38 of 51 are live for every client. 3 are switched on client by client. 7 are built and switched off. 3 are not built.
The Platform, thing by thing
Your AI team7 of 8
Know your company3 of 4
Automate7 of 9
Stay in control9 of 12
Measure and improve5 of 7
Everywhere7 of 11
Each mark is one thing the Platform does or does not do. Point at one to read it, or press it to go to its row.
Live for every client 38
Switched on per client 3
Built, switched off 7
Not built 3
51 marks in six rooms. 38 are solid, live for every client. 3 have a dot, switched on per client. 7 are hatched, built and switched off. 3 are coral and dashed, not built.
The Platform is configured for each client on day one, on infrastructure that is theirs alone. Here is what it is made of, what it does today, and what it does not do yet.
Every row was checked against the Platform's code on .
2,358,190lines of application code in the Advizr PlatformBuild artefact15 Oct 2026
3,721test files that run on every changeBuild artefact14 Sep 2026
54workflow templates a client can start fromBuild artefact15 Oct 2026
32text models a client's admins can allow, from 9 labsBuild artefact15 Oct 2026
Counts of what exists, read from the code. They describe the Platform, never a result for a client.
01Isolation
One client, one stack.
Every client gets isolated infrastructure: its own repository, database and deployment. Your business data never shares a table with another company.
Sign-in service: your team's accounts
Your company
Repository
Deployment
Database
Another client
Repository
Deployment
Database
Another client
Repository
Deployment
Database
Model gateway: requests go out to the model makers
Three clients, drawn as examples. The stacks do not touch. Sign-in and the model gateway are the two shared parts, and the page says so.
What is shared
Sign-in, and our own engineering access. Sign-in runs on one Advizr service in Canada that holds your team's accounts. Advizr engineers can open your workspace to build and support it, and what they change is recorded like anyone else's.
Where it runs
Your database can sit in Canada or the US, chosen when it is created. The app servers and the AI models run in the United States today. Ask us and we show you where each part runs, before you sign.
Who can open what
Every request is checked against the person's role on the server. Viewers read and admins decide. Remove someone from your workspace and their sign-in stops working within a minute.
What leaves
Model requests go through one gateway to the model makers and the clouds that host their models. Advizr never trains a model on your data unless your team turns that on, and the database refuses a training job without recorded consent. The default models come from makers whose business terms rule out training on your data.
51 things in six layers. Each one says whether it is live for every client, switched on for some, built and off, or not built. The notes say where it stops.
Live is on for every client. Per client is built and switched on workspace by workspace. Built, off is finished and left off until it is proved on our own work. Not built is not built.
Your AI team
Who does the work, and how you ask for it.
7 of 8 live for every client
Live for every client. Chief of Staff leads the teamA new chat starts with the Chief of Staff, who answers or hands the work to the right teammate.
Live for every client. Coworkers by departmentThe org chart groups coworkers by department and shows who hands work to whom.
Live for every client. Ready-made coworkers to hireEach card says what the coworker does in its first week and which apps it needs. Roles whose craft needs a switched-off capability, such as design or code, are not counted.
Live for every client. Hire a coworker from one sentenceThe draft coworker does a practice task before you hire, and the hire receipt carries an undo window.
Live for every client. Teach skills in plain wordsAnyone on a client workspace can teach a coworker how the business does something. The skills store is off for clients.
Live for every client. Hand-offs between coworkersA hand-off card shows who took the work, and the result comes back into the same chat.
Live for every client. Team runs for big jobsA Team run step in a workflow splits the work across coworkers inside a set budget. Starting one straight from chat is off for clients.
Decks and pages in your brandBuilt, offStudio and the design tools are off, so a designer coworker cannot do the work yet.
Know your company
What it knows about your company, and how it says where an answer came from.
3 of 4 live for every client
Live for every client. Answers that show their sourcesEach Brain answer cites what it read. When nothing covers a question, it says the brain does not have enough on it yet.
Live for every client. Sources that re-sync on their ownConnected sources are read again on a fixed interval by default. Checking as often as a source changes is off for clients.
Live for every client. Memory that carries across chatsWhat you tell the team is recalled in later chats, and you can keep, fix or forget it.
Memory of people and companiesPer clientThe Relationships tab in the Brain shows who you deal with and what coworkers remember about each. It is on by default but newer than the release clients run, so it reaches each client with the next update.
Automate
Work that runs on a schedule or a trigger, and how you try it first.
7 of 9 live for every client
Live for every client. Workflows from a sentenceDescribe the work and the AI drafts the whole workflow on a canvas you can read before it runs.
Live for every client. Practice runs before going livePractice runs every step on real data and skips anything that sends. On most workspaces Test sends for real.
Live for every client. Systems Advizr runs, with statusEach workflow the Advizr team built shows a running mark and how its last run went. The fuller systems roster is off for clients.
Live for every client. Routines set up from chatA coworker checks in on a schedule and brings back what it finds. People who manage coworkers set these up. Members setting their own is off for clients.
Web page watchersPer clientA coworker watches a page and wakes only when its text changes. It is new, and it reaches each client with the next update.
Live for every client. Forms, triggers and bulk runsClient admins and the Advizr team set these up in the admin pages. A bulk run works through a spreadsheet's rows at the pace you set.
Live for every client. Standing orders on a scheduleA mission runs one coworker on a schedule inside the autonomy level and monthly budget you set.
Live for every client. A task board shared with AIYou and your AI team line up work, hand it out and close it on one board.
Record a task onceBuilt, offBuilt and off by default, and not yet used on a client's real work.
Stay in control
What stays with you: approvals, limits, receipts, isolation and the models that may run.
9 of 12 live for every client
Live for every client. Approve in app, email or SlackRisky outward actions wait for a person by default, and an admin can relax that. An email link approves only an ask that needs no sign-in. Slack buttons appear on work started in Slack.
Live for every client. Rules and limits you setRules decide what waits for you, blocked actions stop outright, and daily send limits cap what goes out.
Live for every client. Spending caps and a pause switchAdmins set a monthly cap and per-person limits, and can pause all AI work at once.
Live for every client. Receipts for actions, with undoOutward actions leave a receipt you can open, with undo while the action can still be taken back.
Live for every client. Finished work checked automaticallyA completeness check grades each coworker's final output before the run settles. Grading by kind of work is off for clients.
Live for every client. A separate database per clientBusiness data lives in a database provisioned for one client only.
Live for every client. No model training on your dataTraining is off unless your team turns it on, and the database refuses a training job without recorded consent.
Live for every client. Choose which models may runAdmins allow models by company and choose whose servers may process requests.
Live for every client. Backup models if one failsEach tier has fixed backup models, and the standard and cheap tiers fall back to another company's model.
Rules typed as one sentenceBuilt, offBuilt and off. Writing a rule as one sentence is on for no client yet.
Single sign-onNot builtNot built, and neither is multi-factor sign-in. Your team signs in with an email and a password, and repeated wrong guesses are slowed and then blocked.
Bring your own model keysNot builtNot built. Admins choose which models may run, but requests go through our gateway and not through a key of yours.
Measure and improve
What the team did, what it cost and what it learned.
5 of 7 live for every client
Live for every client. Goals with pace and forecastEach goal shows a pace bar against its date, and a forecast once there are enough readings.
Live for every client. A weekly update on your teamWhat the team did, how goals are going and what needs you.
Live for every client. Value and cost per coworkerAnalytics shows what each coworker delivered in dollars and hours against its share of spend. Per person it shows usage, not value.
Live for every client. A monthly report to forwardReports turns the same numbers into a monthly edition you can read, download and forward.
Live for every client. A learning loop you approveAfter a run, a background review re-grades what was remembered and proposes skill changes that wait for approval.
A fix loop that tests fixesBuilt, offBuilt and off. It runs only where its switch is on, and that switch ships off.
Rehearse a decision firstBuilt, offBuilt and kept off.
Everywhere
Where your team meets it: chat, Slack, email and the apps you already use.
7 of 11 live for every client
Live for every client. Plan-first chatPlan first drafts a plan, and nothing runs until you say go.
Live for every client. Talk instead of typingDictate into the message box, keep listening for a spoken back and forth, or have replies read aloud.
Live for every client. Slack and email channelsYour team answers in Slack and at its own email address once Advizr connects them. Connecting Slack yourself is off for clients.
Live for every client. A catalogue of 1,500+ appsThe curated apps connect directly. Every other app offers Request this connection, and the Advizr team sets it up. One-click connect for apps with managed sign-in is switched on per client.
Live for every client. Claude and Cursor over MCPAn admin copies a workspace key and address from Settings. Signing in from an AI tool is off, so ChatGPT's own connectors are not claimed.
Live for every client. A portal for your customersYour customers sign in separately to see their requests, outcomes, approvals and invoices.
Live for every client. Chat on your websiteOne script tag puts a coworker on your site, answering only on the sites you list.
A desktop app on a shortcutPer clientAdvizr Desk puts the Chief of Staff in a small box over any app. It is turned on per workspace.
Talk modeBuilt, offBuilt and off. The full voice conversation is on for no client yet.
Microsoft Teams as a channelNot builtTeams is a catalogue app with no sending path, so it is not a channel. Slack and email are.
Forge for each person's weekBuilt, offBuilt and off. It stays off everywhere until the first full session is proven on Advizr's own workspace.
Three of these are not built. If one of them decides your purchase, it is better to know before the first call.
03The stack
What it is built on.
The tools that run inside the Platform. The model makers sit behind one gateway and are named once, on the security page, with every other vendor.
The appOne codebase, typed from end to end.
Next.js + ReactTypeScriptTailwind CSSVercel AI SDK
Data and retrievalYour documents are indexed inside your own database.
Where it runsServices and schedules on a standard platform, shipped through CI.
RailwayVercelGitHub ActionsDockerPython (FastAPI)
ConnectionsApps connect through a broker, and the Platform speaks one open protocol.
ComposioMCPSlack
ChecksNothing ships on hope.
Golden datasets and an eval harnessPlaywright
04Reliability
Judgment where judgment is the job.
A chain of steps multiplies its errors. Five steps that are each right nine times in ten finish right 59 times in a hundred. So the Platform spends model judgment only where judgment is the job. Everything that can be code is code, and code works or fails loudly.
1 step
90%
2 steps
81%
3 steps
73%
4 steps
66%
5 steps
59%
Compound success rate of a five-step workflow at 90 percent per-step accuracy. Arithmetic, not a benchmark.
Outward actions leave a receipt you can open, with undo while the action can still be taken back.
Where a person decides
Coworkers ask before they send anything outside your company, move money, delete data or change access. Your admins can relax that, one conversation at a time. Workflows run the steps your team approved when they built and turned them on, and each step that sends something leaves a receipt.
05Proof
How we know it works.
Two checks we run on ourselves, with the method in the open.
An acceptance run
Answers from a firm's own documents, graded.
48 of 51answers grounded in a matching sourceSystem logged17 Jul 2026
8 of 8questions it had no source for, declinedSystem logged17 Jul 2026
Every answer was graded by Advizr's harness: grounded, cited and drawn from the source that holds the answer. Of the three misses, one cited the wrong source and two declined a question they could have answered. The client's own grading sheet is blank, so this is our measurement, not their sign-off.
Breadth claims across hype domains predict mediocrity in all of them. We build AI systems for operations. That is the whole list.
02
Fine-tuning when retrieval solves it.
A fine-tuned model is expensive to build, hard to audit, and stale the day training ends. If the job is knowing your documents, the answer is retrieval, and we will say so in the first call.
03
Agents where a script will do.
Model judgment is spent only where judgment is the job. Everything that can be deterministic becomes deterministic code that either works or fails loudly.
04
Customer-facing AI without a gate and a test suite.
No eval suite, no human approval path, no launch. We size the harness to the risk, but it never rounds down to zero.
05
Guaranteed accuracy percentages.
Anyone promising a number before measuring your data is guessing. We commit to a measurement method first, then to the numbers it produces.
07Field notes
The thinking, published.
Field notes on how these parts were built and where they broke.
Your custom strategy and systems, delivered in under a month. Your named people using them every week by day 30. Or we keep working for free until both are true.
The page ends on a hand drawn chart of the sky that draws itself as it comes into view, with a small sail ship crossing its open sky from a far star to a beacon and a small planet.