What we learned building Maya, Kulissa's hard AI buyer

Maya started as a gatekeeper on the homepage and became a product thesis with a face. She is not a mascot. She is a hard buyer on purpose. If you can convince her under interruption, you understand what Kulissa is for. If you only want a polite chatbot that flatters your pitch, you want a different website.

This is a building note from the inside: what we learned making a public AI buyer that creates stakes without tricks, and how those lessons shaped the team product. Try her on /maya. Product shape: /product. How we disclose AI: /ai-transparency.

The demo had to be the argument

Most B2B demos explain the product, then maybe show it. We inverted that. The first experience is a conversation. The visitor speaks. Maya pushes back. Credits, if earned under the rules shown in the moment, are promotional stakes for a workspace conversation. Team access stays a custom enterprise engagement scoped with us.

That choice forced honesty. If the buyer is easy, the demo lies. If the buyer is impossible, the demo punishes curiosity. Hard but fair is the narrow ridge. We tune for interruption, objection, and temperature change. We do not tune for humiliation.

Bridge Group 2026 still shows long AE ramps and thin quota attainment. ATD citing Gartner still shows unused training fading fast. A soft demo would have been brand-safe and strategically useless. Teams do not fail because they lack another polite practice partner. They fail because rehearsal never looked like the call.

Interruptions are not a bug

Early versions of voice agents wait their turn like workshop actors. Real buyers do not. They cut in on the value prop. They change the subject to security. They drop a competitor and go quiet.

We made interruption a first-class behavior. That spilled into scenario design for paying teams: park-and-return, stacked objections, recovery after losing the thread. The public field notes became playbooks like objection handling that survives interruption. If Maya always waits, she trains the wrong muscle for the homepage and for the product.

Engineering consequence: turn-taking cannot be purely "model speaks until done." Product consequence: scoring must reward yield-and-recover, not verbal dominance. Coaching consequence: managers finally hear the failure mode they see on calls.

Credits as stakes, not as a dark pattern

Credits are how Kulissa prices AI work for teams. On Maya, promotional credits are the game piece: win them under disclosed rules, redeem them onto a customer workspace, understand that they are not cash and may expire.

What we refuse:

• Hidden timers that steal a win after the fact
• Fake scarcity banners unrelated to real campaign limits
• "Free plan" language that is actually a crippled product
• Silent microphone capture beyond what the browser permission already governs

Microphone access stays optional and under the visitor's control. Abuse gets credits revoked. That is enforcement, not theatre. Team access is scoped with us as a custom enterprise engagement. Maya remains the public rehearsal.

Stakes work when they are legible. People try harder when a rematch costs something they understand. People disengage when the rules feel like a casino UI. We want competitive, not manipulative.

Structured outputs, validated twice

Every model call that matters asks for structured output against a schema. Every answer is validated with Zod before the product trusts it. Free-form prose with hand-rolled parsing exists only when a human is meant to read the text.

Why this became non-negotiable while building Maya and the debrief path:

• Providers accept only what their schema dialect can express. Length bounds and richer constraints get stripped on the way out. The Zod check on the way back in is the real gate.
• A schema-valid object can still cite a transcript span we never had. Identity and evidence checks matter as much as shape.
• UI that renders a score without evidence is how managers learn to ignore AI coaching.

We derive provider JSON Schema from the Zod contract with z.toJSONSchema rather than keeping two hand-copied shapes. Two copies drift. The one that drifts is always the one nobody tested.

Maya's conversational turns and the team's scored debriefs share that philosophy even when the schemas differ. The public AI transparency note says the quiet part: evidence for a score is checked against the transcript we actually have; a model claim about data we never supplied is rejected. Read /ai-transparency.

No dark patterns in the learning loop

It is tempting to juice demo conversion with tricks: flatter the pitch, hide the fail, inflate the score, auto-opt marketing into the microphone path. We treated those as product bugs.

Learning products die when trust dies. Managers already distrust vibes. If Maya lies to impress a visitor, the team product inherits the smell. Prefer:

• Clear AI disclosure before and during the talk
• Scores and outcomes that can be explained
• Campaign rules visible at claim time
• Acceptable use that bans farming and deception

The same ethics shows up in enterprise diligence. Voice, transcripts, and subprocessors get real pages, not footnotes. Buyers who care should read GDPR questions for AI sales training vendors and our /security, /privacy, /dpa, /subprocessors set.

What Maya taught us about scenario fidelity

Homepage Maya is one persona. Paying teams need many. The lesson transferred: generic buyers teach generic habits. Playbook-native scenes teach the show the field actually runs. That pushed Scenario Studio and document-to-scenario work so authors start from methodology language, not from a blank "skeptical CFO" prompt.

Conversion craft for customers: turn your playbook into roleplay scenarios. Onboarding sequencing that assumes hard practice early: first 30 days plan. Scorecards that survive manager skepticism: sales scorecard managers trust.

Maya also taught us that beyond-sales conversations belong on the same engine. Support escalations need ownership and policy honesty more than AE advances. The use case is public: /use-cases/customer-support-escalations.

Product principles we kept

1. The call is the page. Do not turn the first experience into a feature grid.
2. Maya is the visual and narrative anchor. Brand without the buyer is another SaaS site.
3. One action at a time. Talk, continue, claim. Not a dashboard tour.
4. Credits create tension. Stakes belong in the loop.
5. Subtraction beats explanation. If Maya can say it on the call, it may not need a paragraph in the UI.

Those lines sound like brand rules because they are. They are also engineering prioritization. Every panel you add dilutes the conversation.

What we still refuse to claim

We will not promise that Maya or Kulissa collapses your AE ramp below Bridge Group's 6.2-month average by magic. We will not invent customer logos. We will not publish a fake CSAT or quota lift without a measurement design. Software creates rehearsal conditions. Teams create transfer. That sentence keeps the roadmap honest.

When we compare with other tools, we name them and link sources on /compare. Category shopping criteria: AI sales roleplay software buyer's guide. Fair landscape naming includes Second Nature, Hyperbound, Yoodli, Mursion, Attensi, and us. Maya is not a comparison chart. She is the feeling of the category done without anaesthesia.

If you are building something adjacent

A few portable lessons:

• Make the demo isomorphic with the product thesis, not a trailer for a different movie
• Put difficulty on a dial you can defend
• Validate model output in your own runtime, not only in the provider's schema dialect
• Treat credits or quotas as game design, then hire a conscience
• Disclose AI like an adult

Kulissa is still early. Maya will keep changing. The invariants should not: hard fair buyers, structured truth, stakes without tricks.

Credits, promotions, and avoiding dark patterns

Promotional credits should feel earned, not like a slot machine. Maya's win condition is conversational, not a hidden form field. We do not want dark-pattern growth tricks. We want a fair gatekeeper scene that mirrors how hard buyers can be. If we ever cheapen that, the brand metaphor collapses into a lead form with cosplay.

Engineering taste: subtract settings

Every settings page we almost added to the demo would have diluted the first session. The visitor's job is to talk. Our job is to make talking work. Configuration belongs in the workspace after trust exists. That taste lesson now influences how we design manager surfaces too: fewer knobs at first, more opinionated defaults, progressive disclosure later.

Maya also taught us to write failure modes into the demo narrative. Visitors who lose still need a dignified next step: book a conversation, read the product page, understand credits. A gatekeeper who only exists to reject people becomes a stunt. A gatekeeper who teaches the standard while protecting the stake becomes a product lesson. We keep editing her brief so she is tough without being cartoonish, and specific without pretending to know every motion on earth. That editorial work never ends, which is appropriate. Live conversation systems are not static landing page copy. They are performances that must be directed continuously as models, latency, and visitor tactics change.

Sources

• The Bridge Group, AE Models, Motions and Metrics 2026 (AE ramp and quota attainment context for why hard practice matters)
• ATD, State of Sales Training 2023 (citing Gartner on forgetting when training is unused)
• Kulissa product and legal references: /maya, /product, /ai-transparency
• Internal engineering practice: Zod-validated structured outputs derived via z.toJSONSchema; evidence checks against supplied transcripts

FAQ

Is Maya a human?

No. Maya is an AI buyer. Surfaces disclose that. She can interrupt, object, and make mistakes. Details: /ai-transparency.

How do Maya credits relate to team access?

They are promotional stakes under campaign rules, meant to start a workspace conversation. Team access is a custom enterprise engagement scoped with us. See /contact.

Why make the demo hard?

Because polite demos train the wrong expectation. Kulissa exists for rehearsal under pressure. Soft buyers create false confidence.

Do you train models on Maya conversations?

Customer and demo content are sent to model providers to produce the response or score for that request. We do not use that content to train our own models, and we require providers not to train on it. See /privacy and /ai-transparency.

How does Maya relate to team workspaces?

Maya is the public proof. Workspaces get playbook-native scenarios, scored debriefs, and manager readiness. Start at /product or bring promotional credits into a team after a win on /maya.