Managed AI support · by PengKernel LLC

The AI support agent
that knows when to shut up

Most AI support bots answer everything — confidently, including the things they're wrong about. DeskPilot answers from your docs, scores its own confidence, and hands off to a human instead of guessing.

See Plans →
0.12 Default confidence floor
2 Escalation triggers
0 Embedding API calls required
3 Steps to integrate
The Problem

Confident wrong answers
cost more than slow right ones

Generic chatbot wrappers will answer a billing dispute, a refund request, or a question with no good source in your docs — just as fluently as they answer "what are your hours." That fluency is the problem.

❓

Out-of-Domain Guessing

A question outside the knowledge base still gets a fluent, plausible-sounding answer — often wrong.

💳

High-Stakes Autonomy

Refunds, cancellations, and legal threats get handled by the same logic as "where's my package."

📉

No Confidence Signal

Most chat widgets expose zero information about how sure the model actually is.

🔌

Heavy Setup

Vector DBs and embedding pipelines are overkill for a knowledge base of a few hundred FAQ articles.

Under the Hood

Retrieval gates every answer

DeskPilot scores its own retrieval confidence before it ever calls the LLM — no vector database required.

01

TF-IDF Knowledge Base

Your docs are chunked and indexed with TF-IDF cosine similarity — no embedding API, no vector store, runs in-process.

02

Confidence-Scored Retrieval

Every query returns top-K chunks and a similarity score. That score doubles as a confidence signal for the next step.

03

Escalation Gate

Keyword match (refund, lawyer, cancel) or low retrieval confidence routes straight to a human — no LLM call, no guess.

04

Grounded Response

Only messages that clear the gate reach the LLM, which answers using retrieved context only and returns its sources.

Capabilities

Everything you need to ship
a support agent you can trust

📚
CORE

Zero-Setup Knowledge Base

Add documents directly — chunking and TF-IDF indexing happen automatically, no embedding pipeline to stand up.

In-memory · no vector DB
🚦
HIGH

Escalation Gate

Keyword triggers plus a configurable confidence floor decide, before any LLM call, whether a human should handle it.

Default floor · 0.12 cosine score
📎
TRUST

Cited Sources

Every answer returns the knowledge base chunks it was grounded in, so you can show citations or audit responses.

Structured · machine-readable
💬
SAFE

Multi-Turn Sessions

Conversation history is tracked per session, so follow-up questions stay grounded in prior context.

Per-session memory
⚙️
CONFIG

Deep Customization

Not a managed black box. Brand voice, escalation rules, confidence tolerance, and retrieval depth are all exposed at the API layer — tune the agent to your own policy instead of a vendor's defaults.

4 tunable parameters · open core
🔒
SAFE

Answers Only From Your Docs

The system prompt constrains the model to the retrieved context — no open-domain answers, no hallucinated policy.

Grounded generation
Quick Start

Two steps to a grounded agent

1

Load your docs

Add your FAQ or help-center articles to a knowledge base. Chunking and indexing happen automatically.

typescript
import { DeskPilot, KnowledgeBase } from "deskpilot";

const kb = new KnowledgeBase();
kb.addDocument({
  id: "shipping",
  title: "Shipping Policy",
  content: "Standard shipping takes 5-7 business days...",
});
2

Chat

Every response tells you whether it escalated, and why.

typescript
const agent = new DeskPilot(kb);

const response = await agent.chat("session-123", "How long does shipping take?");
console.log(response.answer, response.escalate);
See It in Action

Grounded answers. Real lookups.
A ticket when it can't self-serve.

Watch DeskPilot answer from its docs, call a lookup_order tool to check a real shipment, open a ticket when the shipment is stuck, and escalate without ever calling the LLM when a message matches a high-stakes keyword.

deskpilot · live demo
DeskPilot live demo: grounded answers, an order lookup that opens a support ticket, and keyword escalation
Tool Use in Practice

Resolving a ticket, not just answering FAQs

Real output from the demo — the agent looks up the order, recognizes the delay exceeds the normal window, and opens a ticket instead of guessing or bailing out.

deskpilot · example:demo 
Q: Where's my order ORD-1004? It's been 5 days with no tracking update.
   🔧 lookup_order(ORD-1004) → delayed
   🔧 create_ticket(shipping_delay) → TCK-1001
A: I've opened support ticket TCK-1001 for your order. A member of our team will
   follow up and initiate a trace with FedEx to locate your package.
   confidence=0.42 escalate=false

Q: I want a refund and I'm talking to my lawyer about this.
A: I want to make sure this gets handled correctly — I'm looping in a teammate who can help.
   confidence=0.00 escalate=true (matched escalation keyword: "refund")
Roadmap

What's coming next

Planned

Embeddings-Based Retrieval

Swap TF-IDF for a vector store once the knowledge base grows past a few thousand chunks or needs cross-lingual matching.

Planned

Channel Adapters

Zendesk, Intercom, and email ingestion/reply adapters so DeskPilot sits behind your existing support tooling.

Planned

Human Handoff Webhooks

Push escalations directly to Slack or email instead of just returning escalate: true to the caller.

Planned

Persistent Sessions

Move conversation history out of process memory and into Redis or a database for multi-instance deployments.

Pricing

Plans that scale with your support.

Plans scale by conversation volume, channels, analytics and integrations, deployment model, and support. The higher tiers aren't just more messages — they're built for customers who need configuration, accountability, and a more controlled operating environment.

Starter
$149/ month

For teams putting an AI support agent live on their docs.

Request a Pilot →
  • ✓ Confidence-scored escalation gate
  • ✓ Cited sources on every answer
  • ✓ Multi-turn session memory
  • ✓ Single channel, starter conversation volume Coming soon
  • ✓ Email support
Most Popular
Growth
$599/ month

For growing support teams across multiple channels.

Request a Pilot →
  • ✓ Everything in Starter
  • ✓ Higher conversation volume Coming soon
  • ✓ Multiple channels (Zendesk / Intercom / email) Coming soon
  • ✓ Analytics & integrations Coming soon
  • ✓ Priority support
Enterprise
$2,000/ mo+

For a configured, accountable, controlled operating environment.

Talk to Us →
  • ✓ Everything in Growth
  • ✓ Custom conversation volume Coming soon
  • ✓ Self-hosted / VPC deployment Coming soon
  • ✓ Escalation policy tuned to your business
  • ✓ Dedicated support

Free to try Reach out and we'll set up a pilot with you directly — self-serve 14-day trial signup is coming soon.

Stop your support bot from guessing.

Drop DeskPilot into your support stack in under 10 minutes.

See Plans →