InterviewAdvancedResponsible AI & Advanced Practice / Internal AI tooling and enablement products / #20
Pitch an internal AI tool to a leadership team, live.
SPARK the pitch: TellerAssist, an internal AI tool drafting the compliance note a teller files after a flagged large-cash transaction, at Meridian Community Credit Union
Interviewer's question: "Pitch an internal AI tool to a leadership team, live." No product was named, so I'm pitching TellerAssist: an internal tool that drafts the compliance note a teller files after a large-cash transaction gets flagged, for the teller to confirm or edit before it's submitted. Priyanka Vashisht leads product for Meridian Community Credit Union's teller operations.
The direct answer
TellerAssist doesn't sell speed. It sells the end of a habit every teller already has: putting off the compliance note until the exact end of a shift, when it's most likely to slip past the 24-hour filing window. Three hundred forty of those notes were late last quarter, the exact finding our last exam cited. TellerAssist drafts the note the moment the transaction flags, so confirming it takes under a minute instead of twenty. It doesn't replace a compliance reviewer, and it isn't ready for every regulation type. It's ready for exactly one: large cash transactions, at one branch, starting Monday.
Do this, in order
Open with the cost leadership already feels, not the tool.Why: 340 late notes and a real exam finding get more attention in the first ten seconds than any feature description could.
Name the habit you're selling, not the model.Why: leadership isn't buying an AI tool, they're buying the end of tellers procrastinating a note until it's already too late to file on time.
Scope the pilot to one branch, one transaction type.Why: a narrow, reversible first bet protects the whole initiative from one bad early impression killing it company-wide.
Build in a floor that survives a wrong draft.Why: the compliance flag still logs even if a teller rejects the draft outright, so a bad first draft costs nothing but a minute.
Say plainly what you are not promising today.Why: naming the limits up front is what makes the rest of the pitch believable instead of a sales pitch.
How to answer this, stage by stage
Nobody is grading whether your slides look polished. They're grading whether you can open a real, skeptical room with a cost they already feel, and close it with limits you're honest about.
Stage 1
Name the product, since none was given
Say it like this
"No product was named, so I'm pitching TellerAssist, a tool that drafts the compliance note after a flagged large-cash transaction, for a teller to confirm."
Why this works
Picks a lane immediately instead of stalling on clarifying questions, which is what a real live pitch demands.
Stage 2
Say your structure out loud
Say it like this
"I'll use SPARK. Situation, payoff, anchor, risk, keep out."
Why this works
Signals a real pitch structure before the numbers start, so it doesn't sound like an improvised sales pitch.
Stage 3
Open with the cost, not the tool
Say it like this
"Last quarter, 340 compliance notes were filed more than 24 hours late. That's the exact finding our last regulatory exam cited."
Why this works
This is the situation step, and the strongest possible opener, since it's a number the room already knows hurts.
Stage 4
Name the habit you're actually selling
Say it like this
"I'm not selling a faster form. I'm selling the end of tellers putting the note off until the last minute of a shift, which is exactly when it slips past the filing window."
Why this works
The payoff step, and the real insight the question is testing for.
Stage 5
Give the anchor, scoped and concrete
Say it like this
"One branch, one transaction type, starting Monday. TellerAssist drafts the note the moment the flag fires, and confirming it takes under a minute."
Why this works
Matches the direct answer, concrete enough for leadership to picture exactly what they're approving.
Stage 6
Name the risk, and how you've protected it
Say it like this
"If the first draft is wrong, nothing is lost. The compliance flag still logs even if a teller rejects the draft outright. One bad first impression doesn't sink the pilot."
Why this works
The risk step. Shows the anchor survives its own worst day, not just its best one.
Stage 7
Say plainly what you're not promising
Say it like this
"I'm not promising this replaces a compliance reviewer. I'm not promising every regulation type on day one. I'm not promising zero audit findings, ever."
Why this works
The keep-out step. Naming real limits is what makes the whole pitch trustworthy instead of a sales pitch.
Stage 8
Close on the one line
Say it like this
"340 late notes, one exam finding, one branch starting Monday. That's the whole pitch."
Why this works
Restates the direct answer, ready for whatever gets pushed on next.
Let's learn
Picture the pitch itself: a leadership team that has already read one exam finding about late compliance notes, sitting across the table, arms crossed, waiting to be sold something.
Right now, when a large-cash transaction flags for compliance review, the teller has to write a note by hand, findings, transaction detail, next steps, before the end of their shift. It takes about twenty minutes, on top of an already full counter.
Four steps, and the third one is where 340 notes a quarter quietly slip past the window.
The turn: tellers weren't forgetting the note. They were putting it off, on purpose, because twenty minutes felt impossible to find between customers, and the note always felt safe to do "in a minute," right up until the shift ended and it didn't get done at all.
Knowledge spark: why does a 24-hour filing window matter this much?
Regulators require certain large-cash transactions to be documented within a set window because delayed documentation makes patterns, like structuring deposits to avoid reporting, much harder to catch in time to matter. A late note isn't just paperwork, it's a real gap in the record.
At its worst: the same habit repeats every quarter, the exam finding repeats with it, and eventually a genuinely suspicious pattern goes undocumented long enough that it isn't caught until real harm has already happened.
What TellerAssist is not promising
Not a replacement for a human compliance reviewer, who still signs off on every note. Not every regulation type on day one, just large cash transactions to start. Not zero audit findings ever, just far fewer late ones.
What I would leave alone: the actual compliance review process itself doesn't need to change at all. A human reviewer still checks every note. This is a filing-speed fix, not a policy change.
The lesson: a late compliance note was never really a compliance problem. It was a twenty-minute problem wearing a regulatory costume, and that's the problem worth pitching a fix for.
Now here is the same thing as a story
The short version above is the pitch itself. Read this one for the story behind the number that anchors it.
Priyanka Vashisht can spot which teller wrote a compliance note just from how it's phrased, a skill from two years reviewing them personally before she moved into product.
Meridian's last regulatory exam cited exactly one recurring finding: compliance notes filed more than 24 hours after a flagged transaction. Three hundred forty of them, across every branch, in a single quarter. Nobody had been ignoring the rule. Tellers were simply putting the note off until the end of a busy shift, the same way anyone puts off a twenty-minute task during a day with no twenty minutes to spare.
The finding wasn't about carelessness. It was about a twenty-minute task landing at the exact moment nobody has twenty minutes left.
Priyanka pitched TellerAssist to leadership with that one number first, before a single feature slide. She scoped the pilot to a single branch and a single transaction type, large cash deposits, the most common flag and the clearest rule to draft against.
Eight weeks, one branch, one transaction type, and the number that mattered most came at the end.
Compliance notes filed more than 24 hours late, before and after the pilot
One branch, one transaction type, and the finding that anchored the pitch nearly disappeared there entirely.
Share of notes filed late at the pilot branch, week by week
The habit changed gradually, week by week, not the instant the tool switched on.
Leadership didn't approve TellerAssist because the demo was slick. They approved it because Priyanka opened with a number they already knew hurt, and closed by naming exactly what she wasn't promising yet.
SPARK, in one screenNot a feature pitch. SPARK is what makes a live pitch survive the room's first skeptical question.
S
Situation. The cost leadership already feels.
340 compliance notes filed late last quarter, the exact finding the last regulatory exam cited.
Opens with something the room already knows hurts, not a hypothetical.
P
Payoff. The habit being sold, not the model.
Tellers stop putting the note off until it's too late to file on time.
Leadership buys a behavior change, not an AI feature.
A
Anchor. The number and the scope.
340 late notes, one branch, one transaction type, starting Monday.
The hardest step, and the direct answer: a real, concrete first bet leadership can actually picture.
R
Risk. What protects the first bad draft.
The compliance flag logs even if a teller rejects the draft outright, so one bad first impression can't sink the pilot.
Proves the pitch has thought past the demo, to the day something actually goes wrong.
K
Keep out. What isn't promised.
Not a replacement for a compliance reviewer, not every regulation type, not zero audit findings ever.
Naming real limits is what makes the pitch believable, not a longer feature list.
Four fields, and the third one is what makes the whole pitch survive a bad first draft.
The recap, one line per letter: situation is 340 late notes and one real exam finding, payoff is ending the habit of putting the note off, anchor is one branch and one transaction type starting Monday, risk is the flag logging regardless of the draft, and keep out is the three things not being promised today.
And if you want to be sure it really works, try it somewhere elseSame five letters, a propane delivery service instead of a credit union. Nothing else about the two jobs is alike.
Halden Propane Delivery pitches an internal AI tool that drafts a driver's safety-incident documentation after a near-miss on a delivery route, replacing a report drivers currently put off until the very end of a shift.
Only one of these two designs survives the first bad draft with the pilot still intact.
Mapped onto SPARK: situation is a real number, dozens of near-miss reports filed late enough that a pattern across routes went unnoticed for a full season. Payoff is drivers stopping the habit of leaving the report for the depot at the end of a shift. Anchor is one route, one incident type, drafted the moment a driver logs a near-miss over the radio. Risk is the incident still logging even if a driver rejects the draft outright. Keep out is not replacing a safety officer's review, not covering every incident type on day one, and not promising zero missed patterns ever.
The safest first bet sits in the corner that's both common and easiest to draft rules for.
Swap the trigger and it still runs.
Speed: an interviewer caps you at thirty seconds. Say "340 late notes, one real exam finding, one branch starting Monday, and here's exactly what I'm not promising yet," and stop.
Cost: if leadership balks at even a one-branch pilot's cost, offer the draft-only version first, no live integration, just a template a teller can copy from, proving the habit change before asking for engineering time.
The model gets better, for real: if TellerAssist's draft accuracy improves past the pilot, the pitch still holds, because the real sell was never accuracy, it was ending a habit of procrastination that existed regardless of how good the draft was.
Three honest limits, and naming them is what makes the rest of the pitch worth believing.
Where people run it wrong.
They open with the tool's features instead of the cost leadership already feels.
They promise a full rollout instead of a small, reversible first bet that survives a bad first impression.
They avoid naming limits, which makes the whole pitch sound like a sales pitch instead of a real plan.
How to use it live. When someone asks you to pitch an internal AI tool on the spot, ask yourself first: what's the one number this room already knows hurts. Open there, and close by naming what you aren't promising yet.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits "pitch an internal AI tool to a leadership team, live"?
Tap to flip
ANSWER
SPARK: situation, payoff, anchor, risk, keep out. Built for design and pitch questions, not a metric or estimation method.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Priyanka Vashisht, who leads product for Meridian Community Credit Union's teller operations, and can identify a teller's compliance note just from its phrasing.
3 · THE SITUATION
What real cost does leadership already feel, before any pitch begins?
Tap to flip
ANSWER
340 compliance notes filed more than 24 hours late last quarter, the exact finding the last regulatory exam cited.
4 · THE ANCHOR
What's the one concrete decision the whole pitch hangs on?
Tap to flip
ANSWER
A pilot scoped to one branch and one transaction type, large cash deposits, starting immediately, with drafts ready in seconds instead of twenty minutes.
5 · THE OLD HABIT
What were tellers actually doing before TellerAssist, and why?
Tap to flip
ANSWER
Putting off the twenty-minute compliance note until the end of a busy shift, when it was most likely to slip past the 24-hour filing window entirely.
6 · THE NUMBER
Fill in the blank: at the pilot branch, the share of late notes fell from 29 percent to ___ percent by week eight.
Tap to flip
ANSWER
2 percent. The habit changed gradually, week by week, not the instant the tool switched on.
7 · THE REPLAY
Same regulatory finding, TellerAssist in place. What changes?
Tap to flip
ANSWER
The pilot branch's late notes drop from a share of 340 company wide down to just 6, since the note now takes under a minute to confirm instead of twenty.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this again for a different company. Which one, and what's the anchor there?
Tap to flip
ANSWER
Halden Propane Delivery. Same anchor shape: one route, one incident type, a near-miss safety report drafted the moment it's called in.
Check yourself Score: 0 / 0
Multiple choice
1. Why does the pitch open with "340 late notes" instead of describing TellerAssist's features first?
A. Because features are hard to describe in a live pitch.
B. Because it's a cost leadership already knows and feels, which earns attention faster than any feature list.
C. Because 340 is a round, memorable number.
D. Because regulators require pitches to open with a statistic.
Show hint
Look at the situation step, S.
Show answer
B. A cost the room already feels is the strongest possible opener for a live, skeptical pitch.
True or false
2. True or false: TellerAssist is being pitched as a replacement for a human compliance reviewer.
True
False
Show hint
Look at "what TellerAssist is not promising."
Show answer
False. A human reviewer still checks every note. TellerAssist only speeds up the draft, it doesn't remove the review step.
Fill in the blank
3. Fill in the blank: the compliance note took about ___ minutes to write by hand, before TellerAssist.
Show hint
Look at Section 1's opening paragraphs.
Show answer
Twenty minutes. The exact twenty minutes tellers didn't have between customers, which is why the note kept getting put off.
Short answer, where it wouldn't matter
4. Name a part of Meridian's compliance process this pitch genuinely didn't need to change.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: The actual compliance review process itself. A human reviewer still checks every note; this is a filing-speed fix, not a policy change.
Short answer, apply it yourself
5. Think of a task you or a colleague tends to put off at work. What number would make a skeptical leadership team take a fix for it seriously?
Show hint
Think about a cost that's already been measured or complained about somewhere, not one you'd have to invent.
Show answer
Model answer: Something like the real count of missed deadlines or late submissions already sitting in an existing report, which leadership has likely already seen and felt.
Short answer, the number question
6. If the pilot branch's late-note share had only dropped to 15 percent by week eight instead of 2, would the pitch's core claim still hold? Why or why not?
Show hint
Think about what the pitch's habit-change claim actually depends on.
Show answer
Model answer: Partially. A drop to 15 percent would still show real improvement, just slower than claimed, and would be honest grounds to keep the pilot scoped narrow a bit longer before expanding.
Before you close the answer
Why this works
Tests whether you can open a real, skeptical pitch with a cost the room already feels, and close it by naming honest limits, instead of overselling the tool.
Follow-up traps
"Why not roll this out to every branch immediately if it works this well?" Response: one branch first is what makes a bad early draft cheap to recover from; a company-wide rollout with the same bad draft could kill the whole initiative's credibility in week one.
"Isn't 'not promising zero audit findings' a weak thing to say in a pitch?" Response: it's the opposite, naming that limit honestly is exactly what makes the 340-to-6 number believable instead of sounding like a sales claim.
If pressed
TellerAssist's actual draft only auto-fills fields that come directly from the transaction record itself, teller name, amount, timestamp, account number, never the narrative judgment section, which stays a blank field the teller has to write themselves every time.
From U2xAI Academy
From answering questions to owning outcomes.
A live workshop where you ship a working AI agent, defend a launch decision, and walk away with a portfolio recruiters can't wave off, not just more questions to study.
