Field guide 06 · Small business

Triage customer inquiries without putting trust on autopilot

Sort common inquiries, retrieve approved answers, and draft replies while people handle complaints, commitments, exceptions, and sensitive situations.

The direct answer

What is a low-complexity AI use case for a BC small business?

A bounded inquiry assistant can be a reasonable first pilot when it drafts from approved information and never sends automatically. Start with a narrow queue, exclude sensitive or high-impact topics, and measure correction work as well as speed.

Good fit when

  • The same questions recur across email or forms
  • Approved answers, policies, and escalation rules exist
  • A person reviews drafts before sending
  • Sensitive categories can be excluded

Keep outside the boundary

  • Autonomous complaint resolution, refunds, pricing exceptions, or commitments
  • Employment, health, legal, credit, or other high-impact decisions
  • Pretending a customer is speaking with a person

The workflow

Assist the work. Keep accountability visible.

Today

A team member reads each message, identifies the topic, searches for the right policy or answer, drafts a response, and routes exceptions to the owner.

AI may assist

  1. 01

    Classify the inquiry into an approved category

  2. 02

    Retrieve relevant information from the approved answer library

  3. 03

    Draft a reply with links and no invented commitment

  4. 04

    Flag complaints, sensitive information, urgency, or low confidence for a person

People must retain

  1. 01

    Review the original inquiry and proposed response

  2. 02

    Handle complaints, exceptions, prices, refunds, promises, and sensitive situations

  3. 03

    Approve the answer library and keep it current

  4. 04

    Disclose automated interaction when customers interact directly with a system

Controlled inputs
  • Customer message
  • Approved FAQ and policy library
  • Escalation categories
  • Tone and response template
  • Current hours, service area, and contact routes
Review-ready output

A review queue with category, source-linked answer, draft response, confidence flag, and escalation reason.

30-day pilot

Small enough to inspect. Real enough to learn.

Use historical or low-risk work first. Preserve the original inputs and current-process result so the comparison remains honest.

Week 01

Set the boundary

Choose one owner, one customer-inquiry queue, approved inputs, and a human approval point.

Week 02

Test side by side

Run a small historical sample through the current process and the assisted process. Keep the original records.

Week 03

Measure the exceptions

Track correction time, missed facts, escalations, and whether the proposed output was actually useful.

Week 04

Decide with evidence

Document the result, update the rules, and continue only if the workflow is safer or meaningfully better.

Measure

Count the corrections, not just the speed.

  • Time to a reviewed response
  • Percentage of drafts accepted with minor or no edits
  • Incorrect claims or commitments
  • Escalations correctly identified
  • Customer complaints linked to the assisted workflow

Stop the pilot if

  • The tool invents policy, price, availability, or commitments
  • Sensitive messages are handled without human review
  • The answer library is not kept current
  • The customer experience becomes less clear or less respectful

A failed pilot is useful evidence. It may point to a narrower workflow, better data, stronger review, or no AI at all.

Before choosing software

Ask questions that expose the operating reality.

The right product should fit the workflow, information boundary, review process, and system of record—not only produce a convincing demo.

  1. 01

    Can answers be limited to our approved content?

  2. 02

    Will each draft show the source it used?

  3. 03

    Can we block auto-send and define mandatory escalations?

  4. 04

    How is customer information retained and used?

  5. 05

    Can staff correct the answer library without technical help?

Common questions

Clear answers before the pilot.

Should the assistant reply automatically?

Not in the first pilot. Draft-only operation makes errors visible and provides the correction data needed for a responsible decision.

What inquiries should be excluded?

Complaints, disputes, sensitive personal information, exceptions, refunds, commitments, and any topic with a meaningful impact on a person.

How do we know if it works?

Measure reviewed response time, correction effort, incorrect claims, escalations, and customer outcomes—not just how many drafts it produces.

Source trail

What informed this field guide

This is a practical workflow synthesis, not reported ROI or a claim that the use case has been validated for your organization.

  1. Office of the Privacy Commissioner of CanadaPrinciples for responsible, trustworthy and privacy-protective generative AI

    Supports necessity, proportionality, transparency, accountability, and privacy review before personal information enters an AI workflow.

  2. Canadian Centre for Cyber SecurityITSAP.00.041: Generative artificial intelligence

    Supports controls for sensitive information, account security, output verification, and organizational guidance for generative AI use.

  3. National Institute of Standards and TechnologyAI Risk Management Framework

    Provides a voluntary structure to govern, map, measure, and manage AI risk throughout a pilot and deployment.

General operational education only. Confirm applicable legal, privacy, cybersecurity, professional, safety, contractual, and workplace requirements for your organization.

Make it specific

Turn this pattern into your Project Brief.

Define the owner, real inputs, approved output, success measure, and constraints before evaluating a provider.

Build this Project Brief