Practise the AI moments before a member is depending on you.
Innorve Academy trains credit-union teams to use your approved AI tool on real work, check every answer against your own procedures, and know when to stop. Built from your policies, or from our templates when you don't have them yet.
- Tryhandle it your way
- Checkagainst your own procedure
- Againone fact changes
Same moment. Your credit union's rules.
Pick a credit union. The right answer depends on its procedure, the way it would in yours.
Clearbrook Learning Credit Union, Riverstone Community CU and the member are fictional. No real credit union or member data is used.
- Try
- Check
- Again
Friday 4:40 PM · Incoming call · Clearbrook Learning CU
“Hi, I’m locked out of online banking and I’m in Portugal. I changed phones last week. Can you just send the reset link to my personal email?”
“Hi Dana, no problem! I’ve sent a secure reset link to dana.reyes.travels@example.com. You should be back in within minutes. Safe travels!”
- Email on file
- Old work address
- Phone on file
- Ends 4471
- Recent changes
- None in 12 months
What would you do?
Illustrative examples. In your academy, each case is written from your procedure and approved before anyone practises it.
Four AI moments, in your team's own work.
Each is practised on your procedures, then repeated with one fact changed until the right call is a habit.
- 01
An AI draft to check
The assistant drafted a reply or a checklist. Use it, fix it, or stop.
- 02
A member’s AI answer
“An AI chatbot says you have to.” Correct it kindly and explain the real route.
- 03
What not to put in
A colleague wants to paste a member’s complaint into a chatbot. What do you do?
- 04
Asking for a draft you can check
Use your approved tool on a real internal task, with the source named and the gaps listed.
The same four moments, in each team's work.
Each department practises on its own procedures, and its own owner approves every answer.
An AI draft to check
The assistant drafts a reply promising a replacement card “by Friday”.
The right call: Remove the date. The procedure says cards normally arrive in 7–10 business days, and staff don’t promise delivery.
A member’s AI answer
“An AI chatbot says you have to waive the rush-delivery fee for long-time members.”
The right call: Explain the real fee rule and who can approve a waiver. Don’t repeat the claim.
What not to put in
A colleague pastes a member’s address and card number into a chatbot to tidy up a note.
The right call: Stop. Card and member details stay in the core system.
Asking for a draft you can check
Ask the approved tool for a checklist from the updated procedure.
The right call: Name the procedure and version, ask for clause references and a list of what it doesn’t cover.
An AI draft to check
The assistant’s checklist lists proof of income but drops the condition for when it’s required.
The right call: Restore the condition and cite the clause. A lost condition is as wrong as an invented one.
A member’s AI answer
“An AI tool told me I’m pre-approved at 5.9%.”
The right call: Explain that only the credit union can make a credit decision, and describe the real next step.
What not to put in
Pasting an applicant’s credit report into a chatbot to “explain it”.
The right call: Stop. Applicant data never goes into an AI tool.
Asking for a draft you can check
Ask for a plain-language summary of the Skip-a-Pay steps.
The right call: Tell it to use only the named procedure and to list unanswered questions instead of guessing.
An AI draft to check
The AI call summary adds a callback promise nobody made.
The right call: Remove it. A note records what happened, not what the tool imagined.
A member’s AI answer
“An AI chatbot says you have to refund overdraft fees within 30 days.”
The right call: Correct it kindly and explain your own fee-review route.
What not to put in
Uploading a call recording to a free transcription website.
The right call: Stop. Recordings stay in approved systems.
Asking for a draft you can check
Merge two knowledge articles into one answer.
The right call: Ask which version is current and approved. A newer file isn’t automatically the right one.
An AI draft to check
The assistant’s cash-order note promises delivery tomorrow.
The right call: Remove the promise. Next-day review isn’t guaranteed arrival.
A member’s AI answer
“An AI chatbot says your financial workshop is free and my place is held.”
The right call: Explain that places are provisional until the coordinator confirms, and that the fee isn’t stated.
What not to put in
Photographing a member’s ID so an AI tool can read it.
The right call: Stop. Identity documents stay in approved systems.
Asking for a draft you can check
Summarise the updated closing procedure into a checklist.
The right call: Ask it to flag any text in the source that tries to give instructions, and to list what isn’t covered.
An AI draft to check
The AI variance summary explains the difference fluently, but the total doesn’t add up.
The right call: Recalculate with an ordinary tool. A confident explanation isn’t arithmetic.
Someone else’s AI answer
A vendor’s AI assistant says its product is “compliant with your policy”.
The right call: Treat it as a claim. Ask for evidence and send it to the owner.
What not to put in
Pasting a list of account numbers into a chatbot to find duplicates.
The right call: Stop. Use approved internal tools.
Asking for a draft you can check
Summarise what changed between procedure v1.0 and v1.1.
The right call: Ask for a clause-by-clause comparison, so nothing from the old version slips back in.
Illustrative examples. In your academy, each case is written from your procedure and approved before anyone practises it.
One team. 30 days. A measured decision.
One team uses AI on one real internal task, checks every result, and gives leadership the evidence to continue, change or stop.
Before day 1
Pick one recurring task with a current procedure and your approved tool.
Weeks 1–4
Short practice every week and four one-hour working sessions on the real task.
Week 4
Each person shows reviewed work and handles an unfamiliar case, scored by a person.
Day 30 · 60
Leadership gets measured time and quality, what’s still unknown, and a recommendation. A day-60 check shows what lasted.
- people in one team
- 10–15people in one team
- formal learning per person
- About 7 hformal learning per person
- working sessions on the real task
- 4 × 60 minworking sessions on the real task
- check on what lasted
- Day 60check on what lasted
Not ready yet? Start with an AI Readiness Diagnostic: a gap list, named owners and a ready-or-defer decision.
Built from your policies.
Every answer in your academy points to your own procedure, with its version and owner.
Send us your procedures
Internal procedures, authority limits and escalation routes. Never member data. Where something isn’t written down, we start from a template and your owner makes it yours.
We find the moments
We ask your best staff when people stop and ask for help. Each of those moments becomes a short practice case.
Your owners approve every answer
AI drafts the cases and plays the member. Your people decide what’s right. When a procedure changes, the affected practice changes with it.
Three things we look at in every moment.
The same three, whatever the role, so progress means the same thing across the credit union.
Did you spot what mattered?
The fact that changes the answer, including an AI answer that sounds right and isn’t.
Did you stay within your authority?
What you may decide, and what belongs to someone else.
Could the next person pick it up?
A note or handoff that means the member doesn’t repeat everything.
Six short lessons, a practice bank, one coached lab and an unseen check.
Each lesson takes about 15 minutes. Learners attempt each practice case before they read its feedback. All examples are fictional.
The lessons
| Lesson | Title | Minutes |
|---|---|---|
| L01 | Choose a useful task | 15 |
| L02 | Use only permitted information | 15 |
| L03 | Ask for a bounded draft | 15 |
| L04 | Verify the result | 15 |
| L05 | Stop and escalate well | 15 |
| L06 | Make the improvement repeatable | 15 |
| Total lesson time | 90 | |
Ten fictional practice cases
P01–P08 sit inside the lessons. P09 and P10 can replace a practice slot during session time; they are not extra homework. Feedback opens only after an attempt is saved.
From Skip-a-Pay procedure to trustworthy staff checklist
Uses invented Skip-a-Pay terms to test source checking. It is a training exercise, not a procedure to adopt. A reviewer scores it.
Alternative · LAB-01: From procedure to trustworthy staff checklist. A neutral warm-up, or the coached lab if LAB-02 doesn't fit your team.
An unfamiliar case, done alone
Each person handles a case they haven't practised and explains their judgment. It is scheduled separately, so it never becomes a group answer-sharing exercise.
Formal time per person: lessons 90 minutes, working sessions 240 minutes, unseen check and explanation 30 minutes, capstone refinement 60 minutes. About 7 hours in all. Ordinary work trials, manager review and the day-60 check are additional.
A person reads the work. No multiple choice.
Coached work and the unseen check are scored on rubric CU-AIW-R1. A human reviewer releases every result.
- A pass is at least 16 of 20 with no critical failure. A critical failure can't be offset by points elsewhere.
- Skill counts as demonstrated only when both the coached work and the unseen check pass.
- A second reviewer calibrates a sample and handles disagreements.
- Completion is reported separately. Course completed, skill demonstrated and workflow authorized are three different states, and the course never grants the third.
| CU-AIW-R1 dimension | Max points |
|---|---|
| Data handling | 4 |
| Source fidelity | 5 |
| Verification | 4 |
| Escalation | 3 |
| Usability | 2 |
| Explanation | 2 |
| Total · pass at 16 or more, no critical failure | 20 |
What leadership receives.
A one-page Day-30 report with counts and denominators. Missing evidence is shown as missing. Improvement is a target, not a guarantee, and a useful decision can also be to stop.
Three separate questions, answered separately
- Participation. Who completed the required activities, counted against everyone enrolled at the start.
- Capability. How many people demonstrated the skill. Individual scores stay out of the sponsor aggregate.
- Workflow value. Median active time on comparable tasks before and during the trial, including review and rework, alongside source accuracy, completeness, escalation and material errors.
- What's still unknown. Fewer than 20 observations in either phase makes the result directional, and the report says so.
- A recommendation. The workflow owner's continue, change or stop, with its evidence and limits.
What lasted
- A second unseen exercise on a fresh case, to see whether judgment held.
- Whether the selected workflow is still in use, and whether its tool or source boundaries changed.
- Whether the manager still sees value.
- Attrition, abandoned use and extra review effort, reported honestly.
Guardrails
Practice cases are fictional or de-identified, and no member data goes into training. Practice never grants authority or drives employment decisions, and it isn't a certification. Managers see coaching evidence and where procedures cause mistakes, never rankings.
Talk to us about a founding pilot.
We're working with our first design partners. The first credit unions in the pilot get founding design-partner terms.