assistant can check an order, change a delivery slot or raise a ticket, within scoped permissions and value limits. Verification uses your existing identity provider, not a new one.
Escalation and agent handover
Handover carries a summary, the transcript, verified identity and the steps already attempted, delivered into Zendesk, Salesforce Service Cloud or ServiceNow as a structured case rather than a pasted log.
Multichannel and multilingual delivery
One intent model across web, WhatsApp, in-app and voice, with per-channel response formatting. Language support is tested per market by native reviewers, because machine-translated policy text goes wrong expensively.
Conversation analytics
Weekly review of unresolved topics, drop-off points and repeat contacts within 72 hours. This is the feedback loop that improves the assistant; without it, quality decays from the day it launches.
Safety, tone and compliance
Regulated wording held in reviewed templates rather than generated, refusal paths for advice the business must not give, and transcript retention to your policy so complaint handling has the full record.
Deliverables
What you get
- Grounded answer service with source citation and coverage reporting
- Authenticated account actions with scoped permissions and value limits
- Structured escalation into your CRM or service desk with full context
- Channel adapters for web, WhatsApp, in-app messaging and voice
- Conversation analytics with an unresolved-topic review workflow
- Content gap report and a maintenance routine your support team can run
Stack
What we build it with
- Rasa
- Anthropic Claude API
- Twilio
- WhatsApp Business Platform
- Zendesk Sunshine Conversations
- Salesforce Service Cloud
- Amazon Transcribe
- pgvector
- Node.js
- Next.js
- Redis
- Grafana
Process
How the engagement runs
Two-week increments against a written definition of done. You can stop at any increment boundary and keep everything built so far.
Contact analysis
We read three to six months of real transcripts and tickets, then rank intents by volume and handling time. The top twelve usually cover more than half of contacts.
Answer sourcing audit
We test whether your current content can actually answer those intents. It frequently cannot, and fixing twenty articles beats any amount of prompt tuning.
Build and internal pilot
The assistant runs with your support agents as the only users first. They break it faster and more usefully than a customer cohort will.
Limited live release
Five to ten percent of live traffic, with a visible route to a human on every turn and daily review of the conversations that ended badly.
Scale and maintain
Traffic increased as resolution rate holds, with a standing weekly session where your team retires, edits and adds intents without us.
When this is the wrong engagement
If your help content is out of date or contradicts itself, an assistant will surface that inconsistency to thousands of customers at once, and the content work has to come first.
FAQ
Questions we get asked
- Will customers be told they are talking to a bot?
Yes, on the first message, and every turn offers a route to a person. Concealing it damages trust and is being legislated against in several jurisdictions. In our experience disclosure barely affects resolution rate; hiding the exit route damages it badly.
- What happens when the assistant does not know?
It says so, offers the closest sourced article, and hands over with everything it has gathered. It does not guess. Unanswered questions are logged as content gaps and reviewed weekly, which is how a knowledge base improves rather than ossifies.
- How much of our support volume can realistically be handled?
For retail and utilities with clean content and account access, 35 to 55 percent of contacts resolved without an agent is a realistic first-year range. Regulated advice, complaints and anything emotional should stay with people, and we scope those out deliberately.
- Can it work in more than one language?
Yes, though not for free. Retrieval works across languages, but policy wording, tone and regulated phrases need a native reviewer per market and a separate test set. Budget roughly two to three weeks per additional language, mostly for review rather than engineering.
Related