Train your AI.
Skip the apps.
Three modes. One feedback loop. The AI evaluates your customers, you correct it in plain English, and it rewrites its own logic so it never makes that mistake again. By the end, you may not need an app at all.
The Mental Model
You're not using an app. You're talking to the AI that runs the automation.
When you give it feedback, it updates itself — its rules, its filters, even its own code. There's no support ticket, no backlog. You correct it, it changes, the next run is better.
How It Works
The loop in three sentences.
The AI evaluates a customer and shows you its working.
You tell it where it got things wrong, in plain English.
It rewrites its own logic so it doesn't make that mistake again.
That's it. Everything else is just mechanics.
How It Runs
Three modes. You pick the level of trust.
Same automation, same logic, same output. The only thing that changes is how much sits in front of a human before anything happens.
Mode 1
Chat
Every run creates a chat in your sidebar. You see everything the AI considered, its reasoning, its self-criticism, and the actions it would take. Nothing happens until you say "go."
When you're learning
Mode 2
Semi-Auto
Clear-cut cases handled automatically. Anything ambiguous lands in your chat for review. You stay in the loop on edge cases — the AI handles the rest.
When you trust 80%
Mode 3
Auto
The AI handles everything. No chat created per customer. You watch a dashboard. You spot-check. The system runs your operations.
When you trust 95%
Mode 1 · Chat
This is where the training happens.
Every time the automation runs against a customer, a chat appears in your sidebar. Open it and you see the entire reasoning chain — like watching the AI think out loud.
The system prompt at the top
All the rules and instructions the AI is operating under. You can read it. You can also rewrite it.
Every data point it pulled in
Emails, jobs, quotes, meetings, tasks, WhatsApp messages, Xero invoices — everything it considered for this customer. If something's missing, you'll spot it immediately.
Its analysis and suggested actions
Customer state, sentiment, churn risk, draft email replies, tasks to create, or things to push to your CRM. All structured. All justifiable.
Its self-criticism
Where it wasn't sure. What context might have been missing. What the system prompt should clarify. This is the goldmine — every problem it flags becomes a training opportunity.
What training actually looks like.
A real example — correcting an AI that mistook a Zoom upload date for the meeting date.
AI · Customer state run
Recommendation: send a reply with the Zoom link. The customer asked yesterday to schedule a Zoom — no record of one happening yet.
Wrong. The Zoom already happened 3 weeks ago. You're using the upload date, not the meeting date.
AI
Got it. I was reading the created_at timestamp from when the transcript was uploaded. Updating the system prompt now to use the meeting_started_at field instead. Future runs will have the correct ordering.
No menus. No settings panel. No support ticket. The next run uses the new rules.
The four things you'll tell it.
Pattern 1 · Wrong data in context
"This auto-reply shouldn't be in here."
The AI updates your email filters so this type of email never enters the context again.
Pattern 2 · Wrong business logic
"A blank quote-sent date means direct booking, not missing data."
The AI rewrites its own system prompt. The rule is encoded permanently for every future run.
Pattern 3 · Wrong action
"Don't push to the CRM here — escalate to the sales lead instead."
The AI updates the action logic. Different conditions now route to different outcomes.
Pattern 4 · Missing capability
"After 5 days with no reply, I want it to draft a follow-up."
The AI calls Claude Code and writes the new feature. New tool added to the automation, ready next run.
The Only Rule
Be specific.
"I don't like this" doesn't help — the AI can't infer the rule. "This auto-reply shouldn't be processed" or "treat blank quote dates as direct bookings" gives it something to encode. The more concrete your feedback, the better it gets.
Where You Trigger Training
The Knowledge app is your control room.
Every customer in one list. Hit "Run now" on any row to trigger the automation just for that customer — the chat appears in your sidebar within seconds.
Knowledge · Customers
John Markovitch
Stockdale Engineering · 2 emails this week
Sarah Chen
Chen & Associates · Quote pending
Marcus Webb
Webb Property Group · No recent activity
The amber Pending re-process badge appears when something changed since the last run — so you can see at a glance what's worth re-evaluating.
Inside each customer record.
Expand any customer to see Current and Previous state side-by-side. Watch how your training changed the analysis.
Same for AI Suggestions — current run vs. previous run. Spot regressions immediately if they happen.
Clear buttons on each panel — wipe a stale state if you want a fresh evaluation from scratch.
Bump the timestamp to force the automation to pick a customer up on its next scheduled run, without waiting for new activity.
When You're Happy
Reply "go ahead" and it executes.
The AI fires off the tasks, sends the emails, hits the webhooks, pushes to your CRM. Everything it suggested in the chat — done. No exporting, no copy-pasting, no separate workflow tool.
AI · Suggested actions
Ready to execute:
• Create quote for John Markovitch ($4,200)
• Reply to email confirming inspection date
• Create task: "Follow up if no reply by Friday"
Go ahead.
AI · Done
Quote created. Email sent. Task created. Anything else?
The Path
Don't start in auto. Earn it.
Run in chat mode for a couple of weeks. Train it on your edge cases. Once it consistently gets it right, flip to semi-auto. Once that earns trust, flip the whole thing to auto.
Week 1–2
Chat Mode
Review every run. Correct everything. Let it absorb your business rules.
Week 3–4
Semi-Auto
Let it handle obvious cases. Stay involved when it's uncertain.
Week 5+
Auto
Spot-check via dashboard. Step back. Get on with running the business.
You can always pull back to chat mode for a specific automation if something starts feeling off.
The Architecture
Your business logic becomes LLM functions.
Every business function — send invoice, create quote, push to CRM, draft email — gets encoded once as an AI tool. Then anything can call it: a chat conversation, an automation, or a button on a webpage. Change the rule once. Everything that calls it picks up the change automatically.
Things that call the functions
Multiple front-ends, one source of truth
The functions themselves · AI tools
Your business logic, encoded once
When your business rules change.
❌ Logic scattered across apps
Same rule, six places to change
- Update the website checkout
- Update the admin dashboard
- Update the mobile app
- Update the automation workflow
- Update the Zapier zap
- Update the standalone script
Miss one and you have a silent bug. The bigger the business, the worse it gets.
✅ Logic in AI functions
One function. One change.
- Update the create_quote tool
- Every caller picks it up automatically
- Chat, app, automation, webhook — all current
- One audit trail. One source of truth.
- Test once. Trust everywhere.
Your business rules live in one place — the place that thinks about them.
Before You Start
Five things to remember.
The chat IS the training.
Every conversation you have with the AI in chat mode updates how it behaves on the next run.
Be specific.
Vague feedback teaches it nothing. Give it a rule it can encode.
Read its self-criticism.
The AI knows where it's uncertain. That's the most valuable part of every run.
Don't rush to auto mode.
Two weeks in chat mode is worth six months of cleanup later.
Context is everything.
If it's not in the system, it doesn't exist. Hook up Zoom, WhatsApp, Xero, anything that informs the work.
Halfway through training, you'll notice you don't need an app.
If the AI can read your context, draft the email, create the task, send the quote — and you can correct it in plain English when it's wrong — then the only thing missing from "an app" is buttons. And buttons just call functions. The chat is already calling the functions.
Open a chat. Start training.
No coding. No prompt engineering. Just plain English feedback to the same AI that's running your operations. Chat with eVA for more details or to get started.