Policy-aware customer support

Every reply
has a reason.

Give support teams a grounded answer, a clear decision path and a human handoff when the situation demands one.

Source of truthYour policies and knowledge base
Decision boundaryDeterministic safeguards

A better starting point

Good support begins with knowing what can be promised.

A useful answer is tied to real policies. A difficult case reaches the right person. The people responsible for support can see the work.

Policy-grounded answersExplicit escalationOperator oversight

The platform / 01—03

One path through
every decision.

ResolveHelm connects message intake, business rules, a knowledge base and the operator console. The AI can propose a response; policy controls the next step.

01 / INTAKE

Context before response.

Read Gmail threads, screen automated mail and collect the right context before preparing any answer.

02 / DECIDE

Rules with the last word.

Ground proposed replies in product guidance. Route safety issues, refund demands and exceptions to a person.

03 / REVIEW

Humans hold the helm.

Review tickets, read threads, add internal notes and write manual replies from the administration dashboard.

Go to the console ↗

A decision you can trace

From the question to
the next right step.

Explore an illustrative workflow. These examples run entirely on this page and do not contact a backend or send any email.

Illustrative workflow / A
INCOMING QUESTION / MAIL

“Which board size is right for me?”

Proposed / reply

A policy-backed answer.

  • Relevant guidance retrieved
  • Rule checked before sending
  • Thread context retained

Model-generated text is checked against deterministic safeguards. Sending depends on the configured operational mode.

Illustrative workflow / B
CUSTOMER ISSUE / MAIL

“The damaged item caused an injury.”

Proposed / escalate

A human takes over.

  • Safety concern identified
  • Needs-human route applied
  • No unapproved remedy promised

Safety and sensitive requests take an escalation path rather than receiving unreviewed automated replies.

Illustrative workflow / C
OPERATOR VIEW / TICKET

“Can somebody check my order?”

Next step / review

The case stays visible.

  • Conversation history available
  • Operator adds an internal note
  • Manual reply when appropriate

The administration dashboard provides ticket status, review and manual-reply controls.

Illustrative customer requests

Different questions.
Clear next steps.

Examples based on categories in the repository's evaluation scenarios. These are interface illustrations, not actual customer data or measured performance.

EXAMPLE / 01Product guidance

Which board fits?

Find the relevant sizing rule and draft an answer without inventing product specifications.

Policy lookup→Grounded reply
EXAMPLE / 02Sensitive request

Arrived damaged.

Route a sensitive product issue to a person instead of promising an automatic replacement.

Rule evaluation→Human escalation
EXAMPLE / 03Assembly help

One missing step.

Use documented assembly instructions and keep the suggested answer within the provided guidance.

Knowledge base→Grounded reply
EXAMPLE / 04Order question

Where is my order?

Give an operator the question so they can check order details that the system does not know.

Ticket created→Operator review

The operating principle

Automation should leave room for judgment.

01 / GROUNDING

Start with the rule.

Use maintained policy and product information as the basis for decisions before suggesting an answer.

02 / BOUNDARIES

Decide what can be sent.

Deterministic rules reject risky promises and give escalations a defined route.

03 / CONTROL

Keep the operator informed.

Conversations, tickets, notes and manual actions remain available to the support team.

Operational controls

Run with the safeguards
you choose.

These describe configuration options from the project, not subscription plans or a claim that a live service is connected.

01 / TEST

Dry-run

Examine the decision path while outgoing messages are simulated.

  • DRY_RUN=true
  • Supports development and evaluation
02 / SUPERVISE

Human review

Keep sensitive requests with an operator and preserve a visible ticket history.

  • Ticket status and conversations
  • Manual reply and internal notes
03 / CONFIGURE

Controlled sending

Live outbound sending requires explicit configuration and operational signoff.

  • AUTO_SEND_ENABLED gate
  • Needs working service credentials

Frequently asked

What happens
behind a reply?

A policy-aware AI support automation project using FastAPI, Gmail, Pydantic AI, deterministic business rules, a knowledge base, and a React administration dashboard.

No. The examples are illustrative. Running the actual system requires a configured backend, mailbox authorization and authenticated access.

The documented rule engine escalates refund demands and sensitive cases to a person. A suggested reply is never an authorization to promise an unsupported remedy.

The React console runs separately from the FastAPI service. Configure its backend URL and authorized credentials. GitHub Pages hosts static files and does not run the Python server.

Review the public repository for the backend, dashboard, architecture and test suite.

From decision to action

Keep the answer close.
Keep control closer.

Open operator console ↗
Local illustrative walkthrough

A decision, in view.

This illustration changes text on the page. It never collects customer information or sends messages.

Ready to inspect an example.