What Beaver is

Beaver is the assistant inside ZaaS. It finds your company's approved knowledge, applies the prompts and skills your team shares, and turns useful work into something the team can reuse.

That last part is what makes it different from a chat window. A conversation is yours and disappears into scroll. A saved piece of work survives, and can be handed to someone else.

Five things people need from it

Most people live in the first three. The last two are what turn individual effort into something the company keeps.

Find it

Ask for the job, not the document title. "What do we do when a client asks us to skip the check-in?" works better than "client onboarding policy".

Beaver searches your organisation's approved knowledge and pulls in the five closest matches.

Prompts and skills are not chosen for you. They live in the AI Library, and you bring the right one into the conversation yourself. That is a fair trade: you know which workflow fits the job, and a wrong skill applied silently is worse than no skill at all.

One thing to know about how it matches. Beaver looks for the words you use, not the meaning behind them. It handles ordinary variations well: sign, signed, signing find each other. It will not guess that time off means annual leave if your policy never says "time off". So:

• Ask using the words your documents use. The policy wording is the best search term.
• If it keeps missing a phrase your team uses every day, that is worth fixing at the source — ask an admin to add the everyday wording to the document. Then it works for everyone after that.

Understand it

Every answer ends with a Sources section. Read it. It is the difference between an answer you can use and an answer you should ignore.

• Sources listed. Beaver found relevant approved material and names it. Open the source if the answer matters.
• "No sources." This message has one meaning: nothing in your approved company knowledge matched, so the answer comes from the assistant's general knowledge. That is fine for drafting and brainstorming. Do not treat it as company policy.

The line above every conversation says it plainly: check the sources before you rely on an answer.

Beaver will also say when it does not have enough to go on, and ask for the purpose, audience and constraints it needs. Take it up on that. A sharper question gets a better answer than a retry.

Use it

Three ways, in ascending order of how much of your company's thinking they carry.

Just ask. Plain conversation, grounded in approved documents.

Use a shared prompt. The AI Library holds prompts your colleagues have written and shared. Open one, copy it, and start from there. A good prompt is a captured way of working: it carries the structure, the tone and the questions that experienced people ask without thinking. Adapting one is faster than starting blank.

Follow a skill. A skill is a multi-step workflow — the steps to do a job, not a single prompt. When your team keeps re-explaining the same process, it becomes a skill and lives in the Library. Read it, then work through it with Beaver.

There is also draft check, on the Beaver page. Paste your draft, and the same assistant uses your company's standards to tell you what is missing before you send it.

Keep it

• Save to My Work. Each answer has a save button. It becomes an artefact you can keep, return to and edit.
• Set up a project from this answer. If an answer is really a plan, turn it into a project with a name and a first step.
• Rename or delete a thread. Your list, your call.

Saved work is individual value. It is also the raw material for the next step.

Share it

A saved artefact is private by default. Nothing is shared until you choose to share it, and that choice is visible on the card.

Share a piece of work. Marking an artefact as shared makes it readable by everyone in your organisation. There is a "share with people" list: it records who the work is for. It does not decide who can open it, so if something must stay private, leave it in My Work.

Put it in review, then sign it off. An artefact carries a state — draft, in review, shared, signed off — so work can get a second pair of eyes without a meeting.

Publish a prompt or a skill. A prompt can stay personal, go to your team, or go to the whole company. A skill is written for the whole organisation. This is the most direct way to give your team something reusable: take the prompt that worked, strip the specifics, publish it.

Raise an improvement. The Ideas board is where the team says what would help most. Describe the problem more than the solution; people vote on what matters.

The step only an admin can take. When a piece of work is good enough to become company knowledge, an admin can turn it into a library document. Then Beaver cites it in everyone's answers, not just yours. That is the loop closing: your work stops being yours and starts being the company's.

Working alongside the assistant you already use

Many people will still use ChatGPT, Copilot or Claude for the first pass. That is fine. Beaver does not need to win the whole job.

The useful pattern is to bring the result back:

1. Think or draft wherever you think best.
2. Bring the result into ZaaS and ask Beaver to check it against your company's standards and approved knowledge.
3. Save it, so it survives the chat.
4. Share the reusable part — a prompt, a structure, a checklist — so the next person does not repeat the work.

The first pass is yours. What comes back into ZaaS is what the team gets to keep.

What Beaver can see

When you ask a question, Beaver works from four things:

1. Your conversation. This thread and what has been said in it.
2. The documents you attached to this thread. Under "Context for this thread" you can choose up to ten documents from the company library.
3. Approved company knowledge that matches your question. Beaver searches your organisation's approved documents and pulls in the five most relevant.
4. Your instructions for this thread, if you have set any.

What Beaver cannot see

• Anyone else's conversations. Not a colleague's, not an admin's.
• Another company's documents. Ever.
• Documents that have not been approved yet. If something is marked as being reviewed, Beaver will not cite it. That is deliberate.

Approved documents are readable by everyone in your organisation. If a document should not be read by everyone, it should not be in the library — there is no per-role or per-department restriction yet.

If a document you expect is missing from an answer, it is almost always one of two things: it has not been added to the library, or it is still waiting for approval. Ask your admin.

Your conversations are private

This is enforced in the database, not just in the interface:

• Threads and messages are yours alone. The database refuses anyone else's request for them, including an administrator's.
• Admins see cost, not content. An admin can see how many questions were asked, which model answered, and what it cost. They cannot see what you asked or what Beaver said.
• Sharing is a choice you make. A saved artefact stays private until you mark it shared.
• Your key is never displayed. If your organisation uses its own AI account, the key is stored encrypted and shown only as the last four characters. Not even an admin can read it back out.

One thing this does not cover: support. If you ask us to investigate a specific answer, we will tell you what we need to see before we look.

The monthly limit

Your organisation sets a monthly limit on AI use, so the bill never surprises anyone. There may be a limit for the whole company and one for each person.

When you reach it, Beaver stops rather than overspending. You will see a clear message rather than a silent failure, and it tells you exactly where you stand — for example "$4.20 of $5.00". You have two ways forward:

1. Wait for the new month. The limit resets.
2. Ask an admin to raise your limit.

The limit applies whichever account is paying, so it is not a technical ceiling you can route around. That is deliberate: it is the thing that makes an AI tool safe to leave switched on.

A useful thing to know: short, focused questions are far cheaper than long ones with a lot of attached documents. If you are near the limit, ask for the one thing you need rather than a comprehensive review.

Asking well

Beaver responds to specifics. A few habits that make a real difference:

Say what you want produced. "Write a one-page brief" beats "help me with the shift". Name the output.

Say who it is for. A note for the warehouse team and a note for the client are different documents.

Give the constraint. "Under 200 words." "No jargon." "List the steps in order."

Attach the right document. If the answer should follow a specific procedure, attach that procedure under "Context for this thread". It is faster and more accurate than hoping Beaver finds it.

Ask for the gap. "What is missing from this record?" is often more useful than "improve this record".

Use the words your policy uses. It searches wording, not meaning. The document's own phrasing is the strongest search term you have.

When something goes wrong

What Beaver is not

• Not a search engine. It reads your company's material, not the web.
• Not an oracle. It can be wrong, fluently. That is why sources exist.
• Not autonomous. It does not send emails, change records, or act on your behalf. It drafts and answers; you decide.
• Not a replacement for judgement. A policy answer is a starting point. If the stakes are high, check the source document, then check with the person who owns it.

The habit that makes it worth having

Find it, understand it, use it, keep it — that is one person getting faster.

Share it is the one that compounds. Save the thing that worked. Strip the client's name out of the prompt and publish it. Put the draft in review instead of emailing it around. Tell the Ideas board what keeps getting in the way.

Every one of those takes a minute and saves somebody else an hour. That is the whole difference between a company that has an AI tool and a company that has company knowledge.

What you are setting up

One decision, then six pieces of administration.

The decision is where Beaver's intelligence comes from. Everything else is administration: switching it on, setting a limit, adding your company's knowledge, choosing models, telling your people, and a short monthly review.

What you are really building is a loop. A person finds an approved answer and uses it. They save the result. They share the reusable part. It becomes company knowledge, and the next person starts further ahead. One person gets faster; the team gets better; the business keeps what it learned.

Most of the value sits in step 4 and the monthly habit in step 7. The rest is setup you do once.

The whole thing takes about an hour, spread over a week while you gather the documents.

What is ready today

This guide describes the setup we are aiming at. Two pieces are still being finished, and both say so where they appear:

• ZaaS providing the AI — being built. Until then, choose option 2 or 3.
• A live spend panel in Beaver settings — being built. Until then, the month's AI cost appears in your monthly snapshot.

Everything else in this guide works today.

Before you start

Three questions worth answering first, because they shape everything else.

Who owns this? One person needs to be able to say yes to a higher limit, and to decide what goes in the company library. Name them now. It is usually the person who would field the "the AI said something wrong" conversation.

What will people ask it? Write down the five questions you expect most. If you cannot name five, this is not ready to switch on — you will get a month of curiosity and no value. Good first questions usually sound like "what is our procedure for X" or "help me write Y".

Who approves company knowledge? Only approved documents get quoted in answers. Someone has to be the approver. It is a fifteen-minute-a-week job.

Who keeps the prompts and skills worth keeping? Anyone can publish one — there is no approval step — so the question is not who permits it, but who notices. One person should look at the Library monthly, promote the good ones and delete the rest. Without that, it fills with things nobody would use twice.

Step 1 · Decide where the intelligence comes from

There are three ways. They are not equally ready — the status column is honest about that.

Which one to pick

If you just want it to work: option 1. Nothing to set up, nothing to explain to IT, one bill. Start here when it is available.

If you already pay for ChatGPT or Claude at the right tier: option 2. It is the cheapest sensible route because you are spending credit you have already bought. Read the trap section below first — most people are on a tier that does not qualify.

If you want control over exactly which models are used, or you already have a provider account: option 3. Ready now, ten minutes to set up.

The traps, in plain terms

This is where most organisations lose an afternoon.

• ChatGPT Plus does not come with API credit. Neither does Claude Pro. Those are the consumer plans most people have. They give you the chat app; they do not give you a key.
• The tiers that do qualify are the business and pro-sumer ones: ChatGPT Pro, ChatGPT Business, Claude Max, Claude Team. Check your actual tier before promising anyone anything.
• Enterprise plans usually do not qualify either. ChatGPT Enterprise and Claude Enterprise are chat products. If your company runs on one of those, option 2 is not your route.
• Google is different. Google AI Pro and Ultra give you quota inside Google's own tools, not a key you can use elsewhere. Do not plan around it.
• Microsoft 365 Copilot is its own case. The add-on does include a way for software like Beaver to ask questions, at no extra cost. It is text-only and still being finished by Microsoft. Treat it as a real option if you are a Microsoft shop, and a distraction if you are not.

If you are unsure which tier you are on, the fastest answer is to open the vendor's admin console and look at the plan name on the billing page. Do not guess from the price.

Step 2 · Turn it on

An administrator does this. In the app: Beaver → Settings.

1. Choose how Beaver gets its AI (from Step 1).
2. Paste the key, or complete the claim.
3. Press Verify & save. It tests the connection before saving. If it fails, the message says why — a wrong key, no credit, or a provider that cannot be reached.
4. Set the budget (Step 3).
5. Tick Members can use the assistant.
6. Save settings.

Until step 5, staff see "an admin has not switched the assistant on yet". That is the correct message, not a bug.

The key is never shown again. It is stored encrypted and displayed only as the last four characters. If you lose it, you create a new one at the provider and paste that in. Nobody, including Zorn AI staff, can read the saved key.

One setting you may not recognise. "Ask the provider for zero data retention per request" is on by default and should stay on. In plain terms: it asks the AI provider not to keep a copy of the question for its own training. Some providers can honour that per request; others can only offer it as an account setting. Leave it ticked and Beaver does the best available thing. Turn it off only if a provider rejects requests with it set.

Step 3 · Set the monthly limit

You can set two numbers: a limit for the whole company, and optionally one for each person.

The company figure is your stop-loss. The per-person figure stops one enthusiastic colleague from spending the month's budget in a week.

A sensible starting point is a low figure you would be relaxed about losing in month one, then raise it once you have real usage. As a guide, our recommendation for an everyday model is around US$1.50 per person per month, which covers several hundred short questions. Heavy users will hit it; that is information, not a failure.

What happens at the limit: Beaver stops. It does not overspend and it does not silently fail. The person sees a clear message with the dollar figure. The refused request is recorded as a cost line — who, when, which model — not what they asked. So the ledger tells you how often people hit the ceiling and who to talk to. It will not tell you what they were after, because it holds cost rather than content. Ask them; the answer takes a minute and is more useful than a log.

A practical rule: if a lot of people hit the limit in month one, the problem is usually the model, not the limit. A cheaper everyday model with a good context library answers most workplace questions well. Reserve the expensive model for the questions that need it.

The day the credit runs out. If you chose option 2 or 3, the provider account can run dry. People then see "the provider account is out of credit", which names the problem exactly. Nothing breaks and no data is lost; Beaver simply cannot answer until you top up the account or switch to a different one. Worth knowing before it happens, because it is the one failure that is entirely in your hands.

Step 4 · Add the company knowledge

This is the part that decides whether Beaver is useful or generic. Everything else is plumbing.

There are three collections, and they do different jobs. Most organisations set up the first and forget the other two, which is why their assistant can answer questions but cannot help anyone do the work.

Start with documents. Add two or three prompts in the first month, once you see which questions keep coming back.

The documents

In the app: AI Library → Context. Write them in plain markdown.

The first five documents most organisations need:

1. How we write. Tone, the words you use, the words you ban, one good example and one bad one.
2. How we do the main thing. The procedure for your core service, step by step, in the order a new person would need it.
3. Who to ask. A short list of roles and what each one owns. This saves people from asking Beaver something only a person can answer.
4. The standards. What "good" looks like for the documents your team actually produces — a brief, a report, a client update.
5. The exceptions. The five situations where the normal procedure does not apply, and what to do instead.

Four rules that matter more than volume:

• Approved only. Every document has a review status. Only approved documents are quoted. This is enforced in the database, not just hidden in the interface. If something is provisional, leave it unapproved and Beaver will politely ignore it.
• Everyone reads everything. Approved documents are readable by every person in your organisation. There is no per-role or per-department restriction yet, so a document that should not be read by everyone should not go in the library. Decide this before you paste, not after.
• Concrete beats comprehensive. A page with real examples outperforms ten pages of policy language. If a new starter could not follow it, rewrite it.
• One subject per document. A document covering invoicing, safety and brand voice will be a mediocre match for all three.

Write the words people actually use. Beaver matches on wording, not meaning. It handles ordinary variations — sign, signed, signing find each other — but it will not know that time off means annual leave if the document never says "time off". Add the everyday phrasing to the document. That single habit does more for answer quality than any other change you can make.

A good first test: ask Beaver one of your five questions from "Before you start" and read the Sources footer. If the right document is not cited, the document is not matching — usually because the question's words do not appear in the text. Add the plain-language phrasing people actually use.

The contribution loop

This is the part that separates a company with an AI tool from a company with company knowledge. It is worth understanding properly, because it is the only part of Beaver that compounds — and the only part that needs a person to keep it honest.

What a person can contribute today:

What only you can do: turn a good artefact into a library document.

A conversation answer is not company knowledge. Someone saved it, it helped, and it is still one person's note. When the same shape of answer keeps being useful, an admin rewrites it as a context document, approves it, and then Beaver cites it in everyone's answers. That is the moment individual effort becomes shared value.

The monthly habit that makes this work: open the AI Library, look at what was published since last month, and do one of three things with the best of it — promote it to a document, leave it as a prompt, or delete it. One decision per item, and the library stays worth reading.

Say who wrote it. Nothing in the software records authorship on a document you rewrite. Put the person's name in the text ("From a prompt by …"). It costs a line and it is the difference between people contributing and people stopping.

Step 5 · Choose which models people may use

You do not need to know which model is which. One rule covers it:

One everyday model, one for hard questions. The everyday model answers the majority of workplace questions at a fraction of the cost. The stronger model is for long documents, careful drafting and anything where being wrong is expensive.

Make the everyday model the default. If your organisation has opinions about which vendor, that is a legitimate reason to choose — but do not pick the expensive model as the default "to be safe". It is neither safer nor better at "what is our leave policy".

You can restrict the list to models you are willing to pay for. Anything not on the list cannot be used.

Step 6 · Tell your people

Do not announce it by email and hope. Twenty-five minutes, together, on one screen.

1. Two minutes: what it is. An assistant that reads our documents, shows its sources, and is private to each person.
2. Three minutes: the sources rule. Read the footer. If it says no sources, it is general knowledge, not company policy. This one habit prevents almost every bad outcome.
3. Five minutes: three real questions. Use three of your five from "Before you start". Ask them live, in front of everyone, including the one that gets a poor answer. Watching it say "I do not have enough information" builds more trust than a polished demo.
4. Five minutes: saving and sharing. Save an answer to My Work. Then share it with the room, so everyone sees what sharing actually looks like. Explain the one rule: private until you choose otherwise.
5. Five minutes: the library is theirs. Show them a shared prompt and a skill. Tell them anyone can publish one, and what makes a good one: take the prompt that worked, strip the client's name out, say what it is for.
6. Five minutes: the limits and the ask. What the monthly limit is, what happens at zero, and who to tell when an answer is missing something.

Then send them the plain guide for users (beaver-for-everyone.md), and one line: "if it could not answer something it should have, tell me — that is how we fix it."

Expect the first week to be mostly curiosity. The value shows up in week three, when people start bringing real work to it.

Step 7 · The monthly fifteen minutes

Once a month, look at four things. That is the whole ongoing job.

You are managing from two sources: the numbers, and what people tell you. You cannot read anyone's conversations — that is the privacy promise, and it holds for admins too. So the numbers tell you where to look, and the conversations tell you what to change.

1 · Spend against limit. Are you near the ceiling? Is one person responsible for most of it? One heavy user is usually a good sign — they found a use. In the app, this is Beaver → Settings: this month's spend against the cap, and by person. Until that panel ships, the figure is in your monthly snapshot.

2 · Turnaways. Requests refused at the limit. You can see who was stopped and how often, not what they asked — the ledger records cost, not content. So ask the people concerned what they were trying to do. That conversation is usually worth more than the usage report: it tells you whether the ceiling is too low or the model is too expensive for the job.

3 · The gaps people report. This is the one you have to collect by asking. Every "it could not answer this" is a missing or badly-worded document. Keep a note during the month rather than trying to remember it; the list is your library to-do list, written by your own staff.

4 · Who is not using it. A whole team with zero usage usually means the answers are not landing for their work — the fix is documents, not training.

5 · The Library, once. What got published since last month? Promote the best artefact to a document, keep the good prompts, delete the rest. This is the contribution loop from step 4, and it is the only item on this list that makes next month easier rather than just cheaper.

Write down one change for next month. One. A long list of improvements does not survive contact with a busy month.

What your IT person may need to do

Usually nothing. Here is the honest list, so you can ask the right question.

• Option 1 (ZaaS provides it): nothing.
• Option 2 (subscription credit): someone with owner access to the AI vendor's workspace creates a key and sets a spend limit on it. Ten minutes, no network changes.
• Option 2, Microsoft route: a Microsoft administrator approves the app once. This is the only option that needs an IT ticket in a locked-down environment.
• Option 3 (your own key): create an account, add a payment method, create a key. Ten minutes.
• All options: nothing on your network needs to change. Beaver talks to the AI provider from our servers, not from your office, so there is no firewall rule, no proxy exception and no allowlist to arrange.

There is no server to install, no agent on anyone's laptop, and no data copied onto your machines. People use it in a browser like any other work tool.

One-page checklist

Decide

• Owner named
• Five questions written down
• Approver for company knowledge named
• Access route chosen (and plan tier checked against the traps)

Set up

• Key added and verified, or claim completed
• Company monthly limit set
• Per-person limit set
• Everyday model set as default
• Model list restricted to what you will pay for
• Switched on for members

Make it useful

• Five core documents added
• Each one approved
• Tested with a real question, sources checked
• Documents rewritten where nothing matched

Tell people

• Twenty-minute session run
• User guide sent
• One named person to report gaps to

Then

• Monthly review in the calendar
• One change per month

Appendix · What happens on our side

For the person who will ask "but where does it actually live". No jargon.

• Your key is encrypted before it is stored. Only the server can turn it back into a key, and only for the length of one request. It is never written to a log, never shown in the interface, and never visible to our staff.
• Every question is recorded as a cost line, not as a conversation. Model, tokens, cost, who. Not what was asked. That is how an admin sees spend without reading anyone's chats.
• The spend check happens before the question is sent. If a limit is reached, the request is refused and logged as a refusal. That is why a limit can never be overspent by a burst.
• There is an off switch. An admin can turn the assistant off for everyone in one action, and nothing else in the product stops working.
• Chats are private by construction. The database policy refuses any request for a conversation that is not yours, admins included. This is a promise we can make because it is enforced below the application, not merely hidden in the interface.
• Your documents stay yours. Company knowledge is searched per organisation and approved-only. There is no cross-organisation path.

If any of that changes, this page changes with it.