A CRM AI Copilot That Proposes, Then Waits
Ask it anything across your workspace, watch which tools it used to answer, and approve the writes yourself. KudosCRM handles the drafting, the summarizing, and the extraction while every output stays labeled and nothing reaches a customer without you.
⌘JKudos Intelligence
over Deal · Meridian Group · step 4 of 8You
Where does this deal stand, and set the close date to the 30th?
Kudos Intelligence
Negotiation since 12 Aug, $42,000, owner Anita. Last touch was a proposal sent 6 days ago with no reply.
Source: model · Labeled AI
⏸ Pending write proposal
not appliedupdate_deal · expected_close_date
12 Sep 2026 → 30 Sep 2026
The Short Version
What Is the KudosCRM AI Copilot?
The KudosCRM AI copilot is an assistant that works on the records you already have. It reads the deal, the thread, and the notes, then drafts the reply, summarizes the history, pulls a document into fields, or answers a question about your pipeline. It plans, it calls tools, it shows its work, and it stops before it changes anything.
Ask and See the Working
A full page and a ⌘J drawer, an agentic loop over audited tools, and a list of which tools it called to reach the answer.
Bring Your Own Agent
An MCP server with 41 guarded tools lets your own Claude Desktop, Claude Code, or any MCP client work your CRM through your own token.
Writing Where the Writing Happens
Compose, reply, rewrite, translate, subject lines, record summaries, document extraction, form drafting, and four AI steps inside workflows.
A Visitor-Facing Agent
A support agent with a confidence threshold you set, handing to a human with a written summary the moment it is out of its depth.
Nothing sends, and nothing saves, on its own. Copilot writes come back as proposals. Agent creations land as drafts. No tool anywhere in the product can email a customer — by architecture rather than by setting.
The Copilot
Ask Across the Workspace, and Watch Which Tools It Used
You ask the copilot a question from a full page or a ⌘J drawer, and it plans an answer, calls the same audited tools the rest of the product uses, and reports which ones it called. Most assistants answer and leave you to trust them. This one shows the path, then stops at the edge of anything it would change.
Two Surfaces, One Assistant
A full page when you want to sit with it, and a drawer over whatever record you are already looking at. Progress streams step by step, so a long answer never looks like a frozen screen.
The Closest Thing to Citations
It reports the tools behind each answer. When the number came from your pipeline, you can see that it came from your pipeline.
Capped on Purpose
Eight reasoning steps, token budgets, repeated-call detection, and a quota check before it starts. An assistant that cannot run away is one you can leave switched on.
Write Approval Is Off by Default
A write tool call returns as a pending proposal for a human to confirm — each proposal approved on its own, never one confirmation covering a whole run. Not a setting you have to find. The default.
Mid-run · tools called so far
⚡ Answered on the deterministic path · no model call
4 of 8 (cap)
passed before start
🔒 Write approval is off by default — not a setting you have to find
Audit rows · connected agent
Bring Your Own Agent
Bring Your Own Claude, and Give It 41 Guarded Tools
You connect Claude Desktop, Claude Code, or any MCP-speaking agent to your CRM over a single endpoint and your existing token, and it works inside the same walls your team works in. Nineteen read tools, twenty-two write tools, and nothing else.
Nineteen Read Tools, Twenty-Two Write Tools
Search deals, contacts, companies and leads, pull a pipeline summary, list workflows, forms, products, audiences, templates, campaigns and articles. Create and update deals, move stages, log activities, and create workflows, forms, pipelines, products, price books, audiences, campaigns and help articles. That is the whole surface — 41 tools, each an explicit, reviewed line in a registry.
Your Token Is the Leash
Authentication is the Personal Access Token you already have, so there is no new credential to manage. Revoke it and the agent's access dies in the same moment. Hand it a read-only token and the write tools are removed from its list entirely, then refused anyway if something reaches for them.
What It Creates Lands Inert, and It Cannot Email a Customer
Workflows arrive disabled. Products arrive as drafts and price books inactive. Campaigns arrive as drafts, and the one send tool can mail a test only to your own address — the recipient is not even an argument. Forms arrive unpublished. There is no delete tool and no bulk tool.
Four Guards, and an Audit Row Either Way
Token scope, resource scope, module entitlement, role permission. Every call writes an audit row whether it was allowed or denied, with emails and phone numbers scrubbed from the logged arguments. The agent sees exactly what that user's own role sees.
A Summary That Refuses to Invent
The pipeline summary returns real stage totals, honoring record visibility and field-level masking, and refuses to produce win rates, forecasts, or percentage-to-target. An invented number is worse than no number.
The Writing Tools
The Assistant Behind Every Draft Button
The assistant is most useful in the places you were already typing, and none of them open a separate window. It drafts, rewrites, summarizes, and extracts, and then it hands the result back to you unsent.
The Email Assistant
Compose from context, reply into a thread, summarize, rewrite, translate, and generate subject and preheader pairs, with block generation and tone adjustment on the builder side. Gated on the email-send permission, so drafting rights follow sending rights.
Summaries on Demand, Never on Load
Company briefs, activity-timeline summaries, and task-comment summaries, each behind a button. The button is the consent. Summarizing every record automatically would bill a model call for a page people open just to scan.
Documents Into Fields
Paste the text or point at an uploaded file, and structured fields come back ready for you to check.
Forms From a Description
Describe the form and the fields are drafted. Common descriptions are answered instantly from an archetype catalog with no model call at all, which is why it feels immediate.
Four Steps Inside Your Workflows
Summarize, classify, draft email, and next-best-action run as real model-backed handlers inside automations, every call usage-logged. Drafts land in the run for a person to send.
The Smaller Assists
Subtask suggestions, tag suggestions that say whether they came from the model or a heuristic, a read-only standup digest for ops, and one-click deal email drafts.
The Visitor-Facing Agent
An Agent That Knows When to Fetch a Human
Your visitor-facing agent answers from four kinds of knowledge and stops at a confidence threshold you set. A visitor-facing agent is only as good as its worst answer, so this one is built around the moment it should stop talking.
A Threshold You Set, Topics You Name
Tune the confidence threshold between 0.3 and 0.9 to your own tolerance, then name the topics it must refuse and the topics it must escalate. Escalation topics ship seeded in English and Nepali.
Handoff With the Story Attached
When it is not confident it hands to a person with a written summary of what the visitor actually wanted. If the model is unavailable the summary degrades to an extractive one built from the visitor's own words, which cannot hallucinate. Handback is automatic.
Published in Versions, Paused Without a Deploy
Draft, live, paused, with a config snapshot for every published version and a pill when there are unpublished changes. Pausing takes effect immediately, with no release required.
Four Kinds of Knowledge
Files, question-and-answer pairs, help-center article sync, and a website crawl that respects robots.txt and noindex. Plus a sandbox tab to test against, an evaluation harness in CI, and an analytics tab.
Every Miss Is Recorded
Questions it could not answer land in a review queue, so the gaps in your knowledge base are a list you can work rather than a feeling you have.
Handoff to a human confidence 0.41 · below your 0.6
handback automaticVisitor · “Can I export invoices in NPR and keep the tax lines?”
Agent · matched 2 articles, neither covers currency on export → escalating
📝 Summary attached for the agent
Visitor wants multi-currency invoice export with tax lines preserved. Already read Billing Basics.
Logged to the review queue as an unanswered questionknowledge gap
What Sends, What Saves, and What We Will Not Call AI
Copilot writes are proposals awaiting confirmation, agent creations land inert, and no tool in the product can email a customer. Every output reports its source, model or deterministic, carries a degraded flag when it fell back, and wears an AI label — because every capability here has a non-model fallback and the default mode answers easy cases deterministically before escalating. Every call is logged against your workspace under daily call and spend caps.
And the things that are not AI, we do not call AI: lead scoring is admin-authored point rules with fixed grade bands, duplicate detection is deterministic matching, knowledge search is lexical rather than semantic with no embeddings anywhere, the pipeline forecast is weighted and rule-based rather than a black-box prediction, and predictive win probability is roadmap. Kudos AI runs on a commercial large language model, called server-side from KudosCRM, and the provider is named in our subprocessor list.
How It Works
How Does the KudosCRM AI Copilot Work?
Asking the copilot takes one keystroke, and here is what actually happens along the way.
Ask From Wherever You Are
Open the full page, or hit ⌘J over the record already in front of you.
It Plans and Calls Tools
An agentic loop over the same audited tools the product uses, capped at eight reasoning steps with your quota checked before it starts.
The Cheap Path Runs First
Easy cases are answered deterministically with no model call, and the result says so.
You See What It Used
The tools it called are listed with the answer, so the reasoning is inspectable rather than asserted.
Writes Come Back as Proposals
Anything that would change a record arrives as a pending proposal, with your approval as the gate.
It Is Logged, Labeled, and Capped
The call lands in your workspace usage log, the output carries its source label, and the spend counts against caps you set.
What's Actually in the Box
Every AI Feature, at a Glance
A copilot on a full page and a ⌘J drawer, with live progress
An agentic loop over audited tools, with the tool list shown
Eight-step cap, token budgets, repeated-call detection, quota pre-check
Write approval off by default, so proposals wait for a human
An MCP server with 41 tools, 19 read and 22 write
Personal Access Token auth, revoke to cut access instantly
Read-only tokens lose the write tools entirely
Agent creations land inert: drafts, disabled, inactive, unpublished
No delete tool, no bulk tool, and no way to email a customer
Four-axis guard on every call, with a scrubbed audit row either way
A pipeline summary that refuses to invent win rates
The email assistant: compose, reply, rewrite, translate, subjects
Record, timeline, and comment summaries behind a button
Document extraction into structured fields
An AI form builder with an instant no-model archetype path
Four AI steps inside workflows, all usage-logged
Subtask, tag, and deal-email assists, each labeled by source
A support agent with a 0.3 to 0.9 confidence threshold
Refusal and escalation topics, seeded in English and Nepali
Handoff with a written summary, extractive if the model is down
Versioned publishing, and pause without a deploy
Four knowledge source types, with a crawl that respects robots.txt
Unanswered questions collected in a review queue
Every output labeled by source, with a degraded flag
Per-workspace usage logs, daily call and spend caps
A non-model fallback behind every capability
KudosCRM AI vs. a Copilot Bolted onto Your CRM
Let It Draft. You Decide What Ships.
The AI module is part of KudosCRM, not the upgrade that unlocks the assistant.
The First Thing Each Industry Asks It
Different teams never get to different writing. KudosCRM drafts the viewing follow-up, the renewal catch-up, the contract pulled into fields, and the midnight support answer, and leaves every one of them waiting for your approval.
Works With the Agent You Already Use
The MCP server speaks to Claude Desktop, Claude Code, and any MCP-compliant client over one endpoint and your own token. Inside the CRM the assistant drafts into your Gmail or Outlook composes, your workflows, and your help center. Predictive scoring and enrichment are not here, and are labeled roadmap rather than implied.
Gmail
OutlookGo Deeper
Explore More Features
Kudos AI runs across the whole product — the AI steps live inside Workflows, the drafts land in Email Campaigns, the extraction fills Forms, and the support agent answers from your Knowledge Base.
All FeaturesFAQ
Frequently Asked Questions
Start With One Draft
Ask It Something Before You Trust It
Open the drawer, ask about a deal you already know, and check its answer against what you know.