Your people stop searching for answers and start shipping work.
An AI work platform is not a chat window with a better model. It is a workspace where the system already knows who you are, what you are working on and what you are allowed to touch — and then delivers finished work: the deck, the spreadsheet, the research report, the legacy system entry. This page is the overview for the business leaders and employees who will actually use it.
Context, already there
Projects, files, messages, meetings, mail, structure — inherited, not briefed
Deliverables
Decks, spreadsheets, documents, research, dashboards and live apps you can send on
It operates your software
GUI-level action in legacy systems with no API, isolated, pausable, handable
Governed by default
Inherited permissions, confirmation levels, quotas, watermarking, behavioural audit
The single biggest cost of enterprise AI is briefing it.
Ask any employee how they use AI today and you will hear the same complaint: they spend more time preparing the prompt than doing the work. Uploading the same policy for the fourth time, re-explaining what their team does, restating the format they need. The workspace exists to delete that step.
A project is a context boundary
Create a project or bind a local folder, and everything inside becomes context for subsequent tasks automatically. No uploads, no repeated instructions. You switch context by switching project.
Reference with @
Type @ or press + to bring in files, images, cloud-drive items — or a previous conversation, which is genuinely useful when the discussion is already done and you only need the conclusion.
Side-by-side browser view
Documents, spreadsheets, slides and web pages the agent produces open in a built-in browser panel on the right. Open more tabs and work several outputs in parallel without losing the thread.
Point-and-edit
Select any part of a document, spreadsheet, slide deck, web page or application and change only that. No full regeneration, no lost work elsewhere, much shorter revision cycles and lower compute cost.
You choose the depth of thinking
Automatic, or an explicit tier for lightness or for hard reasoning, with reasoning strength set in three levels. Use judgement for the quick question; spend the budget where the answer actually matters.
Select-and-ask, anywhere
Highlight text in a browser or Word and a floating toolbar offers search, translate, explain, summarise and polish. Turn it off where it is unwelcome — global shortcuts bring the workspace up from anywhere, pinned so it stays available.
Screenshot, circle, ask
Capture a region of the screen and ask about it. The fastest way to read a chart, decode an error message or get an opinion on a layout you are looking at rather than describing.
Read a PDF with it
Bilingual side-by-side translation, ask questions as you read, one-click summary of the whole document. Long reports become usable in a meeting rather than the night before.
It follows you
Same account across desktop, browser and messaging client, sharing one session. A task started on your phone can be finished at your desk with the context intact.
What the enterprise context actually includes. When an employee signs in with their corporate account, the platform inherits what that person is already entitled to see: messages and chats, cloud documents, knowledge spaces, meeting records, schedules, structured data tables, spreadsheets, mail, and the organisation structure itself — who reports to whom, which team owns which project. It can also act: send messages, create calendar entries, create documents and tasks, update task status, start approval flows. That last list is the difference between a tool that advises and a tool that works.
The work platform in numbers
The employee inherits theirs; the agent inherits nothing extra
A human can take any running task back
Employees, external partners, digital employees
Captured, versioned and made available to the estate
Long-running work is only trustworthy if a human can interrupt it.
An agent that runs for twenty minutes on your machine without telling you what it is doing is not autonomous, it is unsupervised. Everything in this section exists so that autonomy stays inside a frame the employee controls — and so that the employee's attention goes to the decisions, not to the supervision.
The breakdown is shown before it runs
The work panel first displays how it has decomposed the task, then proceeds step by step. If the decomposition is wrong, you say so before twenty minutes are spent.
Every step is visible while it runs
Which tool was called, which file was read, which code path executed, what the intermediate result was. Not a spinner.
Pause and take over, at any node
Stop it, take the controls, fix or redirect it manually, and let it continue. Hand-off in both directions.
Correct it mid-flight
Send a message while it works — add a requirement, change a scope, fix a mistake — and the task does not restart. This is the feature experienced users notice first.
Sensitive actions confirm
Irreversible actions prompt. Forms can be screenshotted for review before submission. Administrator credential prompts are a system-level protection that cannot be dismissed.
Squads, not just one agent
A coordinating agent decomposes the work, dispatches to specialist sub-agents in parallel, and consolidates the result — and multiple tasks can progress concurrently across tabs.
Confirmation levels: three settings
How much the platform interrupts you is a configuration decision, and it should be made per user and per task class, not once globally.
| Setting | Behaviour | Use it for |
|---|---|---|
| Ask when needed | Confirms on risky or irreversible actions. The default posture and the right one for most people. | General use, mixed task types |
| Always ask | Confirms before every consequential step. Slow, and appropriate. | Regulated work, systems of record, junior users |
| Allow all | No prompts. Only defensible inside a sandboxed environment on a bounded scope. | Isolated virtual desktop, read-only or low-risk scope |
Two habits we teach in every rollout. First, ask for an outline and approve it before the finished artefact — the revision loop is far cheaper at outline stage than after a full document exists. Second, ask it to restate the task in its own words before it starts. If its understanding is off, you have lost a minute; if you skip it, you have lost an afternoon.
Three things it does that a chatbot cannot
Research is the capability everyone knows about. The other two are the ones that decide whether employees keep using this after the trial ends.
The productivity gap in most enterprises is not writing. It is the assembly: finding the right numbers, putting them in the right shape, checking them against something, and producing it in the format a specific person asked for. This is where an agent that persists beats an assistant that suggests.
- Presentations. Page outline first, approval, then an editable file — with speaker notes editable inside the page and element-level changes.
- Spreadsheets. Cleaning, batch calculation, anomaly detection, new sheets, cross-sheet aggregation, chart generation. We ask teams to keep the source file untouched and attach a separate statistical-basis note — reviewers trust work they can audit.
- Documents. Generation and targeted local revision, including long-form work at handbook length.
- PDF. Reading with translation and questioning, extracting conclusions, converting.
- Web. Search with sources, extraction across sites, rule-based collection and archiving, batch download with renaming rules.
- Meetings. Transcription to speakers, action items and deadlines, structured minutes, mind map — and then tasks created from the conclusions.
- Mail. Drafting in several registers (formal, conciliatory, firm), or sending once a mailbox is connected.
- Data analysis. Cleaning, calculation, metric decomposition, visualisation, dashboards, narrative reports.
- Applications. Building, modifying and deploying web applications, including ones with a backend, with access control and a shareable link.
Where it stops
Data analysis is oriented towards business interpretation — what the numbers say and why they moved. Serious mathematical modelling, econometrics and statistical inference still belong in your specialist tools, and we would rather say so than let a plausible chart pass unchallenged.
Local file editing boundary. Local in-place editing of Office files currently covers presentations and spreadsheets — open in place, save in place, synchronised to and from the cloud copy. Word local editing is announced but not yet available, so Word workflows run in the browser environment for now. We confirm the current state at deployment rather than assuming it.
A large amount of enterprise value sits in systems nobody has integrated: an old inventory tool, an industry-specific client, a back-office portal that only renders correctly in one browser. The approach here is GUI grounding — understanding buttons, input fields and menus from the rendered interface, so no API or plugin is required from the vendor of the other side.
- It does not take your keyboard or mouse. It runs in an isolated sandbox desktop. You keep working; it does not minimise the document you are editing.
- It can hand over. Pause it, complete the sensitive step yourself, and let it continue.
- It works when your laptop is off. A cloud machine can carry the task on with the local machine closed, and you can move a running task between local and cloud without starting a new conversation or losing context.
- Your phone can direct your computer. Issue instructions from the phone, have the connected machine read local files and operate software, and get progress and results back — including drawing on several connected machines' files at once.
- Screenshots before submission. Web forms can be completed and reviewed as an image before anything is sent, which is the single most important control for bulk external submission.
The pattern to copy
These read as small. They are not. Each one is a task an employee currently does manually with copy and paste, and each is one your RPA programme failed to automate because the system has no API.
Cross-system document transfer
Export from one unsynchronised system, import into the next, verify totals page by page.
Bulk entry into legacy systems
Populating an old stock or distribution tool by rule, without an integration project.
Rule-based web monitoring
Checking daily whether orders have passed their dispatch window and raising a chase; or collecting a fixed weekly publication and archiving it.
Field extraction at volume
Pulling key fields from a folder of inconsistent documents into a single table.
Platform support differs by operating system and by what has been granted at the device level — screen control, accessibility and audio recording permissions are all explicit. We map that during deployment.
Deep research is not "search and summarise". It is a loop: decide whether a search is needed, expand the query across several formulations, evaluate what came back, identify what is still missing, search again, then synthesise with attribution. That loop is what produces an answer you can act on rather than a summary of the first five results.
Research skills in the market cover academic and clinical literature, evidence-based clinical reading, medical progress tracking, market and industry analysis, company and announcement analysis, and user research. Outputs are deliverables, not essays: reports, comparison tables, dashboards, slide decks, mind maps.
When it uses enterprise knowledge, answers carry sources and are traceable back to what they came from. That is the difference between a research tool and a corporate one, and it is the reason this capability belongs on a governed platform.
How sources are scored
Not all sources are equal, and the ranking is explicit rather than a black box. This weighting is what lets an employee tell why one document was used and another was not.
| Signal | Weight |
|---|---|
| Authority — publisher and author standing | 30% |
| Content quality and originality | 25% |
| Semantic relevance to the question | 20% |
| Source tier — primary against secondary | 12% |
| Recency | 10% |
| Interaction signals | 3% |
Research is iterative by design. Producing a first pass, reacting to it, and refining over several rounds is the expected interaction — not a sign that the first attempt was inadequate.
Three things this does that a chatbot cannot
The difference is not intelligence. It is that the work platform is a place where work happens, rather than a box you type into.
- You prompt it, it answers, the conversation ends
- No memory of your actual work context
- No access to the systems you already have
- Cannot be handed a task and left to finish
- You supervise it by reading what it said
- Every user rediscovers the same workarounds
- You assign a task, it runs, you approve or interrupt
- It works from your documents, systems and history
- It inherits the employee's permissions and nothing beyond them
- Long-running work can be taken back mid-flight
- You supervise it through a decision log, not a transcript
- Skills a person teaches it become shared estate capability
The employee inherits their permissions. The AI inherits nothing extra.
A work platform that operates a computer and reads documents is a privileged actor on every employee desktop. The governance model is not a feature list — it is one rule applied consistently: the machine can never see or do what the person cannot.
Tenant isolation, two layers deep
Enterprise and personal data are logically separated, and separation goes to the key level: each tenant has its own tenant key, business data is encrypted with random data keys, and sessions and memory are encrypted under multiple layers of keying. Encryption in transit uses managed key management.
Permission inheritance
The employee's identity and object permissions are followed exactly. Document, knowledge space, chat and meeting permissions continue to constrain what is reachable. Generated files inherit the permissions of the underlying file. The phrase we use in every configuration review: if an employee cannot see it, the AI cannot see it.
Team agents are separate actors
A team agent has its own identity, permissions and memory. It does not share a person's. Administrators configure its capability boundaries, tools and access. Its memory belongs to the team and is shared by team members — and isolated from everyone else.
Watermarking and DLP
Watermarking covers conversations and generated files across multiple points. Data loss prevention covers the critical content surfaces, including cloud documents. Together they are what make an accidental leak traceable rather than merely unlikely.
Behavioural audit — with a limit
Audit covers the highest-value interaction points and can be queried by person, time and operation type. Administrators cannot read employees' conversations with the assistant. Audit is for high-risk actions — permission changes, account changes, agent modification, high-consumption calls — not for supervision of every turn.
Seats, quotas and scope
Seats are allocated and reclaimed centrally; a member without a seat cannot start an agent task. Quotas are allocated per member or per group, with warnings as limits are approached and configurable escalation to administrators. Administrators can also restrict which kinds of task the assistant may perform, narrowing tool use and web retrieval.
Data handling position. Enterprise data is not used for base model training without your authorisation. Where data is stored is fixed by the deployment topology you choose, which is a decision we make with you at design time and document contractually — not an assumption. We also do not quote published interception-rate figures from security vendors as if they described this product; if you need assurance on injection defence, we will show you the controls and let your own testers attack them.
The employee inherits their permissions. The agent inherits nothing extra.
This is the single most important design decision in the whole platform, and the one most often got wrong.
One work surface, three kinds of user
Most platforms only really serve the first. An agentic business needs all three to work in the same place, with different rules.
Full internal knowledge, internal systems, and a supervisor chain. The agent works with the same permissions they do — which is exactly why the learning system matters: they are the ones who supervise it.
Scoped access to a bounded slice of your process, with their own identity and their own training record. Partner enablement is a learning problem before it is a technology problem, which is why it sits in the enterprise learning business.
The agents themselves. They need an onboarding curriculum, a performance record, and a supervision log — treated as managed assets with a job description, not as software with a version number.
Skills, connectors, and the things your people teach it
The market ships a few hundred skills and connectors, organised by tool type and by role. The more interesting half is what your organisation adds to it — and what employees end up using it for in the first month.
Skills and connectors, two different jobs
A skill answers how to do something. A connector answers where to get data or execute an action. Confusing the two is why implementations stall — you end up with excellent instructions and no way to reach the system.
Teach it by demonstration
Perform a task once in the built-in browser — filling a form, uploading a file, walking a checkout — and it learns the sequence and writes it up as a reusable skill, including the conditions for using it, the inputs it needs, and how to verify the result. This is how a non-technical expert becomes an automation author in an afternoon.
Enterprise skill library
Skills your organisation publishes become available company-wide. Skills can also be created by uploading a package, by describing the procedure in conversation, or by letting the system mine your own past tasks for candidates.
Custom MCP
Give it an MCP server address and headers and your own tools and data sources become callable. Boundary: custom MCP currently works in the local desktop environment only — it does not take effect in the cloud sandbox. Anything that must run unattended in the cloud needs a different route, and we design that at implementation rather than discovering it later.
Work partners
Preconfigured roles — data analyst, developer, designer, writer, presentation specialist, scriptwriter — each with a name, a specialism and a standard skill set. Useful for adoption, because an employee picks a colleague rather than a feature.
Work squads
Several specialist agents composing a team for one outcome: a research report that pulls in a data agent, a retrieval agent and a design agent in parallel, then consolidates. The unit of delegation stops being a chat and becomes a team.
A note on where skills live. Skills installed on a local machine persist. Skills that exist only in a cloud sandbox are destroyed when that sandbox is idle, and packages or command-line tools installed into it do not survive. Skills uploaded to the shared store and skills in the managed user directory are unaffected. This is not an implementation detail to discover in month three — it determines whether your automation estate is durable, so we design the deployment to put skills somewhere that persists.
Fifteen things employees do with it in the first month
None of these are transformation. That is the point — adoption comes from tasks people already dislike.
Weekly reporting
Pull chats, documents, meeting records and data tables into one progress report instead of four copy-paste tabs.
Requirement extraction
Distil a customer's key requirements from the conversation record, organise the product strengths, produce the pitch deck and the supporting material.
Policy interpretation
Assemble scattered data and search historical precedent, judge whether a change is favourable or risky, and flag it.
Sales diagnosis
Analyse the business data tables and produce a visualised diagnostic report with the reasoning shown.
Multi-platform consolidation
Every marketplace exports a different header format. It builds the monthly summary, ranks the top sellers, and charts it.
Folder governance
Archive the week's meeting materials by date, extract each file's title and owner, and return one table.
Cross-system transfer
Stock system out, ERP in, totals verified page by page. The classic manual integration.
Bulk web forms
Complete questionnaires under a given identity, draft the open-ended answers, screenshot for review, then submit.
Scheduled monitoring
Check daily whether an order has missed its dispatch window and raise the chase; collect and archive a fixed weekly publication.
Meeting to execution
Minutes become tasks, task status updates, approval progress reported, and the owner chased by message.
Dashboards and ledgers
Turn a structured data table into a live web dashboard that refreshes with the underlying data.
Deep research reports
Industry analysis, competitor mapping, literature review with a citation count, company and announcement analysis.
Content production
Product imagery, posters, comics, knowledge shorts and advertising clips — with storyboard and revision rounds.
Lightweight business systems
Renewal and health tracking, event registration, customer feedback — deployed and shared by link with partners outside the company.
Professional review work
Expense pre-review, resume assessment, organisational diagnosis — done by a team agent inside its authorised scope.
Document and compliance review
Policy questions, customer submission preparation, document review, all handled within inherited permissions.
How a skill travels from a person to the estate
This is the mechanism that turns your best operators into organisational capability rather than a training slide nobody reads.
An experienced operator handles an exception the standard process does not cover.
The reasoning, not just the steps — why this exception, why not the obvious path.
Versioned, with an owner and a review date. Available to every agent in that role.
Usage is logged. Where the skill fails often, that is a process problem the skill cannot fix.
Skills rot. A skill nobody has used in a year is either obsolete or nobody noticed it was needed.
What it will not do, however you ask
These limits are structural, not tuning issues. They will not change with a better model, and any vendor who tells you otherwise is describing a different product. We would rather set them now than discover them at the worst moment.
Four categories of action it will not take
- Destructive and cleanup actions. Deleting and bulk removal are excluded. There is no setting, role or confirmation pattern that turns this on, because the failure mode is asymmetric — a mistaken deletion is not recoverable by the person who notices it.
- Authentication and payment flows. One-time passcodes, verification codes, payment authorisation. It cannot read the code, so it cannot complete the flow, and we will not build a workaround that would let it.
- Documents that require a signature and carry liability. Contracts, quotations, financial documents. Your name goes on those; a model must not.
- Error dialogs needing judgement. A modal that requires deciding what the failure means and what the business consequence is — that is a person deciding, not a pattern match.
Product limits we will confirm at contract stage
- Custom MCP is local-environment only. It does not function in the cloud sandbox. Unattended cloud execution needs an alternative route, designed in advance.
- Word local editing is not yet available. Local in-place editing covers presentations and spreadsheets. Word runs in the browser environment. We verify the current position during deployment rather than promising a date.
- Desktop work mode and private deployment are separate questions. Dedicated or on-premise options are discussed for the API layer. For the desktop work mode specifically we will not assert a private deployment posture until it is confirmed in writing for your version.
- Some administrative settings are documented, some are not. The complete administrator menu tree and retention periods are not published in full. We design the governance model around what is documented and tell you which part of your requirement depends on what.
And one thing about data analysis. It interprets business data well and will produce a confident, well-formatted, entirely wrong conclusion if the underlying figures are wrong. It is not a modelling tool and does not replace one. Where the stakes are high, keep the human calculation chain — use it to assemble and visualise, not to certify.
What this will not do, however you ask
The work platform will not make your people faster on its own. It will not replace the briefing you currently do verbally, and it will not learn your business from watching — someone has to author the skill. What it does is make a well-briefed task completable without a human present for the whole of it.
Five questions that settle whether you need this
Five yes-or-no answers. If you cannot answer the first three, the capability is not your bottleneck yet.
- Can you name the specific workflow this serves — not the department?
- Can you name the person accountable when it is wrong?
- Can you show me historical cases where a human did this task, so we can build an evaluation set?
- Can you tell me what this capability explicitly will not do?
- Can you tell me what it costs per task, and how you would reduce that cost?
What the first ninety days look like
The same shape on every capability. The content differs; the checkpoints do not.
| By day 30 | By day 60 | By day 90 |
|---|---|---|
| Baseline captured on your metric | Workflow running with a named approver | Metric moved, against the baseline you did not choose |
| Scope boundary documented — including what it may not do | Evaluation harness re-running on every change | Decision log reviewed by your risk function |
| Historical cases assembled for testing | Training delivered, completion in the LMS | Next workflow scoped from what we learned |
| Data lineage for every source the agent reads | Cost per task measured and modelled | Evidence pack complete and accepted |
Where this capability lands, by industry
Depth is our own 0–3 assessment of how strongly this capability matters in that sector today.
| Financial services | Manufacturing | Healthcare | Telecoms & tech | Public sector | Professional services | |
|---|---|---|---|---|---|---|
| Workflow documentation | ||||||
| Data readiness | ||||||
| Regulatory intensity | ||||||
| Value at stake | ||||||
| Learning dependency | ||||||
| Time to first workflow |
What this page answers
Every topic below is covered in the sections above, and each links to the page that goes deeper on it.
Bring a task your team dreads. We will watch it get finished.
Ninety minutes on your own task, on your own machine, with your own permissions. You will see the boundary conditions we have described above in practice — including the ones that will stop it.