Which everyday tasks should an internal assistant simplify?
Repeated information searches and routine preparation. A new employee needs to request access. A project manager needs the agreed constraints of a contract. A technician needs current instructions for a device. Each situation requires a specific source and a next step.
Start with real employees' questions. Ask them to describe recent occasions when they had to interrupt someone or search several tools. Also establish whether they merely found information or could use it to finish the task.
The assistant does not need to cover the whole company immediately. One well-designed onboarding or service process can demonstrate more value than hundreds of uploaded documents without a clear purpose.
What do international experiences show about workplace context?
In its Zoom case study, Anthropic cites a Morning Consult survey commissioned by Zoom: 82% of leaders and 72% of employees reported struggling to find enough time between meetings and messages; 35% of leaders spent at least three hours a day on meetings and email. These are respondents' reports from a specific survey, not measured savings after AI deployment. Zoom case study.
“We wanted to provide room for experimentation.” — Samantha Garrett from Canva's IT AI team, in the Canva case study.
Canva describes introducing AI into employees' work. For this design, the useful point is being able to test specific uses. Providing access to a tool alone does not prove that a company has eliminated repeated searches or improved decision quality.
Where should the assistant live so people use it?
As close as possible to where questions arise. That may be an internal portal, team communication tool or existing service platform. The choice depends on working habits and permissions, not just an attractive interface.
| Situation | Useful output | Required context |
|---|---|---|
| Employee onboarding | Correct process and access request | Role, team, approved policies |
| Meeting preparation | Briefing and open questions | Project, history and current status |
| Internal IT issue | Instructions or a created ticket | Device, permissions and service rules |
| Document work | Draft using the company template | Current version and specific brief |
Different roles need different functions. An operations employee may value concise mobile instructions; an analyst needs source links and document comparisons. Test real work situations, rather than just asking whether people like the chat.
How do you prevent inappropriate sharing of internal information?
Permissions must apply during retrieval. Finding every document and asking the model not to answer from sensitive parts is insufficient. Even a private document's title or short summary may not be intended for all employees.
In the 2026 Atlassian Community material discusses Rovo and the user's available context. For a custom solution, verify the specific connector and execution mode: an action under an automation account may have a different scope from a signed-in person's query.
Test ordinary employees, managers and users whose access has been revoked. Permission changes must affect retrieval and existing indexes. Chat history also needs its own retention and access rules.
What if documents are out of date?
The assistant should show its source and validity. Two conflicting instructions are not an invitation for a model to choose whichever sounds more convincing. The company needs an information owner and a way to resolve conflicts.
In the August 2026 OpenAI Community discussion raises the issue of connected context that is not current or useful enough. This is an individual experience. It suggests a simple pilot test: update an internal policy and check when the assistant stops using the old version.
A further thread in the XWiki community on Reddit explores preparing internal knowledge. Because it comes from a vendor's community, we do not treat it as independent research. Questions of ownership and access nevertheless belong in every pilot brief.
How do you add actions while keeping the experience simple?
A natural conversation can collect information and prepare an access request, for example. The action itself must follow the relevant system's rules. Employees must not gain greater permissions simply by phrasing a request well.
The agent should distinguish a prepared draft, a submitted request and approved access. If it has only created a ticket, it must not claim the problem is solved. Users need a case number and an explanation of what happens next.
Choose automated steps based on repeatability. Filling out routine requests may be suitable early on. Changing employment conditions, granting sensitive data access or modifying production systems requires a more precisely defined process and different tools.
How can you measure whether the assistant really helps employees?
Choose recurring tasks and compare time to correct completion. Include finding sources and making corrections. Message counts or active users show adoption, not automatically productivity.
Illustrative example: thirty people each save ten minutes of searching twice a week. Over four weeks, that is forty hours of capacity. If someone must continually fix documentation and check answers, some of that capacity goes into administration. That need not undermine the project, but the entire process must be counted.
The first pilot needs a knowledge owner, a set of common and difficult questions and a clear list of available actions. If disorganised information is the biggest issue, begin with the related design of an AI company knowledge base.
Frequently asked questions
Would a standard paid AI chat be enough for employees?
It may be enough for general writing and analysis. An internal assistant adds company sources, permissions and connections to specific processes. It makes sense where a general tool cannot finish the required work.
Will everyone see every document?
They should not. Search and subsequent actions should use the individual user's permissions. A shared technical account without role separation could expose information that employees cannot access in the original system.
Can the assistant create internal requests directly?
Yes, once the relevant system is connected. It needs required fields, permissions and confirmation of creation. A pilot can begin with drafts and gradually automate well-defined steps.
Research and solution design: Tanduva with AI assistance. External case studies are identified; illustrative examples are not measured results from our clients. Editorial methodology and corrections.