The business OS · Agentic AI

Run the company
by talking to it.

Over a thousand actions across all eleven modules are exposed to an agent as tools — raise the invoice, approve the leave, add the vendor, score the call. It calls the same endpoints the screens call, sees only what your role allows, and shows you every write before it happens.

Not a chatbot bolted on. There is no second, looser path into your data. The copilot is another client of the same API, under the same authorisation.

Copilot · any screen, any module

Raise the March invoice for Palm Cove Resorts from last month’s approved timesheets, and tell me if they still owe us anything.

list_timesheetsMarch, approved only · 6 members
get_customermatched one customer, GSTIN on file
list_receivables2 invoices open, 1 overdue 19 days
create_invoiceheld — preview shown, waiting for you

Ready to raise: ₹4,86,000 across 6 members at their project rates, tax applied from the customer’s state. Separately, they are 19 days overdue on an earlier invoice — worth mentioning before this one goes out. Nothing has been written yet · the invoice is a preview until you confirm

Every write works this way — preview first, then verified after

Where it already is

Not one assistant. Fifteen jobs.

Every one of these runs in the product today, on the same key store and the same permission model.

Ask, and it does itMore than a thousand actions across eleven modules are exposed to the copilot as tools — raise an invoice, approve a leave, add a vendor, log time, move a deal. Not a search box over your data: the same operations the screens perform.
It can only do what you canThe tool list is filtered to your tenant’s licensed modules and your roles before the model ever sees it, and calls are routed to the owning service with your own token. The copilot has no privileges you do not.
Nothing is written without youWrite actions are intercepted, previewed and held for confirmation. Once you approve, the copilot re-reads the record to verify the change actually landed instead of reporting its own success.
An analyst for your survey dataOn the research side it is not answering from memory — it runs cross-tabs with significance, key-driver regression, TURF and segmentation as tool calls, checks the base, and answers. Every figure traces back to the function that produced it.
It builds the dashboard tooPointed at a dataset it works in seven passes, and the hard ones are mandatory — cross-tabs with significance, advanced analytics, then a verification pass over its own numbers before anything is assembled. It cannot skip to four pie charts.
And can run it again next waveEvery tool call is recorded in order as a recipe stored beside the dashboard, so the same analysis replays against new data — numbers only, narrative rewritten, or from scratch. Reproducible AI analysis, which is rarer than it sounds.
A second pair of ears on the callIn a live meeting it follows the transcript, waits for the other side to finish, and puts an objection, a missed next step or a suggested close in front of the host — privately. Three modes, from prompt-me-only to autonomous.
And in the interview roomWhen the meeting is an interview it pulls the candidate and the job description from HR and rates answers as they happen — demonstrated, partial or not yet — with the follow-up worth asking next.
Calls scored against your playbookRecorded calls are transcribed, summarised and scored on discovery, objection handling, closing and next steps — weighted the way you weight them, against your own sales motion, ICP and value proposition. Upload your own rubric and it is validated before it scores anything.
A WhatsApp agent that cannot make a price upAuto-replies are checked after generation, not just instructed beforehand: every significant number must exist in your knowledge base, the conversation or the customer’s stored facts, and each quoted price must sit on the same row as the product it was quoted for. A reply that fails either check is blocked before it reaches the customer.
It tells you when it failedFailed reply jobs raise an admin alert within the hour, drafts left unapproved past the messaging window expire rather than sending late, and a weekly digest lists what the AI could not answer so the knowledge base gets fixed.
Quotes compared, never decidedIn procurement the deterministic comparison is always computed and always shown. A weighted score sits on top of it, and the AI explanation on top of that, with a confidence figure. The model describes the trade-off; it does not pick the vendor.
Your keys, your spendAPI keys are held encrypted in the platform key store and used at the moment of the call, so the model spend is on your account. Where cost is a cost of sale — a coding run on client work — it is reported per job.
It hears the roomLive captions are produced per speaker during a meeting, transcripts and summaries after it, and focus-group recordings come back as a diarised transcript. Cloud speech by default, local models where the audio must not leave your infrastructure.
It opens with what needs youThe copilot knows which screen you are on and arrives already holding the relevant figures — and on open it offers what is overdue, what is waiting on your approval and what is about to be late, before you have asked anything.

This is what the AI does across the business OS — the modules themselves are on the platform overview. The same seat opens Meetings, Mail, Chat, Drive, CRM, HR, Accounts, Projects, Procurement, Learning and Research, and the copilot reaches into all of them from wherever you happen to be standing.

Straight answers

The questions a careful buyer asks.

01

Is this a chatbot bolted onto the side?

No. It calls the same endpoints the screens call, routed to the owning module with your own token, so it is subject to exactly the authorisation the interface is. There is no second, looser path into your data — the copilot is another client of the same API, not a shortcut around it.

02

Can it change things, or only look them up?

Both. Roughly half the tool surface writes. Every write is intercepted before it runs, shown to you as a preview and held until you confirm it — and after it lands the copilot re-reads the record to check the change is actually there rather than assuming its own success.

03

What stops it doing something it should not?

Three things, in order. The tool list is filtered to your tenant’s licensed modules and your roles, so a tool you are not entitled to use is never offered to the model at all. Parameters are machine-checked against the schema before execution. And the write confirmation gate sits in front of anything that changes state.

04

Does it invent numbers?

It is engineered not to, and in the places where that matters most the prevention is deterministic rather than a polite instruction. On the research side every figure comes from a statistical function call or a validated query. On WhatsApp auto-replies, a guard checks that every significant number in the outgoing message literally exists in your knowledge base, the conversation or the stored customer facts, and blocks the reply if it does not.

05

Whose AI keys does it use?

Yours. Keys are held encrypted in the platform key store and used at the moment of the call, so the spend is on your account and the data goes where you have chosen to send it. Where cost matters — coding runs, for instance — it is reported per job rather than aggregated into a monthly surprise.

06

Does it work during a live call?

Yes, and only for the host. In a sales meeting it watches the transcript for the moment the other side stops talking and puts an objection, a missed next step or a suggested close in front of the host — never in the shared room. In an interview it pulls the candidate and the job description from HR and rates answers as the interview runs.

07

What happens when it cannot answer?

It escalates rather than improvising, and the failure is visible. Failed auto-reply jobs raise an alert to admins within the hour, drafts left unapproved past the messaging window expire rather than sending late, and a weekly digest reports what the AI could not answer so the gap gets filled.

08

Which models does it use?

Anthropic’s Claude for reasoning and tool calling, on a two-tier setup — the stronger model where the decision is hard and a fast one for mechanical passes like rewriting a narrative around new numbers. Speech uses a cloud provider by default with local models available where audio must not leave your infrastructure.

Ask it to do the thing.

Rather than finding the screen where you can do the thing.

AI is not a module you buy on top. It is included in the seat, it runs on your own API keys, and it is bounded by the same licence and the same roles as everything else. You are not buying an AI product; you are putting the company on one system and then talking to it.

See pricing Book a walkthrough →