Chatbots & assistants
Your customers have met a chatbot before. Statistically, it wasted their time. You're not starting from neutral — you're starting from a grudge.
Working against a reputation
For a decade, "chat with us" meant a decision tree that couldn't understand a sentence and existed to make you give up before reaching a person. People learned. Now the first thing many users type is a request for a human.
That's the bar. Not "can it hold a conversation" — models cleared that some time ago — but "will this specific user, who has been burned before, get a real answer faster than they would have any other way."
Which changes what you optimise for. Deflection rate — the share of people who don't reach a human — is the metric that produced the bad chatbots, because you can maximise it by being hard to escape. We measure resolution and satisfaction instead.
What we build
- Customer support assistants
- Grounded in your help centre, policies, and product documentation, answering with citations and handing over cleanly when confidence drops.
- Internal helpdesk assistants
- For HR, IT, and operations questions. Often a better first project than customer-facing: real value, forgiving audience, no reputational exposure while you learn.
- Sales and onboarding assistants
- Qualifying, answering product questions, and guiding people through setup.
- In-product assistants
- Sitting inside your software, aware of what the user is looking at and what they're entitled to do.
The rules we build to
- Escalation is one message away, always
- Never buried, never gated behind three refusals. Counterintuitively this improves satisfaction *and* resolution, because users who trust they can escape are willing to try the bot first.
- It says when it doesn't know
- A grounded assistant that admits uncertainty and hands over is worth more than a confident one that invents a returns policy.
- It answers from your content
- Retrieval over your real material, with links to source. Not the model's general impression of how businesses like yours usually work.
- Context comes with the handover
- When it escalates, the human gets the full conversation and what was already tried. Making a customer repeat themselves is the specific failure that makes people hate chatbots.
- We measure what matters
- Resolution rate, satisfaction after the conversation, and escalation quality. Not deflection.
Before you ask.
Grounding and citations make it substantially less likely, and confidence thresholds catch much of the rest. It's not zero, which is why escalation paths and human review on sensitive topics exist.
AI Agents
A chatbot answers a question. An agent does something about it. That difference is the entire engineering problem.
Intelligent SystemsKnowledge Search (RAG)
Your organisation already knows the answer. It's in a PDF, in a folder, that someone left in 2022.
Intelligent SystemsWorkflow Automation
Your systems are fine. It's the space between them where the week disappears.
Tell us what you're trying to build.
A 30-minute call, no charge and no pitch deck. Describe the problem and we'll tell you how we'd approach it, roughly what it costs, and whether we're the right team for it. If we're not, we'll say so.
30 minutes · No charge · No deck