For most organisations, the first wave of generative AI looked the same: employees opening a public chatbot in a browser tab, pasting in text, and hoping nothing sensitive went with it. Two years on, many companies are trying to replace that improvised approach with something more structured. The new goal is AI assistants that are connected to company knowledge, follow the organisation’s permission rules and can be managed centrally, rather than a scattered collection of personal tools.
A growing group of platforms is built around that idea. One of them is Dust, available at dust.tt, which lets companies create custom AI agents that draw on internal documents and workplace tools. Its rise reflects a broader question that operations, IT and knowledge-management teams are now asking: how do you give staff useful AI help without losing control of data, quality and accountability?
Why the “one chatbot for everything” model is fading
General-purpose chat assistants are good at writing, summarising and answering broad questions. They are less helpful when the answer depends on information that sits inside a company: the latest product specifications, a client’s contract history, an internal policy or a support ticket thread. Copying that information into a public tool by hand is slow and creates obvious risks.
As a result, interest has shifted towards assistants that connect directly to the places where knowledge already lives, such as document stores, wikis, chat tools and ticketing systems. Instead of asking employees to bring context to the AI, the AI is given governed access to the context. The promise is fewer repetitive questions, faster onboarding and less time spent searching across fragmented systems.
What platforms like Dust offer
According to an independent first-look assessment published by 1stPage Agency, Dust is aimed at knowledge-work teams that want assistants operating across internal data and tools with governance built in. The features highlighted include:
- Custom company agents that can be set up for specific tasks or departments.
- Connections to workplace knowledge, so agents can draw on internal content rather than only general training data.
- Model choice and agent instructions, allowing teams to decide which underlying AI model an agent uses and how it should behave.
- Team spaces and sharing, so useful agents can be reused rather than rebuilt by every employee.
- Enterprise permissions and governance, intended to keep access aligned with who is allowed to see what.
The assessment describes the platform as focused on a defined operational problem and as offering a structured workflow rather than an isolated AI feature, which it sees as a strength for teams trying to reduce repetitive work and messy handoffs.
The caution behind the enthusiasm
The same review is notably measured. It is framed as research-based rather than hands-on testing and does not assign a numerical score. Instead, its verdict is that the tool is worth a controlled trial, and it lists open questions that apply well beyond a single vendor:
- Public claims about productivity gains need independent validation.
- Results depend heavily on implementation quality and on how well organised a company’s data already is.
- Full commercial terms are not always published, and the platform’s pricing is sales-led, so seat, usage and model allowances need to be confirmed directly.
- Human accountability and governance remain necessary even when an assistant is connected to trusted sources.
Readers who want the full breakdown of strengths, limitations and fit can read the complete dust.tt review, which also outlines who the platform is and is not best suited for.
How companies are approaching evaluation
The cautious tone mirrors how many organisations now approach AI purchases. After a period of fast, enthusiastic adoption, buyers increasingly want evidence before committing to company-wide rollouts. Advisers who work with technology teams commonly recommend a few steps:
Start with a bounded problem. Rather than trying to transform every workflow at once, pick a specific, repetitive task, such as answering common internal questions, drafting first responses to support tickets or summarising long document sets.
Measure a baseline first. Record how long the task takes today, how often errors occur and how satisfied users are. Without a baseline, it is impossible to tell whether an assistant has actually helped.
Use real data in the pilot. Demonstrations with sample content rarely reveal the problems that appear when an assistant meets a company’s actual documents, with their gaps, duplicates and outdated versions.
Check permissions carefully. A connected assistant is only as safe as the access rules it follows. Testing whether users can see information they should not is a basic but essential step.
Keep a human in the loop. Outputs that affect customers, contracts or compliance still need review. The aim is to reduce effort, not to remove responsibility.
Data readiness is the hidden factor
One theme comes up repeatedly in discussions of enterprise AI: the quality of the answers depends on the quality of the underlying knowledge. If internal documentation is out of date, contradictory or scattered across dozens of folders, a connected assistant may simply surface that confusion faster. Many organisations discover during pilots that they need to tidy up their knowledge bases before AI tools can deliver consistent value.
That is not necessarily a drawback. Several teams report that preparing for an AI rollout forced long-overdue clean-ups of wikis, shared drives and process documents, which benefited employees whether or not they used the assistant.
Cost and contract questions
Because platforms in this category often price through sales conversations rather than public price lists, buyers are advised to clarify details early. Useful questions include how seats are counted, whether usage is capped, which AI models are included, how costs change as adoption grows, and what happens to company data if the contract ends. Comparing these answers across several vendors gives a clearer picture than headline features alone.
A more mature phase for workplace AI
The interest in governed assistants signals a shift in how businesses think about AI. The question is no longer whether employees will use it, but how to make that use reliable, secure and measurable. Platforms like Dust are part of that transition, and the measured way independent reviewers are assessing them reflects a market that has moved beyond novelty.
For companies considering the step, the practical advice is consistent: pick a clear use case, test with real data, measure the results honestly and scale only when the evidence supports it.
