- Home
- AI agents
- Enterprise AI agent examples
- Knowledge search
Plan · Enterprise AI agent examples
Knowledge search AI agents: examples, the permissions behind them and where to start
A knowledge search agent answers an employee’s question from company content, cites the documents it used and shows only what that person is allowed to open. It is the retrieval layer most other agents stand on, and the published examples measure one thing above all: how long it takes a person to find the answer.
In this guide: 83 pages
Key takeaways
- Time to find is the measure that repeats. Wells Fargo reports 10 minutes to 30 seconds for procedure lookups, Textron Aviation up to 20 minutes to 1 to 2 minutes for maintenance troubleshooting, Toyota 5 to 6 hours to 2 to 3 minutes for robot diagnosis, and our own Rockwell Automation deployment 12 minutes to under 4.
- The agent returns only what the signed-in person could already open, provided permissions are enforced at query time and kept in sync, which means oversharing in SharePoint or a file share becomes visible on day one. Clean it before launch.
- Hours-saved figures need a method. Capita’s 9,000 hours a month is credited to Microsoft 365 Copilot overall, and DBS’s 10% of work hours is an "up to" figure from Glean.
- Measure retrieval and answers separately. A wrong answer with the right documents is a generation problem; a wrong answer with the wrong documents is a retrieval problem, and the fixes differ: prompt and model on one side, chunking, index and permissions on the other.
What do knowledge search agents do?
They find the passage, cite it and answer in the user’s language, across intranet, document libraries, ticketing and line-of-business systems, under the user’s own permissions. The newer ones plan several searches for a hard question and can take a small action afterwards, such as opening a ticket.
Knowledge search tasks by level. "Answers" means question and answer only; "assists" means it drafts and a person acts; "acts" means it changes a record.
| Task | Level | What the agent touches | Seen in |
|---|---|---|---|
| Procedure and policy lookup | Answers | Internal procedures and policy libraries, with citations | Wells Fargo (1,700 procedures); Capita |
| Technical troubleshooting from manuals | Answers | Maintenance documentation and parts diagrams, in several languages | Textron Aviation (about 60,000 pages); Toyota GearPal |
| Multi-step research questions | Assists | Several searches planned and merged; full documents fetched when a chunk is not enough | Azure AI Search agentic retrieval; Bedrock managed knowledge bases |
| Action after the answer | Acts | Ticketing or HR systems, as the user | Coca-Cola Andina creates a ticket for HR (Microsoft customer story, 19 March 2026); DBS support agents on Glean assist with tickets |
| Multilingual answers | Answers | The same content, translated at query time | Capita; Textron Aviation |
| Employees building their own search agents | Answers | SharePoint and connector content the maker may see | Capita (more than 169 agents built by employees) |
Most of these agents answer rather than act, and the figures show why that is enough: the measured gain is the minutes a person no longer spends looking.
What have companies published about knowledge search agents?
Three organizations publish before-and-after times, two publish hours saved with caveats, and the rest publish scope. The Rockwell Automation figures are our own, from a deployment we delivered, and they are labelled that way.
Publicly reported knowledge search deployments, as published by each source and opened on 2 November 2026, plus QueryNow’s own Rockwell Automation deployment. No other company here is a QueryNow client.
| Organization | What the agent does | Published figure | Source and date | How to read it |
|---|---|---|---|---|
| Wells Fargo | A Microsoft Teams app with language models gives 35,000 bankers in 4,000 branches guidance on 1,700 internal procedures | 75% of searches go through the app; response times from 10 minutes to 30 seconds | Microsoft Cloud Blog roundup, 24 July 2025 | A one-line vendor roundup entry; the underlying model is not named |
| Textron Aviation | TAMI lets maintenance technicians ask questions in several languages across about 60,000 pages of documentation for more than 50 aircraft models | Troubleshooting lookups from up to 20 minutes to 1 to 2 minutes | Microsoft customer story, 20 March 2025 | Company-reported through a vendor story; answers only |
| Toyota Motor North America | GearPal gives technicians plain-language diagnostics, service history and repair guidance for robots and line assets; one of more than 50 agents in production | Diagnosis from 5 to 6 hours to 2 to 3 minutes; "at least 6 figures" of annual savings per line, shop and plant | LangChain blog, 24 August 2026 | Toyota staff quoted by the vendor; the multi-million-dollar savings per facility on the same page is a projection |
| Capita | AskMeAnything, a Copilot agent, finds policies and procedures across the SharePoint estate with real-time translation | 9,000 employee hours saved per month; 3,000 licences; more than 169 agents built by employees | Microsoft customer story, 24 June 2025 | The hours are credited to Copilot overall rather than to the search agent; no method given |
| DBS Bank | Glean enterprise search across hundreds of connectors, plus support and HR agents that staff can build on | Over 40,000 users; "up to 10%" of employees’ work hours freed | Glean customer story, undated, read 2 November 2026 | Vendor claim with no named person; an "up to" figure |
| Rockwell Automation (QueryNow deployment) | A new intranet and enterprise search for 28,000 employees: unified search over intranet, OneDrive, Teams, email and third-party systems, custom verticals and a dictionary of company terms | Time to find information from 12 minutes to under 4 | QueryNow case study | Our own figure from a search deployment: the layer an agent reads from |
EY Global Tax’s research agent in Copilot Studio reaches 21 million documents and publishes no outcome figure (Microsoft Cloud Blog, 24 July 2025), which is the honest form of a scope claim: it tells you what was indexed, and nothing about what changed.
What controls does a knowledge search agent need?
Permission-aware retrieval first, then the clean-up it exposes, then evaluation of the retrieval step on its own, and a weekly look at what the agent could not answer.
- Identity and permissions. The query runs under the signed-in user and the index enforces the source’s access lists, so the agent returns only what the person could open. Azure AI Search does this with the user’s Entra token and synchronized access lists (SharePoint and Purview label enforcement were preview on 2 November 2026); Microsoft 365 Copilot and its Retrieval API honour the user’s permissions on the Microsoft 365 index; Bedrock managed knowledge bases filter by the user you pass in and re-check access live for SharePoint, Confluence and other connectors, failing closed. AWS says plainly that its filtering "does not constitute true authorization": your application still authenticates the user.
- Oversharing. A permission-aware agent surfaces whatever is shared too widely, which in most tenants means sites open to "Everyone except external users". Run the SharePoint data access reports, Restricted Content Discovery and site access reviews before launch, and treat the first week of answers as an audit.
- Data access for actions. If the agent can also open a ticket or draft a reply, that tool runs as the user through the narrowest connector, and the read side stays read-only.
- Evaluation. Score retrieval with labelled relevance judgments (Foundry’s Document Retrieval evaluator reports NDCG and holes; Bedrock evaluations run "retrieve only" jobs) and score answers for groundedness and completeness separately. Build the set from real questions. At Rockwell the search needed custom verticals and a dictionary of company acronyms, which only usage data revealed.
- Monitoring. Watch the unanswered and low-confidence questions weekly, the time to find on a sampled set of questions, and which sources are cited. Permission sync lag matters too: a revoked right disappears from the index only after the next sync, inherited SharePoint changes in Azure AI Search need an explicit refresh, and only the supported Bedrock connectors re-check live, so set the sync interval from the sensitivity of the content.
Poisoned content is the risk specific to this agent. A few planted documents can steer answers, and a retrieved email can carry instructions, which is what CVE-2025-32711 in Microsoft 365 Copilot showed in June 2025. Input checks on retrieved text reduce the risk, a read-only answer path removes the actions an injection could trigger, and you still need controls on where data can leave, such as trusted URL lists for links and images, because several 2025 exploits leaked data through image URLs.
Where should you start?
Start with one content set people already search for and fail to find, measure how long finding takes today, and launch answers only, under the user’s permissions.
- Pick the set: procedures, maintenance manuals, HR policy. Textron chose maintenance documents, Wells Fargo chose procedures.
- Time ten common questions with ten people before anything changes. That number is your 12 minutes.
- Fix oversharing and stale content in that set. At Rockwell the content audit took more effort than the technical migration.
- Launch under the user’s identity, with citations shown, in the channel people use. Re-time the ten questions after a month.
- Add agentic retrieval for the hard questions, and a first action, only when the unanswered list is short and the citations are trusted.
How do Microsoft, AWS and Google implement it?
How each cloud builds permission-aware knowledge search for agents, with statuses as the vendor pages showed on 2 November 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Managed knowledge layer | Foundry IQ knowledge bases on Azure AI Search (GA 2 June 2026); one knowledge base shared by many agents | Amazon Bedrock managed knowledge bases with agentic retrieval (no preview label); Amazon Quick as the successor to Amazon Q Business | RAG Engine and Agent Search in Gemini Enterprise Agent Platform (Agent Search is the renamed Vertex AI Search, 22 April 2026) |
| Permission trimming | Azure AI Search document-level access with the user’s Entra token (access list and Purview label enforcement in preview); the Microsoft 365 Copilot Retrieval API returns permission-trimmed chunks | Allow and deny lists ingested per document, filtered by the user passed in, with live re-checks for SharePoint, OneDrive, Google Drive, Confluence, Box, Zendesk and Salesforce | Not verified in the research set; check each connector’s permission handling before launch |
| Agentic retrieval | Azure AI Search agentic retrieval GA for minimal, extractive retrieval (API 2026-04-01); LLM query planning and answer synthesis still preview | AgenticRetrieveStream on managed knowledge bases: planning, retrieval, full document expansion, answer with citations | RAG Engine; Agent Retrieval (the renamed Vector Search 2.0) |
| Context from the work day | Work IQ over Microsoft 365 mail, meetings and documents (CLI and MCP server GA 16 June 2026; preview as a Foundry knowledge source) | AgentCore Memory integration with agentic retrieval since 14 August 2026 | Gemini Enterprise connectors to Workspace, Microsoft 365, Salesforce and SAP |
| Retrieval evaluation | Foundry RAG evaluators: Document Retrieval (NDCG, holes), Groundedness, Relevance | Bedrock RAG evaluation (GA 20 March 2025): retrieve-only and retrieve-and-generate jobs, citation metrics | Gen AI evaluation service with adaptive rubrics (the GenAI Client SDK in preview) |
Glean and Microsoft 365 Copilot carried the DBS and Capita examples as SaaS. Wells Fargo, Textron and Toyota built on platforms (Microsoft Teams with Azure OpenAI, Azure OpenAI, LangChain), and Rockwell’s search runs on Microsoft Search in SharePoint.
Enterprise example · QueryNow
Rockwell Automation: enterprise search for 28,000 employees
QueryNow deployment. Rockwell Automation, an industrial automation company with 28,000 employees, replaced its legacy intranet with a modern SharePoint intranet and unified search.
The problem was the one every knowledge agent inherits: disconnected sites, search results that were often irrelevant or outdated, and decentralized publishing with inconsistent governance, so employees asked colleagues or gave up and used email.
We rebuilt the information architecture around how people work, then configured Microsoft Search with unified results across the intranet, OneDrive, Teams, email and third-party systems, custom search verticals for policies, people and projects, and a custom dictionary that translates company acronyms. The technical architecture covered the hub structure, the permission model and the integration points, so search returns only what the signed-in employee may open.
The measure was time to find information: 12 minutes to under 4. Two lessons from the Rockwell Automation project shape how we approach every knowledge search agent. Content curation took more effort than the technical migration, and search needed ongoing tuning from usage data rather than a one-time configuration.
- Owner
- A governance framework with named roles, publishing workflows and content standards for each hub site
- Identity
- Search runs as the signed-in employee under the SharePoint permission model, so results are trimmed to what the person may open
- Data access
- Unified search across intranet, OneDrive, Teams, email and third-party systems; read only
- Risk tier
- Lowest: answers only, from content the user already has rights to
- Approval
- Publishing workflows in the governance framework; a content audit separated high-value from outdated content before migration
- Evaluation
- User acceptance testing with representative users; time to find information measured before and after
- Monitoring
- Usage analytics, surveys and support tickets reviewed after launch, with the experience refined from usage data
- Retirement
- Outdated and duplicate content retired in the audit; selective migration rather than a lift of everything
Frequently asked questions
Is a knowledge search agent the same as RAG?
Mostly. Retrieval-augmented generation is the pattern: find passages, put them in the prompt, answer with citations. The agent adds planning (several searches for one question), permissions enforced at query time and, sometimes, an action afterwards. The guide’s knowledge topic covers agentic RAG against plain RAG.
Will the agent expose documents people should not see?
It shows what the signed-in person could already open, and nothing more, if retrieval is permission-aware. The exposure it causes is of content that was already overshared. The fix is the data access reports and access reviews before launch, and a sync interval that matches how fast rights change.
How do we know whether a bad answer is a retrieval problem or a model problem?
Score the two steps separately. Labelled relevance judgments (NDCG and similar) tell you whether the right documents came back; a groundedness check tells you whether the answer used them. In our experience most early failures are retrieval: missing verticals, acronyms the index does not know, stale duplicates.
What does "hours saved" mean in the vendor stories?
Usually a self-reported estimate with no method. Capita credits 9,000 hours a month to Copilot overall, and DBS reports "up to" 10% of work hours. Time to find on one kind of question, compared before and after, is the figure you can defend: Wells Fargo for procedure lookups, Textron for troubleshooting lookups, Toyota for diagnosis, and our Rockwell deployment for finding information. Run the same questions before and after in your own test.
Should the first version take actions?
No. Every measured example above answers or assists; none changes an operational system, and the gain is the minutes nobody spends looking. Add a first action, such as opening a ticket as the user, once the unanswered list is short and people trust the citations.
More in Enterprise AI agent examples
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- AI-powered success with more than 1,000 stories of customer transformation and innovation
- Textron Aviation enhances maintenance efficiency with Azure AI
- Toyota scales enterprise AI with Deep Agents and LangSmith
- Capita customer story: Copilot agents built by employees
- How DBS enables its staff to focus on meaningful work through AI integration
- Coca-Cola Andina customer story (Copilot Studio HR agent)
- Rockwell Automation: modern intranet and enterprise search for 28,000 employees
- Document-level access control in Azure AI Search
- Agentic retrieval in Azure AI Search
- What is Foundry IQ
- Microsoft 365 Copilot Retrieval API overview
- SharePoint Advanced Management features included with Copilot licenses
- Work IQ CLI
- RAG evaluators in Microsoft Foundry
- ACL-aware retrieval in Amazon Bedrock managed knowledge bases
- Agentic retrieval in Amazon Bedrock Knowledge Bases
- Amazon Bedrock document history
- Amazon Bedrock RAG evaluation is now generally available
- Amazon Q Business availability change
- Gemini Enterprise Agent Platform release notes
- Gen AI evaluation service overview
- CVE-2025-32711: Microsoft 365 Copilot information disclosure vulnerability
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner