AI chatbots & assistants
The bar for a customer-facing assistant is not "can it hold a conversation" — it is whether it resolves the question, knows when it cannot, and hands over cleanly when it should. Most deployments fail on the last two.
Why people distrust chatbots
A generation of decision-tree bots taught customers that the fastest route to help is typing "agent" until something happens. That expectation is now the default, and any new assistant inherits it.
Language models fixed the fluency problem and introduced a subtler one. A bot that answers wrongly but articulately does more damage than a menu that answers nothing — particularly if it invents a refund policy, quotes a price that does not exist, or states something you are regulated on.
A deployment worth doing therefore starts from constraint, not capability: grounded strictly in your approved content, explicit about uncertainty, and quick to escalate with full context rather than making the customer repeat themselves.
What we build
Grounded answering
Responses drawn from your documentation, policies and product data rather than the model's general knowledge, so the assistant states your policy rather than a plausible industry-standard one.
Clean escalation
Handover to a human with the full conversation, the customer's identity and the assistant's best understanding of the problem, into your existing helpdesk. No repeating the story.
Account-aware assistance
For logged-in users, access to their own order, booking or account data through scoped API calls — which is the difference between a FAQ search box and something genuinely useful.
Tone and boundaries
Brand voice defined and tested, plus explicit refusal behaviour on topics you do not want it discussing — competitor comparisons, legal or medical advice, anything with regulatory exposure.
Internal assistants
The same architecture pointed inward: staff assistants over policy, process and product knowledge, respecting existing document permissions.
Measurement
Resolution rate, escalation rate, satisfaction and the questions it fails on — the last being the most useful, because it tells you what content to write next.
Typical stack
Chosen per project, not by habit. If your team already runs something that works, we use it.
- Channels
- website widget
- Slack
- Microsoft Teams
- in-app
- Helpdesk integration
- Zendesk
- Intercom
- Freshdesk
- HubSpot
- ServiceNow
- custom APIs
- Grounding
- hybrid retrieval over your content
- structured product and pricing data
- account APIs
- Safety
- topic boundaries
- output validation
- PII redaction in logs
- full conversation audit
How a deployment runs
We launch narrow. An assistant that handles your ten most common questions excellently earns the right to handle more; one that attempts everything on day one teaches customers not to use it.
- 01
Content and intent review
Your actual support volume analysed to find what people ask most, and whether the content to answer it exists. Gaps in documentation surface immediately.
- 02
Shadow mode
The assistant drafts answers your agents see but customers do not, so accuracy is proven against real conversations before it faces anyone.
- 03
Narrow launch
Live on a defined set of topics with fast escalation everywhere else, monitored closely for the first weeks.
- 04
Widen on evidence
Scope extended where resolution rates justify it, using failed conversations as the roadmap.
Common questions
Will it replace our support team?
Not in our experience, and we would be wary of anyone promising it. It absorbs the repetitive volume, which lets the same team handle growth without proportional hiring and spend their attention on the conversations that need a person.
What stops it from inventing policies?
Strict grounding in your approved content, output validation, and explicit refusal when the content does not cover the question. Tested against a set of adversarial questions before launch, including the ones your customers will actually try.
Can it handle multiple languages?
Yes, and it is one of the strongest arguments for this approach — the same assistant can serve customers in languages you do not staff for. We recommend restricting to languages you can escalate in, so a handover is not a dead end.
How does it integrate with our helpdesk?
Through the existing API. The assistant becomes a first tier that resolves or enriches, creating tickets with full context when it escalates, so your agents see one queue rather than a separate channel.
How do we know it is working?
Resolution rate, escalation rate and customer satisfaction, compared against your pre-launch baseline. We instrument this from the start, including the questions it handles badly, which is the number most vendors do not show you.
Start with a scoping call.
Thirty minutes, no obligation. If we are not the right fit we will tell you on the call rather than after a proposal.
Related services
LLM & RAG integration
Getting a model to answer questions about your own content is easy to prototype and hard to trust. The differe…
AI agent development
Most AI pilots stall because a demo that answers questions is not a system that does work. We build agents tha…
AI workflow automation
The processes that resisted automation for twenty years were the ones needing judgement: reading a document, d…