Slack-native RAG bot trained on your docs. Source citations on every answer, permissions enforced at the document level, freshness monitored. Monthly retainer delivery.
- Analyze
- Automate
- Monitor
monthly retainer
Who this is for
Ops or IT lead at a 100+ person company where Notion and Confluence search return too much noise, and your engineers or HR team spend real hours each week fielding questions that are already answered somewhere in the docs. The internal knowledge base AI bot exists specifically for that problem.
The pain today
- Notion/Confluence search returning irrelevant results for specific process questions
- Engineers and HR answering the same Slack questions day after day
- New hires taking weeks or months to navigate documentation
- Critical policies known only to two or three people
- No way to tell if a doc is current or quietly outdated
The outcome you get
- Slack-native bot answering questions from your docs in seconds
- Source citations on every answer so employees can verify
- Permissions enforced — doc-level ACLs respected in every response
- Freshness signals — stale docs flagged in answers before they mislead
- Question analytics that reveal gaps in your documentation
Why keyword search fails and what RAG does differently
Notion and Confluence have keyword search. It works fine for 'find the Q3 budget doc.' It breaks the moment someone asks 'what's our policy on expensing home office equipment for contractors.' That's a reasoning question, not a lookup.
Retrieval-Augmented Generation pulls the relevant chunks from across multiple docs, feeds them to a language model with the user's question, and returns a coherent answer with citations. The bot isn't guessing. It's reading the actual source material your team wrote, condensing it, and telling the employee exactly where to look if they want to verify.
For an internal knowledge base AI bot, the pipeline matters: ingest quality (how well the source docs are chunked and indexed), retrieval accuracy (which chunks get surfaced for a given question), and generation quality (whether the model answers from the retrieved content rather than hallucinating). I tune each layer separately. Most off-the-shelf tools treat it as a black box.
Slack-native bot answering questions from your docs in seconds
Slack and Teams integration patterns
I build with two primary deployment targets. Slack: a bot user with slash-command support (/askbot) and @mention in any channel. Threaded responses keep the channel clean. DMs work for sensitive questions. Microsoft Teams: equivalent bot via the Teams Bot Framework, @mention in channels or direct chat.
Both deployments use SSO sign-in so the bot knows who is asking — that's mandatory for permissions to work. Conversation context persists within a thread, so follow-up questions resolve naturally without the user re-stating the full context. No separate web app required. Employees never leave the tool they already use.
I start with Slack as the primary target because it covers the majority of customers I work with. Teams is a standard second deployment for organizations that are Microsoft-first.
30+: Active users.
Permissions — why this is not optional
Not every employee should see every document. HR files belong to HR. Engineering runbooks aren't for finance. Executive compensation docs aren't for interns.
If a knowledge bot ignores those boundaries, the first interesting question from the wrong person becomes a data incident. I handle this by syncing document-level permissions from the source system (Notion, Confluence, Google Drive, SharePoint) into the retrieval index, tagged per document. Every query filters the retrieval pool to docs the asking user has access to, verified against their SSO identity.
This gets tested explicitly during QA — I probe the bot with questions that should fail for specific user roles and confirm they fail cleanly. An audit log records every question, who asked it, and which documents were accessed. That log satisfies most internal compliance reviews without additional work.
Freshness, reindexing, and onboarding acceleration
A policy doc last updated 18 months ago may be wrong. The bot needs to signal that rather than confidently cite stale guidance.
I ingest documents with their last-modified timestamps and surface that context in answers: 'Based on a document last updated 4 months ago.' Retrieval weights fresher documents when multiple sources match. Docs older than 12 months get flagged for review in the monthly report.
Reindexing runs on a schedule matched to the change rate of each source: daily for engineering runbooks, weekly for company policies, on-demand when someone pushes a significant update. The result is a bot that employees can actually trust over time, not just in week one.
One concrete use case that shows the compounding value: new hire onboarding. A standard 30-day ramp involves dozens of questions that interrupt senior colleagues. With the bot fielding those questions, new hires get answers instantly and senior engineers stay focused. The interruption cost is real — it just rarely shows up on any dashboard.
Custom build vs. an off-the-shelf platform
Most SaaS knowledge bot platforms charge $500–2,000/mo and handle simple ingestion well. The problems appear when your permissions model is non-standard, when you need to combine three different doc sources with different auth systems, or when the bot's answer quality degrades because you can't tune the retrieval layer.
A custom build costs more upfront but the configuration is yours. The indexing pipeline, the chunking strategy, the prompt engineering, the model routing — all tunable. When a doc source changes format or an auth system updates, you're not waiting for a vendor to ship a fix.
My $3,999/mo retainer covers the initial build (4–5 weeks for a core set of doc sources) and ongoing iteration: new doc sources added, retrieval quality refined, answer accuracy monitored. LLM API costs run separately at roughly $100–1,000/mo depending on question volume. That range is wide because 100 employees asking 5 questions a day costs about $300/mo in LLM fees, while 1,000 employees scales toward $1,500–2,500/mo. I optimize model routing — simpler questions get cheaper models — to keep that number predictable.
What the question logs reveal
The bot generates a side benefit that most customers don't anticipate: it shows what employees are actually asking.
Questions the bot can't answer well point to doc gaps. Frequent questions it answers correctly but that still get asked daily point to onboarding or training gaps. Time-sensitive clusters reveal operational patterns — spikes in questions about Q4 planning in October, or sudden volume around a system change.
I deliver a monthly report of top questions, answer quality scores, and suggested doc updates. Over the first 90 days, customers typically discover three to five categories of missing or outdated documentation they didn't know existed. That intelligence is worth building the bot for, independent of the time savings.
The Instill product I run — an AI skills library with 1,000+ skills saved across 45+ projects — uses the same retrieval and quality evaluation patterns. Operating it through real user iteration has shown me where these systems fail in production: retrieval drift when docs accumulate without reindexing, permission edge cases in multi-workspace setups, and user trust eroding when the bot cites a visibly outdated source. I know these failure modes because I've debugged them, not read about them.
Recent proof
A comparable engagement, delivered and documented.
An AI knowledge base your whole team uses via MCP
A personal library for Skills, Agents, and Rules, built once, used across Claude, Cursor, and any MCP-compatible AI tool. Save your best workflows once. Run them anywhere.
Read the case studyKeep reading
Frequently asked questions
The questions prospects ask before they book.
Notion, Confluence, Google Docs, SharePoint, Dropbox Paper, GitHub Wiki, GitBook, and internal HTTP APIs. Custom sources like PDFs on a shared drive or a legacy internal wiki are supported with additional ingestion pipeline work. I scope which sources to ingest in week one based on where your actual knowledge lives — not where it's supposed to live.
Document-level permissions from the source system are synced into the retrieval index. The bot can only return content from documents the asking user has access to, verified against their SSO identity. Additional redaction rules — auto-redact patterns like employee IDs or financial data — can be layered in per your compliance requirements. Every query is logged with user identity and documents accessed.
Four to five weeks to ingest your core doc sources, wire Slack or Teams, and ship with permissions working. The first version is intentionally scoped — two or three primary doc sources, core question types covered. Additional sources and refinements continue through the retainer. Trying to ingest everything in week one produces poor retrieval quality; phased ingestion gets the bot useful faster.
Roughly $300–500/mo for 100 employees asking around five questions a day. At 500 employees the range is $800–1,500/mo, depending on question complexity and how much is served from cached results. I route simpler questions to cheaper models to keep costs predictable. LLM fees are billed directly to your account, separate from the $3,999/mo retainer.
Yes. Code-aware sources — GitHub, GitBook, internal runbooks — integrate the same way as other doc sources. The bot can answer 'where is the deployment pipeline defined' or 'what's the escalation path for a database alert' from engineering docs. This is particularly useful for onboarding engineers who shouldn't need to interrupt a senior colleague for questions that already have a written answer.
Reindexing runs on a schedule matched to how often each source changes — daily for runbooks, weekly for policies. Every answer surfaces the source document's last-modified date so employees can judge freshness. Docs older than 12 months get flagged in the monthly report. The goal is a bot employees trust over time, not just during the first demo.
For simple setups with one or two doc sources and standard permissions, a SaaS platform at $500–1,000/mo can work fine. When your permissions model is non-standard, when you're combining doc sources with different auth systems, or when retrieval quality matters enough to tune, a custom build gives you control the platforms don't. The retainer covers that tuning continuously, not just at launch.