Everyone knows what bad customer service feels like: the hold music, repeating your story five times, a script that never solves anything. This is a support assistant built to do the opposite. It listens, answers from the company's real help articles, and hands you to a human the moment one is needed, with the whole conversation attached so you never start over. Compassion and competence, tuned to any business.
It is the same call from two ends. The customer wants one thing fixed and to feel heard. The business wants to help everyone without going broke or burning out its people. The system in the middle usually lets both down.
Hold music and "your call is important to us." Telling the whole story, then telling it again to the next person. A script that never quite answers the question. Being transferred, disconnected, and told to call back. And the worst one: hanging up with the problem still not solved.
Live help is expensive and almost impossible to staff around the clock. Good agents burn out repeating the same fifty answers. Quality swings between a great rep and a tired one. The knowledge to fix things exists, but it is scattered, and routine questions bury the cases that genuinely need a human.
Think about the last time customer service actually went well. It almost certainly felt like this.
That is the whole bar: compassion, competence, and the problem actually solved. This assistant is built to clear it on the everyday questions, and to know the exact moment a person should take the rest.
Take each thing people dread about contacting support, and here is what this assistant does instead.
A customer whose gear failed right before a trip. Watch the tone, the grounded answer, and the clean hand to a human the moment the stakes rise.
No hold music. No repeating the story. A grounded answer, real warmth, and a person brought in at the right moment with the whole thread in hand.
When someone is frustrated, or something genuinely went wrong, it leads with a real apology and slows down instead of firing back a canned line. Specific, honest empathy, never a hollow "I understand your concern."
Short, direct answers in everyday words, framed around what can be done right now. It sets expectations honestly and tells people exactly what happens next.
The tone is set per company, from a warm outdoor shop to a calm clinic to a precise software team. Same engine underneath, your personality on top.
It says it's a virtual assistant up front, and treats handing you to a person as a feature, not a failure. Being straight about what it is earns more trust than faking it.
The most common way AI support fails is the handoff itself: the customer gets bounced to a person and has to start the whole story over. That moment of lost context is the biggest single reason people say they hate support bots, so it is the part this is built around.
What was asked, which article answered it, what the assistant already tried, and the customer's own words all travel to the agent. The person picks up where the bot left off, and nobody repeats themselves.
A frustrated tone, the same question asked twice, or a high-stakes topic trips a handoff on its own. The customer never has to fight a menu to reach a person.
The assistant clears the repetitive volume so a person spends their hours on the cases that genuinely need one, arriving with the easy parts already done.
# the company powering the transcript above company: display_name: "Northwind Outdoors" tagline: "Gear for people who'd rather be outside" branding: colors: { primary: "#1f6f54", accent: "#f4a23b" } tone: "warm, outdoorsy, practical" support: in_scope: [orders, shipping, returns, gear care, warranty] out_of_scope: [wholesale pricing, payment-card details] escalation: triggers: ["speak to a human", "lost package", "refund over $300"] contact: { phone: "1-800-555-0142", hours: "Mon-Fri 8-6 PT" } knowledge_base: path: "knowledge" # 6 articles, chunked and embedded llm: tier: "fast" # a local Ollama model, no external API
The profile compiles into the assistant's system prompt and its runtime behavior. Swap this file and the knowledge folder, and the same code speaks for a different company.
Retrieval over the company's documents, a model allowed to answer only from what it pulled, and a policy layer that decides when to hand off. That core holds whichever engine powers it, so a business is never locked to one vendor.
Most support SaaS runs on a hosted model: Claude, GPT, or Gemini, reached by API or kept inside the company's own cloud through Bedrock, Azure, or Vertex. Best answer quality, fastest to launch. A cheap model handles routine questions and only the hard ones escalate to a larger one.
The same app runs on self-hosted open models such as Llama, Qwen, or Mistral, inside the company's own walls. No conversation leaves the building and there is no per-token bill. It is the right call for healthcare, finance, or anyone who cannot send customer data to a third party.
Returns, shipping, a warranty claim, where an order is, or tell it you're frustrated and want a person. It resolves what it can from Northwind's documents, shows its sources, and gets a human the moment it should.