Ebbinghaus documented it in the 1880s, and researchers who repeated his experiments over a century later got the same result: without reinforcement, people forget roughly 70 percent of new information within a day and are down to about 20 percent retention after a month (Ebbinghaus, 1885). A workshop can change how someone thinks about a hard conversation for an afternoon. Whether it changes how they actually handle that conversation three weeks later depends almost entirely on what happens after the workshop ends.

That’s the real problem AI is being asked to solve in training right now, and it’s worth being precise about it. Not “make training more efficient.” Not “replace facilitators.” The actual gap is reinforcement, giving people enough repeated, low-stakes practice that a new behavior survives past the room it was taught in. Ariel’s own reinforcement approach starts from that same premise: a workshop sparks awareness, but consistent practice is what turns it into lasting behavior change.

The research backs up something trainers already knew

Ariel held an internal Ask Me Anything session with a senior solution designer and the co-founder of an AI role play platform the company partners with, and one point they kept returning to matched what the research shows: practice beats lecture, by a wide margin. A 2025 meta-analysis covering 12 studies and more than 900 participants found that role play training produced a large effect on skill development compared to traditional instruction (Fu & Li, 2025). Separate research on retrieval based learning found that testing yourself, actually attempting the skill and getting it wrong, builds longer term retention better than reviewing material passively does.

None of that is new information to anyone who has run a training program. What’s changed is the ability to offer that kind of repeated, low-stakes practice at scale, without needing a facilitator in the room every single time. That’s the actual opening for AI, and both speakers were careful to frame it that way. Not a shortcut around human coaching. A way to make the parts of coaching that require constant repetition available more often than a coach’s calendar allows.

Keeping a person in the loop isn’t a compromise, it’s the design

The term the Ariel team used repeatedly was hybrid intelligence: a human sets up the skill and stays reachable throughout, while AI handles the repetition. This isn’t a hedge or a way of soft launching AI cautiously. It reflects something specific about how soft skills get built.

An instructor led session does something a private tool cannot. It introduces the skill, gives someone a real person to ask questions to, and lets them try a new behavior in front of others who can react in the moment. Some people are genuinely ready to do that from day one. Others need to fail privately first, a dozen times if that’s what it takes, before they’ll risk trying the same thing in front of colleagues. A program that only offers the classroom version has already lost that second group.

What AI adds is a place for that private repetition to happen. The Ariel solution designer described it as a role play tool where people can run the same difficult conversation as many times as they want, get coached, and try again, with no recording saved and no audience beyond themselves. Ariel’s own AI Practice Portal is built around exactly that idea: a secure space where teams rehearse high-stakes moments and get real-time feedback before they walk into the real thing. That combination, a strong classroom foundation plus unlimited private practice, is what neither piece can do on its own.

Accountability doesn’t happen by accident

Here’s a problem that has nothing to do with the technology itself. Inside a live session, accountability is automatic. People are visibly there, participating or not. The moment the session ends, that visibility disappears unless someone deliberately rebuilds it.

The Virtual Sapiens co-founder made a point worth sitting with: a genuinely useful AI tool can still fail if nobody designs for adoption. Managers need enough visibility to ask a real question about someone’s practice, not necessarily every score behind it, just enough to know whether the work is happening. Individual contributors need reminders built into the system, because good intentions fade fast once a workshop ends.

Even framing matters more than most programs give it credit for. Call an assignment “pre-work” and most people treat it like optional homework. Call the same activity a baseline assessment, positioned as the first step in something the company already invested in, and people show up differently because it feels like the start of something rather than a chore attached to it. That reframe alone appeared to do more for actual usage than any feature of the platform itself.

Trustworthy feedback has to be built, not assumed

An attendee asked a fair and pointed question during the session: can AI generated feedback ever be 100 percent accurate? The honest answer, and the one given, was no. No human coach hits 100 percent accuracy in a single conversation either. The standard worth building toward is the same one applied to people: reasonable, professional, and consistent.

Getting there takes real work that’s easy to skip. It means building a specific rubric for each scenario, defining what strong performance looks like, what’s mediocre, and what needs improvement, tied to what that particular person is actually trying to develop. It means testing the AI by deliberately pushing conversations off script, because a tool that only handles the polite, expected version of an interaction isn’t ready for the messy version that shows up in real life. And it means the system has to be willing to say a session went poorly if someone didn’t genuinely engage with it. A tool that hands out a passing score just for participating isn’t measuring anything, and everyone using it will figure that out fast.

Reinforcement outlives the trainer

The forgetting curve doesn’t pause because a live trainer isn’t available anymore. If anything, the weeks right after a workshop are exactly when reinforcement matters most, and exactly when most programs go quiet. This is the piece hybrid intelligence is actually built to solve. Nudges remind people how much practice access they still have. Coaches or managers can be added to a cohort for ongoing visibility into who’s using the tool and how often. And AI reinforcement doesn’t have to be the only option. Post program coaching by phone or email still has a place for organizations that want it.

None of this is about AI outperforming a human trainer. It’s about giving a skill somewhere to keep developing after the trainer has moved on to the next room, which, based on what the research says happens to untouched learning after 30 days, is exactly when it needs the most help staying alive.