Zoom’s Onyx Acquisition: A Practical Knowledge Search Plan for Small Businesses

ClayoCraft Insights · 24 October 2026 · 7 minute read

Zoom’s Onyx acquisition highlights a familiar problem: useful business knowledge is scattered across tools. Here is how smaller teams can make it searchable without opening every document to everyone.

The answer exists somewhere. One colleague remembers a client email, another has a newer policy in a shared folder, and the person answering the customer is looking at last year’s proposal. This is the problem AI knowledge search for small businesses should solve. A convincing answer is useful only when it draws on the right records and respects who is allowed to see them.

On 23 October 2026, Zoom announced its acquisition of Onyx, an open-source enterprise search and knowledge management platform. Zoom says Onyx connects organisational information to people and AI agents, and will remain a standalone, open-source, self-hostable platform. The announcement points towards the importance of business context; it does not establish that any particular small company needs a new platform or that an integration is immediately available on its subscription. Read Zoom’s announcement.

Start with a question your team repeatedly struggles to answer

Do not begin by connecting every application. Begin with a recurring question: which version of our cancellation policy applies to this account, what did we promise during onboarding, or where is the approved delivery checklist? Choose a question that wastes noticeable time and has a reasonably clear, verifiable answer. That gives your pilot a purpose beyond demonstrating a chatbot.

Interview the people doing the work, including someone who joined recently. Experienced staff often compensate for poor documentation through memory. A new colleague exposes the actual search experience. Ask them to find an answer without help and note the folders opened, contradictory versions encountered and questions escalated. Those observations become the baseline against which improvements can be assessed.

Build a small collection of trusted business knowledge

A search system cannot decide which of two conflicting policies your business intended to approve. Before indexing documents, identify an owner for each collection. Record the current version, approval date and review date. Archive superseded copies in a way that keeps required history accessible while making their status unambiguous. Correct the source instead of adding instructions that ask AI to ignore obvious disorder.

A useful first collection might contain current service descriptions, a customer onboarding checklist, approved support answers and internal escalation routes. Leave contracts, payroll records and personal customer information outside the initial scope unless they are necessary and properly controlled. Smaller scope makes permission testing, answer checking and ongoing maintenance more manageable for a team with limited capacity.

Separate finding information from making a decision

Searching for the approved refund procedure is different from approving a refund. Finding a contractual clause is different from interpreting its legal effect. Keep the pilot focused on retrieving and summarising information, with clear links back to the original source. An employee should be able to inspect the passage, its date and its context before relying on the answer.

For every question, define what happens when the source is missing, outdated or contradictory. A useful system should say that it cannot establish an answer and direct the question to an owner. Requiring an answer every time rewards confident invention. In customer-facing work, a short delay for verification is often preferable to an immediate promise that the business cannot honour.

Permission checks belong in the acceptance criteria

Someone who cannot open a finance document directly should not gain its contents through a search summary. Test this explicitly using accounts with different roles. Include a former project member, an external collaborator and a new starter. Check both search results and generated answers, because hiding a document title while exposing its contents still defeats the access boundary.

Permission changes also need testing after the initial setup. Remove a person’s access to a test collection and verify that the search system reflects the change. Ask providers how long indexing and access updates take, how cached answers behave and how logs are retained. These are practical procurement questions, not a substitute for reviewing the provider’s actual terms and technical documentation.

Compare conventional search with AI-assisted retrieval

Sometimes a better folder structure and ordinary search are sufficient. If staff need an exact invoice number, a reliable database filter may outperform a conversational interface. AI becomes more interesting when questions use everyday language, relevant information spans several approved documents, or staff need help understanding unfamiliar terminology. The tool should fit the question rather than the other way around.

Run the same set of questions through the existing method and the proposed system. Include simple lookups, ambiguous questions, missing answers and deliberately outdated documents. Record answer quality, source relevance and the time taken to reach an approved response. Do not compare a polished demonstration against an unusually difficult manual task and call the difference a business-wide productivity gain.

Measure useful outcomes instead of counting prompts

In an illustrative test, twenty questions might take an average of twelve minutes each to answer manually. If assisted search takes five minutes, the apparent saving is 140 minutes across the sample. That is a hypothetical calculation, not a benchmark or a claimed client result. Include time spent checking sources, correcting mistakes and maintaining documents before deciding whether the pilot is worthwhile.

More importantly, inspect the failures. An answer that is fast but uses an expired price list may create more work than it saves. Track the proportion of answers supported by the correct current source, the number requiring escalation and any cases of inappropriate disclosure. Decide the acceptance threshold before testing, especially where errors can affect customers, money or sensitive records.

Keep the architecture proportionate to the business

Self-hosting can offer control, but it also brings responsibility for updates, backups, monitoring and support. A managed service may reduce some operational work while introducing subscription costs and provider dependencies. Neither choice automatically makes a system private, compliant or cost-effective. Compare the complete workload and the capabilities your team can realistically maintain after the project is handed over.

For companies serving North America, the UK and Europe, include the location of users, customers and data processing in the review. These regions have different legal frameworks and contractual expectations. Ask a qualified adviser about the obligations relevant to your situation. A product’s availability in a region does not by itself settle questions about international transfers, retention or customer permission.

Give somebody responsibility for answers after launch

Knowledge management is an operating task. Name the person who reviews stale content, approves new collections and investigates poor answers. Give employees a simple way to report a source problem from the answer they are reading. A monthly meeting is not enough if a pricing change can make a frequently used document wrong tomorrow.

Maintain a short change log containing the collection changed, the reason, the approver and any follow-up tests. If you replace a search provider, retain your original documents, ownership records and test questions in portable formats. The most valuable asset is the organised business knowledge and the team’s ability to judge answers, not the configuration of one interface.

A realistic first month

Spend the first week choosing one knowledge problem and gathering baseline examples. Use the second to clean the source collection and confirm permissions. During the third, test retrieval with a small group. In the fourth, compare the results, document the failure cases and decide whether to expand, revise or stop. These are suggested stages, not a promise that every implementation can be completed in four weeks.

ClayoCraft’s AI & Business Efficiency support connects technology choices with practical workflow design. Where the underlying issue is unclear ownership, explore our Operations & Business Systems services. A useful starting conversation identifies one recurring question, the records behind it and what a reliable answer would change for your team.

Source: Zoom announcement, 23 October 2026. Editorial review: 24 October 2026. This is independent business analysis. Examples and suggested processes are ClayoCraft guidance, not results claimed by the source organisation. Publication dates are verified; exact publication times were not available for every source.