"Where is the latest version of that contract?" "Who decided this, and in which Slack thread?" Every company asks these questions dozens of times a day, and the answers sit in four or five different systems. Since late 2025 the big AI assistants have moved into this gap: Anthropic, OpenAI and Mistral all let their chat apps search a company's own documents, mail and chats. This guide compares what they do, where a specialist such as Glean still has the edge, and what the survey data says about the size of the problem.
The problem: company knowledge is scattered
Two surveys frame the problem. Coveo's 2025 employee experience report, based on 4,000 employees of companies with more than 5,000 staff in the US and UK, found that people use about four systems a day just to find information (five or six in technical roles), that a quarter do not know where to look, and that 42% of the content they come across is not relevant to their role. Slite's smaller survey of just over 100 knowledge workers, published in January 2026 and half of them in Europe, puts the time cost at 3.2 hours per week, or about 166 hours per person per year.
Source: Slite Enterprise Search Survey Report, January 2026 (100+ knowledge workers, ~50% in Europe). Vendor survey; small sample.
Lost time is the headline, but the second-order effects matter too: delayed answers to customers, duplicate work because nobody found the earlier version, and decisions made on outdated documents. Slite also found that only 27% of organisations had any enterprise search tool, and that security was the most common reason for holding back.
Old tools are not trusted, and neither is AI yet
Coveo's respondents rated intranets (31%) and classic enterprise search (24%) as ineffective more often than generative AI (15%). At the same time, trust in AI answers is falling, not rising: the share of employees who fact-check AI output climbed from 36% in 2024 to 42% in 2025, and only 17% fully trust their company-approved AI tools. Frustration with unhelpful work tools rose from 28% in 2022 to 40%.
Source: Coveo 2025 Employee Experience Relevance Report (4,000 employees, US and UK, companies with 5,000+ staff), as summarised by SiliconANGLE. Earlier values are from 2022 and 2024 editions of the same report.
That is the real design brief for AI search: answers have to come with citations a person can click, and the system has to respect who is allowed to see what. Both are now standard in the products below.
What Anthropic, OpenAI and Mistral now offer
| Product | Plans | How it works | Notable points |
|---|---|---|---|
| Claude enterprise search (Anthropic) | Team, Enterprise | A pre-built "Ask Your Org" project queries connected tools live through MCP connectors: one for documents, one for chat, optionally email | Launched Oct 2025 with a Microsoft 365 connector. Anthropic says connected data is not indexed to answer queries; each user signs in with their own account |
| ChatGPT company knowledge (OpenAI) | Business, Enterprise, Edu | A GPT-5 variant trained to combine several sources; answers show which sources and snippets were used | Launched 23 Oct 2025 with Slack, SharePoint, Google Drive, Gmail, GitHub and HubSpot. Must be switched on per chat; web search is paused while it is on |
| Le Chat Enterprise (Mistral) | Team, Enterprise | Enterprise search across Google Drive, SharePoint, OneDrive, Gmail and Calendar, plus document libraries and no-code agents | Can run self-hosted, in a private cloud or hosted by Mistral. A community-maintained list counted over 60 connectors after a June 2026 update |
| Glean (specialist) | Enterprise contracts | Indexes company systems into its own knowledge graph and lets you pick the model (Bedrock, Vertex AI, Azure OpenAI) | Initial contracts typically $100K–500K a year (Sacra); aimed at large organisations |
The key technical difference is live querying versus indexing. Claude and ChatGPT mostly ask each connected system at question time, using the user's own permissions. That is easy to set up and keeps data where it is, but it depends on the search quality of each source system and gets slower with many sources. Glean and classic search platforms build an index first, which costs more and takes longer to deploy, but can rank across all sources at once. Our posts on RAG and MCP explain both building blocks.
Does the specialist still win?
The only head-to-head test we found comes from Glean itself, published in February 2026: about 280 internal queries, graded blind by four people on correctness. When graders had a preference, Glean's answer won about 1.9 times as often as ChatGPT company knowledge and 1.6 times as often as Claude enterprise search, with the same underlying model on both sides. Glean did not measure speed or cost, and did not test Microsoft Copilot or Google Gemini.
Source: Glean, "Enterprise search evaluation 2026" (Feb 2026). ~280 queries from Glean's internal deployment; vendor-run test of its own product. Treat as indicative only.
Read it as a vendor result, but the direction makes sense: an index built for search will usually beat on-the-fly queries for complex, multi-source questions. The market agrees that the category is real. Glean passed $100M in annual recurring revenue in early 2025, $200M in December 2025 and an estimated $300M in May 2026.
Sources: Sacra research and company announcements (2025–2026). May 2026 is a Sacra estimate.
What the lost time is worth
Take the conservative Slite figure of 3.2 hours per week and an illustrative fully loaded cost of €30 per hour. If AI search recovers only a quarter of that time, the numbers already justify a pilot for most teams above ten people:
Source: AI Workshop calculation. 3.2 h/week × 52 weeks × €30/h; recovery assumption of 25% is illustrative, not measured. Licence costs not included.
How to roll it out without surprises
- Clean up permissions first. AI search respects existing access rights, so a folder shared with "everyone" years ago becomes searchable for everyone. Review SharePoint and Drive sharing before you connect them.
- Start with three sources. Documents, chat and one business system (CRM or tickets) cover most questions. More connectors make answers slower and noisier.
- Build a test set of 30–50 real questions with known answers, in Croatian, German and English, and score each candidate on it. Our guide to AI evaluation shows how.
- Teach people to click the citation. 42% already fact-check AI answers; make that the habit, not the exception.
- Decide where data may go. If documents cannot leave the EU or your own servers, Mistral's self-hosted option or an open-weight stack is the way; see our post on sovereign AI.
What this means for Croatian and DACH companies
Many companies in the region already pay for Microsoft 365 and either ChatGPT or Claude licences, so the cheapest win is to switch on the search features they have and connect them properly. Two points need extra care: GDPR and works-council rules in Germany and Austria, because searching mail and chat touches employee data, and language quality, because most vendor tests are run in English. A two-week pilot with one department, a Croatian or German test set and a clear permission review answers both questions before you commit.
Want your team to find answers in seconds?
AI Workshop sets up and tests enterprise AI search with Claude, ChatGPT or Mistral: permission review, connector setup, a test set from your own questions and a measured pilot with one team.
Talk to an AI consultant