Agentary AI
← Back to the website

› Lead-to-meeting automation walkthrough

From a qualified website lead to an approved appointment

This walkthrough documents the mechanism running behind the Agentary AI audit form. It demonstrates how deterministic controls, AI-assisted analysis and mandatory human review can work together without granting a model unsupervised access to customer communication or calendars.

Illustrative walkthroughProduction architectureNo client data shown
Open the audit form Review security and privacy

› Visible system model

The same stages, represented with synthetic data

The animation below uses fictional input and events. The public audit form uses the production mechanism described on this page, but private lead records and review decisions are never exposed in the demonstration.

demo #4821 · synthetic
  1. 01 / TRIGGER

    Lead submits the secure audit form

  2. 02 / AGENT

    Gemini structures the opportunity

  3. 03 / HUMAN

    A person reviews before anything sends

  4. 04 / RESULT

    A secure appointment link reaches the lead

› validated · rate-limitedsequence preview · not a client result
synthetic audit trail

› End-to-end sequence

Seven controlled transitions, each with a named responsibility

  1. 01Browser → server

    Secure intake

    The form validates required fields and collects only the context needed for an initial fit review.

  2. 02Deterministic

    Abuse controls

    Turnstile, a honeypot, request-size limits, same-origin checks and per-IP rate limits run before a lead is accepted.

  3. 03Private storage

    Private record

    The submitted lead is written outside the public webroot before any model processing or notification begins.

  4. 04Advisory AI

    AI qualification

    Gemini returns structured fields including score, confidence, priority, summary, risks, next action and a draft response.

  5. 05Explainable review

    Reason codes

    The reviewer receives the proposed score together with the reasons and uncertainty needed to judge it.

  6. 06Mandatory human

    Human approval

    A private token-protected page lets the owner reject the lead or approve a future appointment. Nothing is sent automatically.

  7. 07Customer action

    Customer hand-off

    Approval sends a branded appointment link. The attendee can add it to Google Calendar, Outlook or Apple Calendar.

› Two deliberately separate experiences

The reviewer sees risk and reasoning. The attendee sees only the confirmed appointment.

Internal review console

Accessible only through the random review token delivered to the private notification channels.

  • Lead details and AI summary
  • Score, confidence and reason codes
  • Risks and recommended next action
  • Reject, schedule or resend controls

Customer appointment page

Accessible through a separate attendee-only token created after approval.

  • Confirmed date, time and time zone
  • Meeting duration and status
  • Google Calendar deep link
  • Protected Outlook/Apple calendar download

› Security model

Controls are placed around the model, not delegated to it

Minimum exposure

Secrets and lead files remain outside the public site. Telegram receives a minimized notification rather than the submitted message.

Separated tokens

Review and attendee links use different high-entropy tokens. Stored records contain token hashes rather than the usable attendee link.

Human before communication

The model drafts and recommends. A person must approve the lead, appointment and customer-facing message.

Bounded retention

Lead, anti-abuse, review-link and attendee-link records follow declared deletion eligibility periods instead of indefinite retention.

› Declared lifecycle

What is retained and for how long

Lead record
Eligible for deletion after 180 days
Anti-abuse record
Eligible for deletion after 24 hours
Private review link
Expires after 7 days
Attendee link
Eligible for deletion 30 days after the meeting
AI interaction storage
Requested with provider storage disabled
Customer email
Sent only after human approval

› What this is evidence of

A working capability demonstration, not a customer case study

It demonstrates

  • A production-style intake and review architecture
  • Structured AI qualification with visible reasoning
  • Human approval before external action
  • A professional appointment and calendar hand-off

It does not claim

  • A customer's conversion-rate improvement
  • Guaranteed AI accuracy or revenue
  • Unsupervised calendar booking
  • Suitability for every CRM, industry or risk level

› Questions

What a prospective client should ask

Does the AI send emails by itself?+

No. It prepares analysis and a draft. A human reviewer decides whether to reject or schedule the lead before any appointment message is sent.

Does the workflow require Microsoft 365?+

No. The current implementation uses authenticated cPanel email and provides Google Calendar plus Outlook/Apple-compatible calendar actions.

Can this connect to our CRM?+

Usually, if the CRM offers an API, webhook or supported import. Authentication, permissions, deduplication and ownership rules would be defined during discovery.

Can we replace Telegram with another approval channel?+

Yes. Teams, Slack, email, a private portal or an internal workflow tool can be used if their security and access model match the project.

Use this pattern as a starting point, not a fixed product

Your qualification rules, systems, approval roles, retention and customer hand-off should be designed around your actual process.

Open the audit form