How a regional media company leveled up their ad sales role play training with a custom AI training platform.
Consolidated Publishing, a regional media company operating multiple newspapers and magazines across Alabama, was investing in a new wave of sales talent. Onboarding new ad sales representatives meant getting them comfortable with the company's consultative selling approach and the best way to do that has always been role play practice.
The problem? Sales teams can only practice with the same colleagues and managers so many times before the sessions lose their edge. The scenarios become predictable. The feedback starts to sound the same. And new reps who need the most repetitions are often the least comfortable asking for more practice time. Plus, not everyone has time to help a colleague role play an upcoming call when they need to practice.
Sales reps needed role play practice to build confidence, but this wasn't always possible with a small and busy team.
Realizing that AI could be the practice partner, with real local context for custom training sessions.
While leading an AI workshop onsite for their sales team and learning more about their process, I saw an opportunity. I'd already introduced the team to AI-powered prospect research and built them a prompt library for pre-call preparation. When the publisher and sales manager mentioned how central role play was to their training philosophy, and how it tended to plateau, I realized AI could be the perfect role play partner.
What if the AI could play the prospect? Every call could be different, every industry accurate, every objection realistic, and it could be available whenever a rep wanted to practice.
They asked me to work up a full proposal. I got it to them within days. The publisher asked if I could compress a four-week build into three to keep the momentum going. I said yes.
I want to be honest about something: when I started this build, I had never used Voiceflow or Make before. I had no coding background and no formal training in AI system development. What I had was a clear vision of what the tool needed to do, a deep familiarity with AI as a collaborator, and a willingness to learn systems.
What still surprises me is how fast I was able to learn and build using AI as my teammate.
Mapped out the full system, including agents, training modes, difficulty levels, and feedback loops, before writing a single prompt.
Built 23 documents across 6 folders, including local business profiles, publication data, sales philosophy, and regional demographic data, to ground each agent in real context.
Built and iteratively trained all 10 agents, each requiring 10-30 testing rounds to reach accurate, authentic role play behavior.
Connected Voiceflow to Google Sheets via Make, solving a tricky JSON-to-structured data problem that took days to crack.
Leadership team tested from November 7–19, providing real sales manager feedback that shaped the final calibration.
Full sales team rollout December 1, 2025. Phase 2 commissioned within weeks. Still in active production use today.
I was learning these platforms in real time, using AI to turbocharge my learning curve. What we were building together kept exceeding what I thought was possible.
SAL-E (Sales Agent Learning Environment) was built entirely using no-code platforms, primarily Voiceflow for the conversational AI interface, with Make handling the API connection to a Google Sheets tracking dashboard. I used Anthropic's Claude as my technical thinking partner throughout: working through architecture decisions, troubleshooting logic problems, and iterating on agent prompts in real time. The build ran from October 19 to November 7, 2025. The complete prototype with a full intake flow, three difficulty agents, feedback coaching, and session logging was finished on Halloween night after two focused weeks of work.
One of the most technically demanding challenges wasn't the logic flow of the system, it was getting the system to reliably log every session. Voiceflow was sending the full transcript and coaching feedback as one long JSON string; Make was expecting a structured object. Initially, I didn't know what either of those things meant. Getting these two platforms to speak the same language required days of iteration and problem-solving. When it finally clicked, the session tracker began capturing everything automatically, including rep name, scenario selected, difficulty level, call outcome, and the full coaching feedback, without any manual input from the team.
After three weeks of leadership beta testing and refinement, SAL-E rolled out to the full sales team on December 1, 2025. In January 2026, the client commissioned a Phase 2 expansion. This involved a Reverse Agent mode and a dedicated Classified Ads training pathway, both delivered in three additional weeks.
Getting a conversational AI agent to sound like a real Anniston restaurant owner — skeptical, time-pressed, grounded in local concerns about foot traffic and competition from chains — didn't happen on the first try. Or the fifth.
Each difficulty agent went through 10 to 15 rounds of testing and refinement before it felt right. The process was iterative and collaborative: I'd run a full practice session, take detailed notes on what was working and what wasn't, including verbatim examples from the transcript, then bring those notes to Claude to diagnose the problem and suggest prompt adjustments. Claude would generate an updated prompt, I'd deploy it in Voiceflow, and we'd run the scenario again. Round by round, the agents got sharper.
Core behavior testing + exit condition calibration for each of the 10 agents
The early "wrong" outputs looked like this: the Warm Lead agent would commit to buying an ad after just two exchanges with no real pressure, no authentic hesitation. Then overcorrection: the Challenging Prospect would pepper reps with questions about digital ad CPMs and ROI benchmarks that no real small business owner would ever ask. Neither version felt true to what sales agents encountered on real sales calls. The calibration came from working directly with the client's sales manager who helped me understand what reps actually encounter in the field and translating that reality into precise behavioral instructions for each agent.
A separate challenge was conversation length and natural endings. Early agents would never end the call. They'd always find one more question to ask, one more detail to probe. I had to engineer explicit exit conditions, essentially training each agent to recognize when a sales call had reached a natural conclusion, rate the outcome, and hand off gracefully to the Feedback Coaching Agent. Each difficulty level has its own logic: Warm Lead sessions run shorter, Challenging Prospect calls run to 10 or more exchanges. Factoring in exit condition testing, each difficulty agent required close to 25 to 30 total rounds of testing and development before reaching production quality.
The Feedback Coaching Agents required 5 to 6 rounds of their own testing, mostly tuning output format and ensuring the agent cited specific moments from the transcript rather than giving generic advice. The goal was feedback that reads like a senior sales coach wrote it, not a chatbot. By the time we were done, it was referencing the client's consultative selling philosophy by name and pulling exact quotes from the session to illustrate coaching points.
The difference between a generic AI role play tool and SAL-E is the Knowledge Base: 23 documents organized into six folders that give every agent the context it needs to simulate an authentic, locally grounded conversation.
When a rep selects "Restaurant — Anniston — Challenging Prospect," the agent isn't improvising a generic skeptical business owner. It's drawing on a detailed Business Owner Brief that describes what it feels like to run a business in Anniston: the cautious optimism about downtown revitalization on Noble Street, the competition from Oxford's retail corridor, the Anniston Army Depot's 3,700 steady-income workers as a customer base, the JSU students who drive down from Jacksonville on weekends. The prospect sounds like they actually live there, because the AI knows what factors into a local business owner's decision making process.
The Industry Brief layer adds another dimension. Each brief covers the real local competitive landscape: specific businesses by name, the seasonal patterns that hit hardest, and the exact objections a local owner is most likely to raise. The agent knows what this owner is worried about before the call even starts. The rule baked into every agent's instructions: use this information to be an authentic local character, never to recite data.
Location-specific character profiles for Anniston, Oxford, Pell City, and Talladega, each with personality, history, and real local context.
County-level economic data, community demographics, and market context for each coverage zone.
One per industry — restaurants, automotive, healthcare, retail, and more — with local businesses by name, seasonal patterns, and common objections.
Rate cards, subscriber numbers, and audience demographics for the Anniston Star, Daily Home, St. Clair Times, Lakeside Living, and Foothills Magazine.
Full advertising product database — formats, placements, pricing, and packages.
The consultative selling framework used by the Feedback Coaches to evaluate each practice session.
SAL-E is built in Voiceflow and accessible in any browser, no app download or IT setup required. Reps open the tool, build their scenario, and jump into a realistic sales call anytime, anywhere.
Select Sales Call Practice, Reverse Agent Mode, or Classified Ad Sales, each with its own dedicated training pathway.
Warm Lead, Cold Call, or Challenging Prospect for progressively harder scenarios that mirror real sales situations.
SAL-E includes ElevenLabs voice mode so reps can practice out loud to nail their tone and phrasing just like a real phone call.
The AI plays a fully contextualized business owner drawing from local economic data, industry briefs, and realistic objection patterns.
A dedicated Feedback Coach agent reviews the full transcript, cites specific moments, and ties observations to the company's consultative selling philosophy.
Every session logs to a Google Sheets dashboard, giving sales reps and managers visibility into rep activity, outcomes, and feedback trends without any manual reporting.
SAL-E's depth comes from a coordinated network of specialized agents, each built and tuned for a specific role. No single agent does everything, each one is focused, calibrated, and connected.
Three distinct difficulty levels — each a different local business owner persona with realistic objections and decision-making patterns.
Reviews the full transcript and delivers structured coaching tied to the client's consultative selling framework.
Same difficulty progression as Sales Call Practice, but focused on classified advertising conversations — different products, different objections, different buyer psychology.
Mirrors the Sales Feedback Coach but calibrated specifically for classified advertising conversations and buyer dynamics.
The roles flip — the AI plays the sales rep and the human plays the business owner. Added in Phase 2 after a sales rep suggested it at launch. Now one of the most-used features.
Highlights the AI's own sales tactics after the call, turning the session into a learning moment about technique.
Every agent is powered by GPT-5 mini and draws from the same curated Knowledge Base of 23 documents across six folders including Business Owner Briefs, Coverage Area Intelligence reports, Industry Briefs for eight business categories, Publication details, and the client's Sales Philosophy guide. The Knowledge Base is what makes each scenario feel grounded and specific rather than generic.
Difficulty levels were carefully calibrated through extensive testing to feel genuinely challenging without being discouraging and to allow reps to practice and grow at their own pace.
Feedback from Consolidated Publishing's leadership team consistently pointed to one thing: the level of customization. This wasn't a generic sales training tool. It knew their publications, their markets, their selling philosophy, and their team's specific development needs. The sales manager described the coaching feedback quality as exceptional, noting that it incorporated their consultative selling techniques in ways that felt like their own training voice.
The Feedback Coaching Agent's output, which referenced specific quotes from the transcript, identified what the rep did well, and suggested exactly what to practice next, was something the team hadn't seen from any training tool before. It reads less like an AI evaluation and more like a debrief with a senior colleague who was on the call.
The leadership team was enthusiastic about SAL-E from the first beta session, specifically calling out how well the system reflected their actual sales process and local market context.
The sales manager was particularly impressed by how the Feedback Coach incorporated the company's own consultative selling techniques — not generic sales advice, but their approach.
A sales rep at the launch session suggested Reverse Agent Mode. It was built into Phase 2 and quickly became one of the most-used features in the entire system.
Within weeks of the December launch, the client commissioned Phase 2 — adding Reverse Agent Mode and a full Classified Ads pathway for an expanded sales team.
Reps practice against AI prospects that know Anniston's downtown revitalization, Talladega's Honda plant workforce, and the exact objections a Pell City restaurant owner is likely to raise. Every call mirrors the real world.
This production-ready system was built entirely using no-code platforms by a single consultant. It demonstrates what's possible when AI and no-code tools are combined with the right idea.
"Our experience was extremely positive — so positive that we have engaged Thomas for other projects, which have proven to be just as effective."
— J.C. Blucher Ehringhaus III, President and Co-Publisher, Consolidated Publishing Company
SAL-E started with a conversation during a workshop. What can we solve in your business with a custom AI tool? Let's connect to find out.