AI Implementation · Victoria, BC
Connect the AI you already pay for. Test it before it goes live.
Fixed-price builds for teams on Claude, ChatGPT or Copilot: a connector to one business system, answers from your documents with the page each came from, and rules your coding agents cannot skip.
Describe your taskSee the demosOur Own Test
229 annual reports, 2,129 figures read by AI.
Solutions
Start with one build.
For Teams That Run Coding Agents
The Test Record
Every build ships with its test record.
- 1What was builtThe system, the assistant and the dates.
- 2What it may doEvery lookup and action, and who can use it.
- 3What failedEach failed test, its cause and the fix.
- 4Pass rateFrom the final run, before go-live.
- 5Sign-offWho approved it, and when.
Founder · Victoria, BC
Marc Charbonneau
Marc spent seven years testing protection systems at BC Hydro before they went into service. Arbutus Cambium brings that habit to AI integration by defining what a pass looks like before a tool enters everyday use.
Have a task in mind?
Tick through one page: the AI you have, the systems involved and what a wrong answer costs. We reply with a scoped proposal.
Open the task checklistSolutions
Two builds for your assistant. One set of rules for your coding agents.
Each has a fixed scope, a price and a test it must pass on your records. Need something else? The checklist takes any request.
Builds
Rules
Demos
Watch each build run on invented records.
No demo touches a real customer's data. Each one ends with its test record, failures included.
About
A testing habit, applied to AI.
Marc Charbonneau
Founder
Marc spent seven years at BC Hydro as a Communications, Protection and Control Technologist, testing protection systems before they went into service. Arbutus Cambium applies the same rule to AI: nothing goes live until it passes a test on your own records.
Outside work he surfs, skis, rides and takes photographs.
The Name
Cambium
Cambium is the thin layer under a tree's bark where new growth forms. We work at the same layer: where AI meets the work your team already does.
- Company
- Arbutus Cambium Inc. [Incorporation pending]
- Based in
- Victoria, BC
Start Here
Describe what you need. We reply with a scoped proposal.
Tick what fits, whether it is one of our builds or not. The proposal names what we would build, the test it has to pass on your records, and a fixed price. [Set the reply time]
The Checklist
Questions
- Which AI does it work with, and in which apps?
- Claude, ChatGPT and Microsoft 365 Copilot for the two builds. In Claude, a build works on the web and in Claude Desktop. In ChatGPT, OpenAI offers custom connections on the web version only. In Copilot, a build arrives as an agent your admin installs, used in Microsoft 365 Copilot and the Microsoft 365 apps that support agents. Claude Code, Codex and GitHub Copilot for the rules. Your team keeps the software and the assistant it already has.
- Is it a one-time install, or does it need upkeep?
- Each build is a one-time fixed price with 30 days of fixes. AI vendors update their apps often, and an update can change how a build behaves, so each build has an optional monthly plan that runs its test again and fixes what fails: $350 CAD a month for a connector, $500 CAD for document answers, and from $350 CAD for the coding-agent rules.
- Do we need a particular AI plan?
- For a team on ChatGPT, custom connections need Business, Enterprise or Edu. In Claude they work on every plan, and on Team or Enterprise an owner adds them for everyone. In Copilot your admin installs the agent; whether staff without a Copilot licence can use it is checked when we scope the build.
- What happens when the AI gets something wrong?
- Before go-live, each build is tested on questions with known answers from your records, and the test record lists every failure and its fix. Document answers say when no page supports an answer instead of guessing. A wrong answer found in the first 30 days is fixed as part of the build.
- Can it change or delete our data?
- Not by default. A connector only reads, and each person sees only what they can already see in the system. Creating or updating records is a separate option, from $7,500 CAD, and the assistant asks the person before each change.
- Where does our data go, and is it used to train the AI?
- Records stay in your systems. What the assistant reads to answer goes to the AI vendor you already use. Files that cannot leave the building are read by a model on your own computer. On business plans, OpenAI, Anthropic and Microsoft each state that they do not train their models on your data by default.
- Do we need an IT person?
- Not for the two builds. Someone with admin rights to your AI account installs the build while we walk them through it, and the install guide covers each step. The coding-agent rules for 26 or more developers go out through your device management tool.
- What if our request is not listed?
- Tick "Something else" on the checklist and describe it in one line. We reply with a fixed quote, a smaller first step, or why we are not the right fit.
- How long does a build take?
- One to three weeks, by build. [Replace with measured times after the first builds.]