Chat surfaces
Slack (Bolt: DMs, mentions, slash commands, Block Kit) and Microsoft Teams (Bot Framework: 1:1, channel mentions, Adaptive Cards).
Pick 1–2
For IT, HR & Ops teams at 50–500+ person companies
Your internal help desk, answered by AI,
in the Slack or Teams your team already lives in.
We build you a secure, company-owned AI help desk that answers employee IT, HR, and ops questions from your own policies, resolves routine requests, and files and routes the tickets it can't. Your team stops answering the same questions all day.
Grounded in your docsCited answersSSO + least-privilege permissionsUS-based, you own it
The problem
Your best people are stuck doing Tier-0 triage.
Industry benchmarks put roughly 35–55% of internal tickets in the same handful of questions: VPN resets, PTO policy, expense limits, "how do I request X," "what's the status of my ticket." Every one pulls someone off real work, and the backlog grows with every new hire.
- New hires wait hours, sometimes until tomorrow, for an answer they could've had in 10 seconds.
- Nobody can find the policy doc, so they ask in Slack again.
- Your ticket backlog grows with headcount, and another support hire is $70K+/year fully loaded.
At 250 employees, that repetitive volume quietly runs an estimated $37K–$55K a year
in loaded time, on tickets that never needed a human. Figures are illustrative benchmarks; your
audit quantifies the real number for your team.
How it works
Answers in seconds. Tickets routed in one step. Humans only when it matters.
Employees ask the way they already ask a coworker: in Slack, Teams, or an SSO-gated web portal.
1
Employees just ask
In Slack, Microsoft Teams, or a web portal. No new tool to learn: a DM, a mention, or a slash command.
2
It answers from your policies and cites the source
Grounded in your handbook, runbooks, and knowledge base. If the answer isn't documented, it says so instead of guessing, and offers to open a ticket.
3
It takes action, not just talk
Routine request? It resolves it or files a properly-tagged ticket in Jira Service Management, ServiceNow, or Freshservice: right queue, priority set, context attached.
4
It escalates cleanly
Sensitive, complex, or "I need a human"? It hands off to the right person with the full conversation. No repeating yourself.
5
You see exactly what it's doing
A monthly dashboard: tickets deflected, time-to-resolution, escalation rate, and where your docs need work.
Every interaction logged · permission checks before retrieval · sensitive categories hard-routed to a human
The payoff
What it's actually worth.
- Deflect 35–55% of repetitive tickets: the Tier-0/1 volume that never needed judgment.
- Seconds, not hours: a grounded first response, 24/7.
- Cleaner tickets: the ones that reach a human arrive complete, tagged, and routed.
- Less than one hire: covers the deflectable load of several support people for less than one fully-loaded salary.
- Your data stays yours: never used to train a model; US-region, tenant-isolated.
- Your team, upgraded: they do the work you actually hired them for.
Deflection ranges are industry benchmarks, not a guarantee. Your actual numbers come from your knowledge base,
ticket mix, and monthly tuning. A named client case study will appear here after the first build ships.
Integrations
It plugs into the stack you already run.
Base build includes one primary from each relevant group. Additional connectors are scoped add-ons.
ITSM / ticketing
Jira Service Management, ServiceNow, or Freshservice: create tickets, read status, route to the right queue with priority set.
Pick 1 primary
Identity / SSO
Okta, Microsoft Entra ID (Azure AD), or Google Workspace via OIDC/SAML: resolves the user, department, and entitlements on every request.
Required, pick 1
HRIS (read-only)
Workday, BambooHR, Rippling, Gusto, ADP, or HiBob for policy context, PTO balance, and org lookups, all permission-gated.
As needed
Knowledge sources
Confluence, Notion, SharePoint / OneDrive, Google Drive, your ITSM knowledge base, Guru, wikis, and PDF/DOCX exports.
Ingest what applies
AI layer
Anthropic Claude under enterprise terms (no training on your inputs). Reasoning for routing, a lighter model for classification, prompt caching for cost control.
Built in
Base build = 1–2 chat surfaces · 1 ITSM · 1 SSO · up to ~4 knowledge sources · extras scoped as add-ons
Built right
Secure, permission-aware, and yours to keep.
This is the section your security team reads. It's built to survive that review.
Least-privilege permissions
SSO-aware retrieval filtered by the caller's entitlements before anything reaches the model. Restricted HR or IT content never surfaces to someone who shouldn't see it.
Your data never trains a model
Enterprise API terms with no training on your inputs. US-region infrastructure, tenant-isolated: no shared indexes, no cross-client data. It's contractual.
Full audit trail
Every question, source accessed, permission decision, and action logged and exportable for compliance. Optional PII redaction and configurable retention.
Human-in-the-loop by design
Terminations, comp, investigations, and security incidents are hard-routed to a human. The AI never adjudicates sensitive categories.
Scoped, revocable access
Least-privilege service accounts and OAuth scopes: read-only where read-only suffices, write scoped to ticket creation. Credentials in a managed secrets store.
You own it
Code, prompts, configuration, and data live in your environment. No black box, no lock-in: a built-and-owned asset, maintained under the retainer.
US-only. Key's Touch serves US-based companies, with infrastructure deployed to US-region. NDA and DPA precede access to anything.
Investment
A one-time build, plus a partnership that keeps it sharp.
Scoped by chat surfaces, ITSM complexity, knowledge-source count, HRIS integration, and permission depth. Every build is a fixed quote after discovery.
Core
100–200 employees, 1 chat surface, 1 ITSM, up to ~3 knowledge sources, standard SSO, no HRIS. Core help desk, RAG, ticketing, escalation, and dashboard.
Multi-surface
200–350 employees, Slack and Teams or a web portal, 1 ITSM, ~4–5 knowledge sources, group-based permissions. Adds a second surface and richer permissions.
Enterprise
350–500+ employees, dual chat surface, HRIS read integration, complex ITSM routing, 6+ knowledge sources, strict security review. Adds HRIS, advanced routing, and the full security pack.
AI Performance Partnership: scoped after implementation. Continuous monitoring, model and prompt maintenance,
knowledge-base tuning, and a monthly performance report with a prioritized knowledge-gap list. Required for the first
6 months while the system tunes to your real questions; month-to-month after with 30 days' notice.
50% of build due at kickoff, 50% at go-live · typically 4–6 weeks to go-live · you own the code, prompts, config, and data
See it work
A routine question answered in seconds. A real request routed in one.
Exactly how it looks in the Slack or Teams channel your team already uses: a cited answer from your own docs, then a ticket filed and routed the moment something needs a human.
MC
Maya Chen9:41 AM
How many PTO days do I have left this year, and do unused days carry over?
AI
Help Desk AIApp9:41 AM
You have 11 of 20 PTO days remaining for 2026. Up to 5 unused days carry into Q1—anything beyond that is forfeited on Dec 31.
Source: Employee Handbook › Time Off PolicyDR
Devin Rao9:43 AM
My laptop still won't connect to the VPN after a restart, and I have a client call in an hour.
AI
Help Desk AIApp9:43 AM
That one needs a human. I've opened a ticket with your device details and error logs attached, and flagged it as time-sensitive.
Ticket #INC-1042 createdVPN connection failure — MacBook Pro
Escalated to the on-call IT engineer with the full thread—Devin won't have to repeat a thing.
FAQ
The questions IT and HR leaders actually ask.
Will it make things up?
No. It only answers from your documented content and cites every source. No documented answer means it says so and offers to open a ticket. It won't invent policy.
What about sensitive HR or security questions?
Hard-routed to a human by design. The AI never adjudicates terminations, compensation, investigations, or security incidents.
Our documentation is a mess. Does this still work?
Yes, and it helps fix it. Week one is a knowledge audit, and the monthly dashboard flags exactly which gaps to close. The system gets sharper as your docs improve.
Does it replace our ITSM (ServiceNow, Jira, or Freshservice)?
No. It sits on top and feeds your ITSM cleaner, better-routed tickets, so your queues only hold work that actually needs a human.
Will it replace my team?
No. It removes the repetitive Tier-0 grind so your team does the work you hired them for. It's a force multiplier, not a headcount cut.
How long to launch?
Typically 4–6 weeks from a signed agreement and system access to go-live.
What does it cost?
Every build is scoped to your headcount, chat surfaces, integrations, and knowledge sources after a short discovery. You get a fixed, phased quote with the payback math before you decide. The monthly AI Performance Partnership that keeps it accurate and improving is scoped after implementation.
Is our data safe?
Least-privilege, SSO-aware retrieval so restricted content never surfaces to the wrong person. Your data never trains a model, infrastructure is US-region and tenant-isolated, and every question, source, and action is logged for audit. You own the code, prompts, configuration, and data.
Book a demo
See it answer your questions.
A few quick questions so we can load a demo with policies that look like yours. Takes about 90 seconds.
Pick a time. That's it.
Choose a slot for a live, no-obligation demo, including the moment it can't answer and files the ticket for you.
15 minZoomYou'll talk to Key directly
Based on your answers, a full internal help desk build may be more than you need right now. If you're under ~50 people or still without an ITSM, an Key Builds AI knowledge assistant or audit will serve you better as a first step, and you can graduate into this later.
See it answer your questions, live.
We'll load a demo with policies that look like yours and show it work end to end: a cited answer from your handbook, the moment it can't resolve something and files a routed ticket, and the permission wall in action. Secure, US-based, and yours to keep.
Book a working demo →