I replaced our sales training with an AI coach that runs in Slack — here’s what broke

At Nutrient, we have a sales wiki that references product names we retired a year ago. We have a university where analytics show most pages have been visited zero or one times in the last quarter. We have onboarding materials spread across Google Docs with varying access levels, Notion pages nobody trusts, and 35-minute video recordings that new hires are still being pointed to despite the fact that someone painstakingly cut them into 37-second clips that — you guessed it — nobody watches either.

We’ve rebuilt our sales enablement three times in three years. Each time, it was great for about two months. Then products changed, people changed, and the content quietly rotted while everyone went back to sending Slack messages to colleagues for information instead.

I was privy to a meeting where our sales ops lead pulled up the Notion analytics and confirmed that, yes, people were clicking the training links. They just weren’t clicking the 37-second clips we’d carefully cut. They were watching the original 35-minute recordings — the ones the clips were supposed to replace.

That meeting convinced me we didn’t have a content problem. We had a system problem. And no amount of reorganizing Notion pages was going to fix it.

So I built an AI coaching agent. It took nine days. Here’s what actually happened.

Why another learning management system wasn’t the answer

Let me be specific about what was broken, because “our enablement sucks” isn’t useful.

In that same meeting, the head of our sales onboarding walked through the current structure. It was genuinely thoughtful: week one covers SDKs and server fundamentals, week two is integrations, week three is workflows, and by week four, you’re shadowing other sales representatives. New hires submit demo recordings at the end of each week. There are assessments tied to certifications. On paper, this is solid.

In practice, the biggest pain point wasn’t structure. It was trust. People didn’t trust that what they were reading was current. The sales wiki had product knowledge views that called back to legacy naming conventions. Process documentation hadn’t been updated since 2024. Sales reps were hesitant to edit Notion because they were worried about “breaking something.” And, well, if you’ve ever told a non-technical person that “you can’t break GitHub, just go mess with code,” you know how that lands.

The deeper issue: enablement was an event, not a system. You get onboarded, maybe there’s a certification at the 90-day mark, and then… enablement never begins. Or never begins again. Our ops lead literally said both phrases in the same breath and laughed.

I didn’t want to build version four of a content library that would rot by summer. I wanted something that would teach people to think about our products — to hear a prospect say “we already have a PDF viewer” and know what to do next. No slide deck teaches that. No quiz tests it.

What I actually built

Coach is a conversational AI agent that lives in Slack. Every day, at a time you pick, it DMs you with a realistic scenario and coaches your response.

Here’s what it’s not: an LMS, a quiz bot, or “ChatGPT answers your sales questions.”

Here’s what it is: a five-minute daily conversation that does what a great sales manager does — asks questions, listens, pushes back when your thinking is shallow, and shuts up when you’ve demonstrated real understanding.

The design principles are as follows.

Teach thinking, not recall — “A prospect just said Y — what do you do and why?” Not “name the five features of X.”

Private by default — Nobody sees your answers. Nobody scores you. This isn’t a performance review dressed up as learning.

Respect intelligence — These are experienced professionals. Coach never says “great job!” when the answer is mediocre. I literally had to write “Coach Carter doesn’t grovel” into the rules file after the AI kept apologizing for giving feedback.

Short and daily — Five minutes compounds. A quarterly training dump decays. We all know this. We just don’t build for it.

The personality file — the soul of the system — has one line I keep coming back to:

“You are not here to deliver content. You are not here to evaluate people. You are here to make their thinking sharper tomorrow than it was today.”

Three AIs walk into a content pipeline

Here’s the part that surprises people: the AI coaching conversations were the easy part. Getting 113 sessions of curriculum right — that was the mountain.

At this point, we should stop calling it “AI” and recognize that it’s still mostly just a person doing the thinking with some very expensive autocomplete helping out. I used three different tools for three different jobs.

ChatGPT for the blueprint — Before writing any content, I needed to get the thinking right. What should we teach? In what order? What does “good” look like for each topic? I spent days in back-and-forth dialogue — not a single prompt, but an extended conversation — building what I call the single source of truth (SSoT): a spec covering coaching philosophy, learning objectives, progression logic, and every design decision. If you can’t define what “good judgment” looks like for a given scenario, you have no business building a session around it.

Nuri for grounding in reality — Nuri is our internal AI agent at Nutrient. It has access to our data, including Gong call recordings. I fed the SSoT to Nuri and had it generate scenarios based on actual sales conversations: real buyer personas, real objections, real deal patterns. The curriculum wasn’t hypothetical. It was drawn from what actually happens on our calls. Generic scenarios are the enablement equivalent of Muzak: nobody learns from them.

Me for architecture and curation — I reviewed everything, restructured it into learning objective clusters, wrote the application questions, defined completion criteria, and made sure each session built on the last. AI generated raw material. I shaped it into something coherent.

The key structural unit is what I call a learning objective cluster — not a topic, not a product area, but a set of learning objectives tied to one real-world decision moment. Example: “Framing the problem before pitching solutions on a first discovery call.” Each cluster defines the role, the decision, what good looks like, and the common failure modes. Then it produces 10–15 daily sessions, each training one micro-judgment.

The litmus test for any session is “What one thing will this person do better tomorrow?” If the answer is vague, the session doesn’t ship.

More than 6,500 lines of curriculum. 113 sessions across 11 learning paths. All grounded in real data. Not a single hypothetical scenario.

The stack is embarrassingly simple

I’m almost reluctant to share this because people expect infrastructure. There is none.

• Platform — OpenClaw (open source AI agent framework)
• AI — Claude Sonnet for coaching conversations
• Delivery — Slack Block Kit messages
• State — JSON files per learner. No database.
• Scheduling — Bash + jq, triggered by cron
• Content — Markdown files in Git

The entire thing is about 20 files and some shell scripts. It runs on a Mac Mini. The infrastructure cost is $0.

The one architectural decision that saved everything: separate the deterministic from the probabilistic. Scheduling? Bash. Progress tracking? Bash. Report generation? Bash. Coaching conversations that require human-like judgment? AI. Everything else? A Markdown file that a human can read and edit without a PhD.

This isn’t just clean architecture. It’s survival. Let me explain.

What broke: The actually useful section

If you’re building anything with LLMs, this is the part that will save you time and money. I learned these lessons so you don’t have to. You’re welcome.

LLMs cannot do cron jobs

I’ll say it louder for the people in the back: do not give an LLM a recurring operational task.

My first version had the AI agent handle scheduling: check learner profiles, determine who needs a session, deliver it. Sounds perfectly natural. For about three days, it worked.

Then it started hallucinating. It would read perfectly valid JSON files and report that all fields were blank. It would flag “corrupted” profiles that were fine. Over the course of a single weekend, it ran more than 200 agent cycles, each burning tokens to produce confident, detailed reports about problems that didn’t exist. Five learners got stuck waiting for sessions that never came.

The fix: 80 lines of bash. It has never hallucinated. It costs zero tokens. It runs every 15 minutes. I wasted an entire weekend and a few hundred dollars learning something that should be a bumper sticker: deterministic tasks demand deterministic tools.

The AI will narrate its own process

“Let me check your profile… I see you’re enrolled in LOC-SDK-SALES-003… Let me load today’s session content…”

This is the LLM equivalent of a waiter explaining their route to the kitchen. Nobody asked. It took multiple rounds of increasingly aggressive instructions to kill — from “avoid technical language,” to “NEVER announce what you’re doing,” to literally banning specific phrases.

Not guidelines. Bans. And you will add more after launch. My Git history is basically a catalog of things I had to explicitly forbid Coach from doing.

Affirming without closing

This one is insidious because the conversation looks correct to a human observer.

Coach says: “Great answer! You’ve clearly demonstrated the concept!” Then… doesn’t run the completion script. The learner thinks they’re done. The system thinks they’re still in session. No new session arrives the next day. The failure is invisible until someone messages you asking why their coaching stopped three days ago.

The fix is what I call the atomic close protocol: affirmation + completion script + next-session preview must happen in a single turn. No separation. No “I’ll close this in a moment.” I enforce this at three levels: coaching rules, delivery rules, and the personality file. Belt, suspenders, and duct tape. The AI will still try to wriggle out of it.

PII routing will bite you

An admin-only summary report containing every learner’s name and progress got delivered to a learner’s DM thread. Why? The agent’s default delivery target was “wherever the last conversation happened.” The last conversation was with that learner.

Imagine explaining that one to your security team.

Every message must explicitly specify its destination. No defaults. No “reply in the current thread.” Pin every output to a named channel. If your AI agent framework has a concept of “current conversation context,” treat it as hostile for any message containing PII.

The AI desperately wants to be nice

This is the one that never fully goes away.

Despite 200 lines of personality instructions telling Coach to be direct, push back on shallow thinking, and never over-praise, it constantly drifts toward validation. “That’s a wonderful point!” sneaks back in. “Excellent thinking!” appears uninvited. The LLM’s deepest instinct is to please, and that directly conflicts with effective coaching.

If you’ve ever worked with that manager who tells everyone they’re doing great and then wonders why nothing improves, that’s what an unconstrained LLM coaching agent becomes. Personality definition isn’t a prompt engineering afterthought. It’s a first-class architectural concern. And it requires ongoing maintenance, because the drift never stops.

The numbers

• 9 days from first commit to production
• 113 coaching sessions across 11 learning paths
• 6,500+ lines of curriculum
• 11 pilot learners in the first two weeks
• ~5 minutes per day per learner
• $0 infrastructure — just a Mac Mini, Slack API, and one AI model
• 3 AI tools in the pipeline: ChatGPT (ideation), Nuri (content generation), and Claude (coaching)
• 80 lines of bash replaced an entire LLM-based scheduling agent

What I’d do differently

Start deterministic from day one — If a task has the right answer, don’t give it to an LLM.

The curriculum is 80 percent of the work — The coaching AI wrapper is the last mile. Getting the SSoT right, generating content from real data, and structuring it into a coherent learning experience — that is where you should spend your time. Great curriculum with a mediocre AI beats mediocre curriculum with a great AI every day of the week.

Budget half your time for teaching the AI what NOT to do — Every fix is obvious in hindsight. None of them are obvious in advance.

The 80/20 rule applies — And so does the corollary: you will spend 80 percent of your debugging time on the 20 percent of problems that the AI creates by trying to be helpful.

The real unlock

This isn’t “AI replacing trainers.” It’s AI delivering what a human designed — at a scale and cadence that humans can’t sustain. Daily, personalized, in five-minute conversations that fit into real workdays.

One person with domain expertise, an open source agent platform, and nine days built a coaching system that scales across an organization. No ML team. No training data. No fine-tuning. Just Markdown files, shell scripts, and very clear instructions about what the AI should never do.

But here’s the insight that matters more than the tech: enablement fails not because companies lack content, but because they lack a system for continuous reinforcement. We had the content. Three times over. What we didn’t have was something that would deliver it daily, adapt to each person, and survive longer than two months without human maintenance.

An AI agent is a compelling delivery mechanism for that system. But only if a human designs what goes inside it. The real unlock isn’t the AI; it’s deciding that enablement is a system, not an event, and then building accordingly.

Nine days. A few spectacular failures. A sales team that now spends five minutes a day getting better at the part of their job that actually matters: thinking.

And a 35-minute OCR video that, mercifully, nobody has to watch anymore.