Explain hallucination in one paragraph a sales team could repeat accurately.

Corvestra builds Fernline, an AI voice agent that answers a company's customer-service phone line. Moraine Home Services dispatches technicians for plumbing, HVAC, and appliance repairs under home warranty plans, and is deciding whether to let Fernline take its calls. Before Moraine signs anything, its VP of Operations wants one honest answer to one question: does the thing ever just make something up.

How to answer this, stage by stage

Nobody is grading whether you know what hallucination means. They are grading whether you can compress it into four sentences a stranger could repeat correctly, under pressure, in a room where the deal is on the line.

Let's learn

The paragraph is four sentences long. It lives on a card taped inside the lid of Tullis Quilliam's laptop, right where he can read it without looking down.

Corvestra builds Fernline, an AI voice agent that answers a company's customer-service phone line instead of a person picking it up. For its first two years, Corvestra had three account executives, and all three had helped build Fernline themselves. When a prospect asked if it ever got things wrong, each of them answered from their own gut, and their gut was usually close enough, because they actually understood what the model could and could not see.

Three reps, three private answers to the exact same question, none of them written down anywhere.

Then the sales team grew to twenty two people. A call-recording audit that spring sampled forty demo calls and found eleven different answers to the same question, ranging from "it can just make things up randomly" to "it never gets things wrong, it's basically perfect."

Here's the paragraph Praveena Yaxley, the AI PM who owns this, ended up writing. It's what every rep now says, word for word, when a buyer asks the question live.

Four parts, in this order, every time. Skip one and the paragraph tips into a warning or a promise.

Both edges lose. One loses the deal today. The other loses the account later.

What it costs at its worst: five months after Cobblewick Appliance Care signed, a caller asked about a roof leak. Fernline told them it was covered under warranty, a detail that had never been loaded into Cobblewick's knowledge base at all. It wasn't a lie the model told on purpose. It was a guess, dressed up as a fact, because nothing stopped it from guessing. Corvestra's account manager spent eleven hours across three calls that month walking Cobblewick's ops team back from canceling, because the original rep had told them, flat out, "it never gets things wrong."

What I would leave alone: a buyer's technical due diligence call, where their own engineers ask about the model in real depth. That's exactly the room for the long version, confidence cut-off points, how often it routes to a person, what the logs actually capture. Forcing that same depth into the live sales demo would lose the room and never even reach the question that was actually asked.

The lesson: "use your own judgment" isn't really an answer. It's a bet that everyone's judgment stays as good as the three people already in the room. That bet pays off right up until the team grows past the room.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one for why the paragraph had to exist at all, and what it cost before it did.

Tullis Quilliam can read a buyer's face before they finish their question. Three years at Corvestra, and he's closed accounts other reps had already written off, mostly by knowing exactly when to stop talking.

For a long stretch, that instinct was enough. Corvestra was small. Every rep had sat in on Fernline's build meetings, knew which customers had thin knowledge bases and which had deep ones, knew where the model was actually strong and where it was guessing. When a buyer asked a hard question, whatever a rep said off the cuff was close enough to true, because they weren't really improvising. They were reporting.

Then the team grew, fast, to twenty two. New hires read the pitch deck and repeated its confident language word for word. A few technical hires read the model card and answered so carefully they scared people. Nobody had ever actually written the real answer down, so everyone just filled the gap with whatever felt right that week.

Word got around about what happened to Cobblewick Appliance Care before anyone at Corvestra called it a pattern.

Same wrong answer from the agent, two different outcomes, depending on what the buyer was told before it happened.

Cobblewick's rep had sold the account eight months earlier, on a demo that went well, closing with a line meant to sound reassuring: "honestly, it never gets things wrong." Five months into using Fernline, a caller asked about a leaking roof. Fernline told them, with total confidence, that it was covered under warranty. Nobody had ever loaded that detail into Cobblewick's knowledge base, because it wasn't in Cobblewick's actual policy. The model didn't know that. It answered anyway.

Cobblewick's ops team didn't hear "the model made an understandable mistake." They heard "you told us it never gets things wrong, and it just cost us money." Corvestra's account manager spent eleven hours across three calls that month, walking them back from canceling outright.

We didn't lose Cobblewick's trust the day the agent got the warranty detail wrong. We lost it five months earlier, in a demo, when a rep promised it never would.

The old decision sat in a much smaller meeting, back when Corvestra wrote its first onboarding guide. Someone asked what a rep should say if a buyer pushed on accuracy. The answer the room landed on was "use your own judgment, just be honest." It made sense that day. Every rep in the building actually knew the product cold. Writing a rigid script felt like it would make the team sound like they were reading from a card instead of actually understanding what they sold.

Cobblewick's incident is what finally moved this from a nice idea to a required one.

Praveena Yaxley wrote the paragraph in the weeks after Cobblewick's renewal call, tested it against the actual audit recordings, the real questions buyers had asked, not a made-up FAQ. It went from optional reading to a required, certified line in onboarding. Every rep had to say it back, unscripted, to a manager, before their first live demo.

Run the same afternoon again, months later, with the paragraph in place. Tullis is mid-demo with Sarolta Grielle, VP of Operations at Moraine Home Services, walking her ops team through how Fernline handles a warranty claim call. Partway through, Sarolta leans forward. "Does it ever just make things up?"

Tullis doesn't pause. He says the paragraph, the same four sentences Praveena wrote, the same words he said back to his manager in training. Sarolta writes one line in her notebook. By 10:14, four minutes after she asked, they're back on pricing.

What I'd tell my past self, sitting in that first onboarding meeting: "use your own judgment" isn't a real answer. It's a wish that everyone's judgment stays as good as the three people already in the room, and that wish runs out exactly when the room gets bigger.

SPARK, in one screen

Not a script to sound rehearsed. SPARK is what forces you to name the one paragraph a whole team owns, instead of trusting twenty two different people to each land on something close enough.

Same underlying facts, two different rooms. The live demo gets four sentences. The technical review gets the rest.

The recap, one line per letter: situation is a buyer asking a hard question live, payoff is every rep repeating one accurate paragraph instead of guessing, anchor is the paragraph itself, tuned against real audit recordings rather than a made-up FAQ, risk is either extreme costing Corvestra on a different day, and keep out draws the line at the room, not the honesty, the deep mechanism goes in a written doc, never in a live guess.

One option Praveena rejected on the way to this design: a longer, fully technical paragraph, six or seven sentences covering confidence cut-off points and routing logic in real detail, taught to every rep as the standard line. She turned it down because a rep under pressure compresses six sentences into whichever extreme feels safe, which recreates the exact problem the paragraph was built to fix, just with more words in it. She also rejected leaving it open, "say whatever feels right, just don't lie," because that was the literal status quo that produced eleven different answers across forty calls in the first place.

And if you want to be sure it really works, try it somewhere else

Same five letters, a clinic's front desk instead of a warranty dispatch line, and this time the wrong guess isn't a roof leak, it's a patient's copay.

Palterra builds an AI phone agent that reschedules and confirms appointments for Rivenwood Clinics, a network of family medical clinics. Manoela Ulyanova hit the same wall eight months into selling it: a front office director asked, mid-demo, "will it ever tell a patient the wrong copay?" and three different reps in the room had three different answers.

Same shape of problem, a different wrong guess. Scripted holds. Improvised drifts.

Mapped onto SPARK: the situation is a front office director asking about copay accuracy, live, at one clinic network. The payoff is every Palterra rep repeating one paragraph instead of guessing. The anchor is a paragraph naming the real risk, a wrong estimate on a plan detail outside that clinic's own fee schedule, the real catch, every answer grounded in the clinic's actual fee schedule with anything uncertain routed to a live scheduler, and the honest expectation, rare, logged, reviewable. The risk is the same shape as Corvestra's: a wrong dollar figure instead of a wrong coverage line, but the same two ways to lose, scaring a director off or setting up a patient complaint later. And keep out draws the same line: the deep mechanism stays for the clinic's own IT security review, never the sales floor.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it, say the paragraph, its four parts, and what it protects against.
Cost: no enablement budget this quarter. Have whoever already owns the paragraph run it as a five minute training in the next team meeting and require every rep to say it back once, free, this month.
The model got better, for real: say Fernline's accuracy on grounded answers doubles overnight. The paragraph still matters, because a smarter model that still occasionally guesses on a detail nobody gave it can still cost a customer real money once. Reliability and model quality are two different jobs.

Where people run it wrong.
They hand a buyer the full technical FAQ as their live answer and bury the actual point under detail nobody asked for.
They write the paragraph once and never check it against real call recordings, so it quietly drifts back into everyone's own words within a quarter.
They read it flat, like a legal disclaimer, which lands worse than either extreme it was built to replace.

How to use it live. Before answering a hallucination question cold in an interview, ask yourself one thing: what's the one sentence I'd want every seller at this company saying the same way, not what's the most complete technical answer I personally could give. Naming that sentence first is usually the exact distinction a SPARK question is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0