Automation and AI Engineering

For Narqix automation and AI engineers building CRM workflows, integrations, lead routing, AI agents and conversational AI. Use this page as your working checklist to ship automation to the Narqix standard.

Track: Specialist · Role: Automation Engineer, AI Engineer · Time to complete: 12 min

The Narqix view of automation

Narqix does not sell "AI". We remove bottlenecks. Before you build, answer three questions: what manual work does this remove, what does it save per month (hours, response time, lost leads), and what happens when it fails? Humans and automation work together: automation handles volume and speed, people handle judgment and relationships.

Definition of done

The workflow does exactly what the Scope Sheet describes, proven with test runs that cover normal, edge and failure cases
Every failure path ends in an alert to a named human or a safe stop, never silence
A kill switch exists, and the client knows how to use it
Logs show every run: trigger, input, actions, result
A runbook explains what it does, when it runs, how to check it, and what to do when it fails
Baseline and after numbers recorded (hours saved, response time, error rate)
The client accepted it in writing

Quality checklist

Design

Process mapped on paper before building: trigger, steps, decisions, owners, handoffs
Broken processes fixed first. Automating a bad process makes the mess faster
One workflow per job, with a clear name: Client_Area_Trigger_Action_v01
No duplicate triggers that could fire the same action twice
Data fields mapped and documented between every connected system

Reliability

Safe to run twice: a retry does not create duplicate contacts, messages or charges
Retries with limits for temporary failures; no endless loops
Rate limits of connected systems respected
Time zones handled correctly for every scheduled action
Bulk actions tested on a small batch before the full list
Monitoring alert if the workflow stops running or error rate rises

AI agents and conversational AI

Written purpose, scope and boundaries: what it may answer, what it must hand to a human
Never claims to be a human. Follows the client's disclosure rules and the law where the conversation happens
Never invents facts, prices, availability or promises. Answers only from approved knowledge
One question at a time, short messages, natural tone
Clear handoff: when a human takes over, the AI stops talking, and the human gets the full context
Fallback when no human replies within the agreed window: a short honest acknowledgment and an alert
Test set of at least 30 real-style conversations, including hostile, confused and off-topic ones, run before every change to prompts or settings
Prompts and settings versioned like code
Usage cost per conversation estimated, with a monthly cap and alert

Messaging compliance (SMS, calls, email)

Consent recorded for every contact who receives marketing messages
Opt-out words (STOP and similar) handled instantly and permanently across every workflow
Quiet hours respected in the recipient's local time
Sender registration completed where carriers require it
Federal and state rules checked with the client; legal review recommended for high-risk use

Pre-launch checklist

All testing done with test contacts, never with real customers
Client reviewed the workflow map and approved every customer-facing message
Backup or export of affected CRM data taken before any bulk change
Kill switch tested
Alerts go to a named person who is on duty
Launch in stages: 10 percent of volume or one segment first, check results, then full rollout
First 24 hours monitored: run logs, error counts, sample of messages sent, AI conversations read
Runbook and change log saved in the client's documentation

Coach's note: read 20 real AI conversations every week, even after launch. Dashboards tell you volume. Transcripts tell you where the AI confuses people, where leads drop, and what the next fix should be. This habit is what separates a working AI setter from an expensive embarrassment.

Reporting KPIs

KPI	Notes
Successful run rate	99 percent or more for production workflows
Errors and failed runs	Count, cause and fix, every week
Hours of manual work removed per month	Compared with the baseline you recorded
Speed to lead	Time from inquiry to first response, before and after
AI handoff rate and handoff quality	Share of conversations sent to humans, and whether they were the right ones
AI answer accuracy on the test set	Tracked after every change
Opt-out and complaint rate	Rising numbers mean the messaging needs work
Running cost per month	Against the agreed cap

Access and security rules

Every workspace, integration platform, AI account and API key belongs to the client's business. You work under a named seat.
Use service accounts and keys with the narrowest permissions the workflow needs. Never connect a client workflow through your personal account.
Keys and tokens live in the vault or the platform's secret storage. Never in workflow notes, chat or documents.
Rotate keys at offboarding and after any suspected exposure.
Customer personal data goes only to AI services the client approved in writing for that data.
Webhooks and endpoints are protected with secrets or signatures so outsiders cannot trigger them.

Common mistakes

Open the list

Testing on live contacts "just once".
No alert when a workflow fails, so it breaks silently for weeks.
Two workflows reacting to the same trigger and double-messaging leads.
AI given broad freedom with no knowledge limits, then inventing prices or promises.
Ignoring opt-outs in one workflow while respecting them in another.
No documentation, so only you understand how it works.
Building complex automation around a process the client has not agreed on.

Red flags to escalate

Red flag	Level	What to do
Workflow sending wrong, duplicate or mass messages to customers	P1	Hit the kill switch, alert Delivery Lead and client within 15 minutes
AI agent gave false information, claimed to be human or made a promise	P1	Pause it, save the transcript, escalate
Opted-out contacts receiving messages	P1	Stop sending, escalate
Leaked or exposed key or token	P1	Rotate it, escalate
Bulk change corrupted or deleted CRM data	P1	Stop, restore from backup with Delivery Lead approval
Usage costs running above the monthly cap	P2	Pause non-critical runs, inform client
Client asks to message contacts with no recorded consent who are not on an approved, scrubbed cold list	P2	Decline, inform Delivery Lead

Never do this: never let an AI agent pretend to be a human, and never send marketing messages to people who opted out, or to contacts who neither consented nor are on the client's approved, scrubbed cold list. These are not settings to "test". They are lines Narqix does not cross.