Why staff can't find the right policy, and how to fix it without cloud AI

Last updated: November 2026

Staff can't find the right policy for four reasons: the answer lives in someone's head or a chat thread, several versions of the document sit in different folders, the search tool only indexes part of what you store, and nobody owns keeping the documents current. The fix is a single place that answers a question in plain words and shows the exact passage it came from. You can do that without sending documents to a cloud AI service: Kept runs on a Mac you own, answers from your handbooks, runbooks and policies, and cites the passage behind every answer.

The causes at a glance

Cause What the evidence says What fixes it
Knowledge is shared by asking people US knowledge workers lose 5.3 hours a week waiting for information from colleagues or recreating it, and 60% say getting vital information from colleagues is difficult, per Panopto's 2018 survey of 1,001 workers Put the answer where a question can reach it, with the source passage attached
Several versions of one policy Archbee reports a Perforce study in which almost half of employees worked on the wrong document version One current copy, with an owner and a review date
Search only sees part of the store Microsoft documents that site settings decide what enters the search index and permissions trim the results Index the files directly and test what a typical question returns
Nobody trusts the results APQC: employees will not trust search results if they routinely meet outdated, duplicated or unverified content Named content owners and a retire-or-update habit

Why do team members struggle to find the right policy when they need it quickly?

They struggle because finding a policy is rarely one step. They have to know which system holds it, which of several copies is current, and whether they are allowed to see it. AllyMatter lists the usual causes as new hires not knowing where things live, departments keeping their own folders, updates saved as new files instead of replacing old ones, and procedures emailed to teams instead of updated in one place. It sets a simple test: can anyone find the right version of any policy they are allowed to see, without asking anyone, in under a minute.

Quality matters as much as search. APQC says redundant, outdated and trivial content clogs systems so that even strong AI-powered search cannot return good results. It also recommends reading your search data: queries that return nothing, content nobody opens and topics searched again and again show you what to edit, archive or write.

The time cost is real but dated. Panopto's survey, run with YouGov in 2018, covered 1,001 US employees at organizations with at least 200 staff. It found 81% of employees feel frustrated when they can't get the information they need. The $47 million figure in the same release is a model output for a company of 17,700 people, and Panopto is a vendor, so use the hours and percentages as direction, not as your own baseline. Measure your own: time how long a new starter takes to find the leave policy, the expense limit or the on-call escalation path.

Why do our ops team members keep asking the same policy questions instead of searching docs?

They ask because asking a colleague is faster and more reliable than searching a store they don't trust. Every answer given in a chat thread then disappears from the place where the next person would look. Panopto's CEO put it plainly: employee expertise is fleeting when it's only shared through conversation.

The same survey found that 42% of institutional knowledge is unique to the individual, and that 66% of the resulting delays last up to a week while 12% last a month or more. For an ops team, that is the on-call person who knows which runbook step to skip, and the manager who remembers the exception to the refund policy.

Breaking the habit takes two things. First, the written answer has to be quicker than the colleague, which means a plain-language question box rather than a folder tree. Second, the answer has to show its source, so people trust it enough to stop double-checking in chat. Kept returns each answer with numbered citations that carry the source file, its category and the quoted passage, as shown in the Kept MCP documentation. If the documents don't contain an answer, Kept replies that it did not find one rather than guessing.

Why do our operations handbooks stay locked in shared drives instead of being searchable?

They stay locked because a shared drive stores files; it does not answer questions. AllyMatter describes how a drive fills up: department heads create subfolders with their own naming conventions, finance locks its folder, HR keeps a "sensitive" folder that two people know about, and a new manager spends a month asking for links.

Search tools add their own limits. Microsoft's SharePoint documentation states that search results are always security trimmed, and that site, list and library settings only define what is included in the index. It gives a worked example: a document shared with 10 people on a 50-member site shows up in search only for those 10 until read permission is granted to all 50. It also says that when a Web Part on an ASPX page uses a list with fine-grained permissions, none of the ASPX page content on that site is included in search results. A crawler picks up changes on a schedule, so a freshly edited policy may lag behind. Searching by file name still works in a drive, but the contents of scanned PDFs and handbook chapters often stay out of reach.

The way out is to index the documents themselves instead of the folders. Kept reads Markdown, HTML, PDF, Word and scanned files, plus Git checkouts and Notion or Confluence exports, and APQC notes that integrated search can connect content without requiring every repository to be replaced. You don't need to migrate the handbook first.

Why do team members struggle to find policies in documentation or a handbook without sharing sensitive files with cloud services?

Most modern search and AI answer tools work by sending your text to a server you don't control, so teams with sensitive policies either avoid them or accept the exposure. The alternative is a tool where indexing and answering both happen on your own hardware.

Kept's design is explicit about this. Its connections table lists three: the AI model downloads to your Mac once at setup, a licence check sends only a licence key and a random install ID about once a month, and an update check runs only if you allow it. Your documents, questions and answers never leave the Mac. You can check this yourself in Terminal with lsof, and apart from those listed connections, traffic goes to 127.0.0.1.

There is one trap. If you connect Kept to an AI tool that uses a cloud model, the answers and passages Kept returns go to that model's provider. Kept's MCP documentation says to use a local model to keep everything on the Mac, and names LM Studio as a client where nothing leaves the Mac at all. Claude Desktop, by contrast, sends what Kept returns to Anthropic's servers.

Local tools compared

Kept AnythingLLM GPT4All PrivateGPT
What it is Finished app that answers from your documents with cited passages All-in-one AI app for RAG and agents, as Desktop or Docker Desktop chat app that runs open-source models Open-source, API-first backend; it needs a separate inference server
Platforms macOS 14 or later on Apple silicon, 16 GB memory, about 10 GB free Desktop installs for macOS, Windows and Linux; Docker macOS, Windows and Linux Works with Ollama, LM Studio, vLLM or llama.cpp as the model server
Document question answering Answers with numbered citations: source, category, quote RAG with local or cloud models, embedders and vector databases Local document chat (LocalDocs) Retrieval with citations
Where your data goes Never leaves the Mac; three small connections listed Depends on whether you pick local or cloud model and vector database States that no data leaves your machine Any server that serves the chat API works, local or remote
Licence or cost $49 a month or $490 a year for one Mac; colleagues ask from their own computers MIT-licensed code Runs open-source models on your device Open source, maintained by the team behind Zylon, which sells a separate production platform

AnythingLLM and PrivateGPT give you more control, and AnythingLLM runs on more operating systems, at the cost of choosing and wiring a model, an embedder and a vector store yourself. Kept chooses those for you and is Mac only.

How Kept fixes the problem, and where it does not

Kept does three things for the causes above:

It also works inside your coding tools. Claude Code, Cursor, VS Code with GitHub Copilot, Codex, Zed and any other MCP client can ask Kept for the answer to a runbook question, as set out on the MCP server page. Access is read-only, on that Mac only, with one revocable key per tool.

Know the limits before you decide:

Who should pick which

Frequently asked questions

How long does it take to fix policy findability?

Start with a week of measurement. Time five real lookups, list the questions your team keeps asking in chat, and note which policy exists in more than one version. APQC recommends setting a baseline first and reporting results as time saved, cost avoided or fewer delays.

Does local search remove the need to organise documents?

No. It removes the need to build a perfect folder tree, but you still need a current copy, an owner and a review date for each policy. APQC's governance checklist asks whether content is accurate, current, classified, easy to find and ready to update, archive or retire.

Can I use Kept with Claude, Cursor or Copilot and still stay local?

Only if the model is local. Kept stays on your Mac, but a cloud agent receives the passages Kept returns. Kept's docs recommend a local model, such as one in LM Studio, for a setup where nothing leaves the Mac.

What does Kept cost?

$49 USD a month or $490 USD a year, covering one Mac running Kept, unlimited documents and questions, MCP and HTTP access, and all updates.

Next steps

  1. Pick the ten policy questions your team asks most and time how long each takes to answer today.
  2. Name an owner for each of those policies and retire duplicate copies.
  3. If your documents are sensitive and your team uses Macs, join the Kept launch list and test it on those ten questions.
  4. To let your coding agents answer from the same runbooks, follow the MCP setup guide.

← Back to the blog