Choosing the right Claude model
Learn when to use Haiku, Sonnet, or Opus to get better results and stay inside your rate limit. A practical guide to picking the right Claude model.
Yoodli turns high-stakes sales conversations into repeatable practice with Claude
Industry:
Software
Company size:
Startup
Product:
Claude Platform
Location:
North America
23% more deals closed
by sales reps who practice three or more scenarios per week on Yoodli
1,600+ manager hours recovered per quarter
at Snowflake after replacing manual roleplay sessions with Yoodli's AI-driven practice
Yoodli is a secure, experiential learning platform that uses AI sales roleplays to help enterprise teams practice high-stakes conversations before they happen. Sales reps rehearse discovery calls, customer success teams prepare for escalations, and leaders practice difficult feedback conversations, all against AI personas that listen, push back, and stay in character. Organizations including Snowflake and Google Cloud use Yoodli to train thousands of employees at a time.
With Claude, Yoodli achieved:
- 23% more deals closed by reps who practice three or more roleplay scenarios per week
- 1,600+ manager hours recovered per quarter at Snowflake, their customer, who previously spent manager time running manual roleplay sessions
- 15,000+ partner reps certified at Google Cloud, scaling a program that would have been extremely difficult with traditional coaching
- 40% faster ramp time for new sales hires across customer organizations
The challenge
Choosing the right Claude model
Learn when to use Haiku, Sonnet, or Opus to get better results and stay inside your rate limit. A practical guide to picking the right Claude model.
Choosing the right Claude model
Learn when to use Haiku, Sonnet, or Opus to get better results and stay inside your rate limit. A practical guide to picking the right Claude model.
Reps learn on live calls instead of in practice
Before AI-powered practice simulations, enterprise training teams relied on classroom sessions, recorded examples, and occasional manager-led practice. A sales rep might rehearse a scenario once or twice during onboarding, but structured practice was hard to sustain at scale. Running roleplays required significant manager time, and feedback quality varied depending on who was coaching. Teams that invested in call analytics could review what happened after a conversation, but had no scalable way to let reps practice and get feedback before high-stakes calls.
"Too often, reps learned how to handle critical conversations through real experience on live customer calls rather than structured, repeatable practice." said Derek Sessions, CTO at Yoodli. The cost wasn't just inconsistency. Enablement leaders had little visibility into who had actually practiced, how well they performed, or whether new messaging was being adopted across teams.
The bar for any AI-powered alternative was high. "These tools often get one real chance to make a first impression with a team," Sessions added. "If the experience feels artificial or unhelpful, users quickly lose trust and are unlikely to come back."
The solution
Introducing Claude Sonnet 4.6
Hybrid reasoning model with superior intelligence for agents, featuring a 1M context window
Introducing Claude Sonnet 4.6
Hybrid reasoning model with superior intelligence for agents, featuring a 1M context window
Introducing Claude Sonnet 4.6
Hybrid reasoning model with superior intelligence for agents, featuring a 1M context window
How conversational quality determined model selection
For a product built around real-time conversation, latency matters. But Sessions and his team found that conversational intelligence mattered even more. If an AI persona drifts out of character, gives unrealistic objections, or responds with generic answers, experienced professionals stop taking the exercise seriously.
"Claude felt the most human and realistic out of the box compared to other models we tested," Sessions noted. Cost wasn't the deciding factor; it was conversational quality. "Our go-to-market team spent a lot of time testing the roleplays with us, and it became obvious quickly which models produced the most realistic conversations." The gap was visible beyond the engineering team. Sales leaders reviewing roleplay transcripts and participating in sessions themselves could tell which model produced the most believable interactions. In several cases, customers chose Yoodli specifically because of the roleplay quality.
Balancing intelligence and speed for real-time practice
Yoodli uses a combination of Claude models, each matched to the demands of a specific task. Haiku 4.5 handles the real-time conversation, where speed is critical for maintaining natural back-and-forth. Sonnet and Opus handle feedback, analysis, and scoring, where deeper reasoning produces better results.
When a rep starts a session, the system loads the scenario, persona, and evaluation framework defined by the organization. The AI persona might be a skeptical buyer, a frustrated customer, or a direct report in a difficult performance conversation. During the roleplay, the system converts speech to text, generates the persona's next response in real time, and tracks signals like pacing, filler words, and conversational structure.
On top of Claude's conversational foundation, Yoodli's engineering team built persona design, evaluation rubrics, conversation memory, and integrations with customer-specific sales methodologies like Sandler and MEDDIC (structured frameworks that define how reps qualify and advance deals). "Claude provides the core conversational intelligence, and our engineering work shapes it into a structured, realistic practice experience," Sessions explained.
The hardest technical challenge was the tradeoff between intelligence and latency. Earlier small models didn't meet the bar for realistic conversation, so the team chose Sonnet for its stronger conversational intelligence and engineered around the latency tradeoff, optimizing time-to-first-token and building fallback strategies to keep voice conversations feeling natural. Haiku 4.5 resolved the tradeoff. "We're finally able to achieve the right balance of strong conversational intelligence and low latency for real-time practice," Sessions said.
The outcome
Scalable practice that changes performance
The combination of realistic AI conversations and consistent, structured feedback has produced measurable results across Yoodli's customer base. At Snowflake, the platform recovered more than 1,600 manager hours per quarter that had been spent running repetitive practice sessions, while achieving 94% participation across more than 3,000 sales reps. Google Cloud used Yoodli to certify over 15,000 partner reps globally, scaling a program that would have been extremely difficult with traditional coaching alone. Across customer organizations, reps who practice three or more scenarios per week close 23% more deals, and new sales hires ramp 40% faster.
The shift to Haiku 4.5 didn't change the quality of conversations, because Sonnet had already set that bar. What it changed was the feel of the interaction. Faster responses keep learners in the flow of the conversation rather than waiting on the system, which makes practice sessions feel more like real calls.
Beyond enterprise sales, the platform has also found traction with individuals practicing everyday communication. "We've had users recovering from strokes, users with workplace anxiety, and users who struggle with stammers all reach out to share how Yoodli has helped them achieve their goals," Sessions said. "For these users, the fact that it's a private, safe space with endless practice has been key." In healthcare and education settings, professionals also use roleplays to prepare for emotionally complex conversations where clarity and empathy are critical.
Looking ahead
Yoodli's roadmap focuses on richer simulations, deeper personalization, and new ways for AI to guide learners through complex scenarios in real time. One emerging opportunity: using Yoodli to help people learn how to use AI itself. Sessions sees AI-powered practice as a natural fit for building those skills hands-on.
"The conversational quality is excellent, especially for realistic roleplay scenarios," Sessions said. "We're continuing to push the boundaries of what real-time AI learning experiences can look like."
"Claude felt the most human and realistic out of the box compared to other models we tested."
CTO, Yoodli