AI is moving from demos into daily work. Bonzai’s AI solutions for business add assistants grounded in your own documents and data where they save real hours, not where they only look clever.
A generic chatbot can write a polite paragraph. It cannot tell a customer what your warranty actually covers, which service you offer in their city, or when to stop and get a person. That gap is the difference between a demo and a tool your team will trust.
Bonzai builds AI that knows your business: assistants trained on your own data, sales and support answers that stay on brand, automatic summaries and drafts, and a hand-off to a person when the question leaves the map. We start with the simplest design that will do the job, and we only add moving parts when they are needed.
That is also Anthropic’s published advice for building with language models: find the simplest solution possible, and only increase complexity when needed (Building effective agents).
What the assistant is for. 1 Your sources; 2 Answers and drafts; 3 The hand-off.
SourcesPolicies, pages, product notes — the files you approve
DraftsSummaries and replies a person can send
Hand-offA person takes over when the assistant should not
What you get
On-brand answers, drafts, and a way out to a human
This is for teams who want an assistant on the website, in support, or beside a sales inbox — and who care that the answers sound like them. It is not a promise that the assistant will never be wrong. It is a way to put your approved material to work, with a clear path to a person.
Assistants trained on your own documents and data
Sales and support answers, on brand
Automatic summaries and drafts
For teams that want an assistant grounded in their own pages, policies and documents — on the website or inside an app — with a clear hand-off to a person when it cannot answer.
Support and sales teams who answer the same questions every day
Companies that have already seen a generic chatbot invent a policy
Operators who want summaries and drafts, reviewed before they go out
Sources, a small system, then evals. 1 Collect the sources; 2 The simplest system that works; 3 Evals from real failures; 4 The hand-off.
Simple firstAdd complexity only when the simple path is not enough
Hand-offDesigned in from the start, not added after a bad answer
How an assistant is built
Start simple, ground it in your material, then test the failures
“We see teams delay building evals because they think they need hundreds of tasks. In reality, 20-50 simple tasks drawn from real failures is a great start.”
Mikaela Grace, Jeremy Hadfield, Rodrigo Olivares and Jiri De Jonghe, Anthropic, Demystifying evals for AI agents (9 January 2026). Demystifying evals for AI agents, Anthropic
We start with the material you already trust: policies, service pages, product notes, FAQs, the phrases your team already uses. If it should not be said, it should not be in the source set.
Then we build the smallest system that can answer from that material and admit when it cannot. Many useful assistants are not a swarm of agents. They are a well-bounded retrieval step, a clear prompt, and a path to a person.
We do not wait for a perfect test suite. Anthropic’s 2026 guidance is that teams often delay evals because they think they need hundreds of tasks; 20 to 50 simple tasks drawn from real failures is a great start (Demystifying evals for AI agents). We write those tasks down and run them as the assistant changes.
Typical timelines for AI That Knows Your Business
Project
Typical timeline
What shapes it
AI assistant
2 to 4 weeks
Grounded in your own documents and data, tested on real questions.
Honest about its range. 1 Stated capabilities; 2 Prompt suggestions; 3 It knows the page you are on.
HonestThe assistant says what it can and cannot do
In contextIt should know the page or record in front of the person
A personThe way out is visible before the user is stuck
What good looks like
Clear about what it can do, and quick to involve a person
“Helpful site-specific AI chatbots clearly state their capabilities, offer relevant prompt suggestions, and quickly signal they know what users are looking at.”
Nielsen Norman Group’s April 2026 guidelines are a fair picture of a helpful site assistant: it clearly states its capabilities, offers relevant prompt suggestions, and quickly signals that it knows what the user is looking at.
A good assistant is also easy to leave. If the question is about a price we have not published, a medical or legal judgement, or a complaint that needs empathy and authority, the hand-off should be obvious — a person, a booking, a phone number — not a longer reply.
Drafts and summaries should be labelled as drafts. A person still sends the email that matters. That is not a limitation of the product; it is how you keep the voice and the facts yours.
20–50
A solid starting set of evaluation tasks, drawn from real failures, before you wait for hundreds of tests.
Confidence is not the same as being grounded. 1 No sources; 2 No evals; 3 No way out.
GroundedAnswers should come from material you approved
EvalsStart with 20–50 tasks from real failures
EscapeA person must be reachable
Pitfalls
The assistants that sound confident and still should not ship
An assistant with no approved source set. It will fill the gap with fluent guesses.
Waiting for hundreds of eval tasks before testing anything. A short list drawn from real failures is enough to start, and far better than none.
Hiding the hand-off. If the only way out is to type “I want a human” after three wrong turns, people will leave instead.
Letting drafts go out as if a person wrote them. Label them, and keep a person on the send.
Building a multi-agent system because it sounds advanced. Complexity should be earned by a job the simple path cannot do.
The same interface, two different jobs. 1 A generic chatbot; 2 An assistant on your material.
YoursAnswers are grounded in your approved sources
On brandThe voice matches the rest of the site
A personThe hand-off is part of the design
Compared
A generic chatbot versus an assistant that knows your work
We will not claim the assistant is always right. We will claim it has a bounded job, a source of truth, a test set, and a way to get a person involved.
A generic chatbot, compared with an assistant that knows your business
A generic chatbot
How Bonzai builds it
What it knows
The open web, or nothing specific to you
Your pages, documents and FAQs
When it is unsure
It often invents an answer
It hands off to a person, and you can review the conversation
How you test it
A few demo questions on launch day
Real questions from your work, tried before it goes live
How it sits on the site
A second chat competing with the first
One entry point, available where people need it
We do not claim the assistant will never be wrong. We ground it in your material, test it on real questions, and keep a path to a person.
Published ranges, then your estimate. 1 The assistant; 2 The source work; 3 Your quote.
QuoteThe exact figure comes from an estimate or a call
Costs and timelines
What it costs on the published price table
These are the published ranges on Bonzai’s price table. Your exact price and timeline come from the live estimate or a short call.
Typical timelines for AI That Knows Your Business
Project
Typical timeline
What shapes it
AI assistant
2 to 4 weeks
Grounded in your own documents and data, tested on real questions.
The notes below come from Anthropic and Nielsen Norman Group pages we fetched on 22 November 2026. They are not a guarantee that any one assistant will be error-free.
Research we fetched on 22 November 2026, used when we design and test assistants for clients.
Start simple; add an agent only when you need one
The most reliable implementations begin with retrieval and in-context examples. We ground the assistant in your material first. We only add more moving parts when a simpler setup cannot do the job.
Waiting for a huge test set usually means you are reverse-engineering success from a live system. We start with a few dozen real questions and known failures, and we grade the outcome, not a scripted path.
People should not have to guess which chat to open, and the assistant should say what it can do, offer suggested questions, and stay available as they move around the site. When it cannot help, it should hand off to a person.
A personA hand-off is part of every assistant we ship
FAQ
Questions about AI that knows your business
Which AI models do you use?
We usually build on Anthropic’s Claude through its API, and choose the model for each job by cost, speed and quality.
Will the assistant make things up?
It answers from your own documents and data, and is tested on real questions before launch. When it does not know, it says so and hands over to a person.
Is our data safe?
Your data stays in your systems and is sent to the model only to answer a question. We use providers whose business terms do not train on your data, and keep personal details out where we can.
How long does an AI assistant take?
Typically two to four weeks, grounded in your documents and data and tested on real questions.
Is this the same as Google Analytics?
No. Analytics tools can be one source. A Bonzai dashboard is a leadership view that can bring traffic, leads and revenue together, with definitions you agree on and a written reading of what to do next.
How often does a dashboard update?
It depends on the sources. The view should say when it was last refreshed. We will tell you, in the quote, whether a number is live, daily or pulled on a longer cycle.
Will a person still be involved?
Yes. Every assistant we ship has a hand-off to a person when the question leaves the approved material, or when a draft still needs a human to send it. The way out should be obvious, not hidden behind three wrong turns.
What material do you need from us?
The documents and pages you already trust: policies, service pages, product notes, FAQs, and the phrases your team already uses. If it should not be said, it should not be in the source set.
Live cards, and recent reading. 1 Related work; 2 Latest in this category.
Our workOnly live cards already on /work
MonthlyThe latest-news list is refreshed by hand
Work and what’s new
Work that already lives on this site
We only link to Our work cards that are already live. The “Latest in …” list is curated, dated, and checked about once a month.