Victor Riparbelli is the CEO and co-founder of Synthesia, a leading enterprise-focused AI video platform.gettyA decade ago, if you walked into a newsroom like BuzzFeed or Business Insider, you would see a giant leaderboard measuring each journalist in real time by the number of pageviews their stories generated. Besides increasing anxiety among reporters, this measurement method incentivized editors to rely on clickbait and virality more often than they would have before. It brings to mind the famous quote from Charlie Munger, “Show me the incentive and I'll show you the outcome.”Today, as the traffic these publications receive from Google and social media has declined significantly, this practice is mostly no longer in place. Media businesses have moved toward more strategic outcomes that drive longer-term value for their business, such as paid subscribers, event attendees and sponsorship revenue growth.History Repeating ItselfAI adoption is facing a similar transition inside companies. Earlier this year, “tokenmaxxing” peaked with software engineers and even other roles. Companies were recognizing employees simply by how many AI tokens they used. It was the vanity metric moment for AI, similar to the pageviews metric moment for media. Unlike the media, which remained fixated with pageviews for years, tokenmaxxing lasted only for a few weeks until most COOs realized that it was deeply unsustainable.Since starting Synthesia back in 2017, one of my core principles has been utility over novelty. I’ve challenged my teams to meet customer needs in the most sustainable way possible. So, from my perspective, it was quite obvious that tokenmaxxing would be a short-lived tech trend. The real competitive advantage for businesses is building measurement systems that tie AI to business outcomes, without incentivizing people to game the system.What I’m seeing in AI-native technology companies is that you measure outcomes differently in different areas of the business. In some departments, there are clear and obvious metrics to track. For example, in customer support, tickets resolved can give you clear insight into progress, as well as cost-to-serve and quality assurance (QA) scores. In corporate training, you can move from simply looking at completion rates to measuring behavior change and skill development over time.The AI Impact ScorecardA simple scorecard can track AI impact across three layers.​1. Efficiency measures time and cost per unit of work. An example here could be reducing the time to roll out a sales training program from six weeks to three days, or dropping the cost per training video from $5,000 to under $500. Many companies are already doing this. Our client, KONE, the global elevator company, changed the way it trains its field technicians and saved a full day per course using pre-learning videos while producing content 30% faster.2. Quality moves beyond time and cost savings to behavior change: Did the learner actually understand the material? Can they apply it in a meaningful way to drive impact? Risk is about knowing how someone applied their knowledge and why.3. Compliance and legal teams need to be able to audit why an employee used AI to draft a customer contract, verify guardrails were in place and ensure the output didn't violate regulatory requirements. Our own legal team built an AI legal avatar for common contract management tasks like scheduling calls, answering questions and talking through points with outside counsel. With strong guardrails in the training, testing and oversight, it's been working well.On top of that, there are three principles that can prevent another tokenmaxxing moment. First, reward verified outcomes and impact, not volume. Measuring "problems solved with AI assistance" and tying it to outcomes like faster decisions, higher customer satisfaction scores and lower rework rates forces honesty. Second, sample for quality, not just quantity. A legal team using AI to draft contracts should QA-check outputs for accuracy. One bad clause could cost millions, so volume doesn't matter. Third, measure at the workflow level, not the model level. An AI model's accuracy score is irrelevant if the workflow it sits in doesn't drive better outcomes.Moving OnJust as the early days of websites and social media had vanity metrics of page views and likes, the early days of AI had token leaderboards. It's good to see the industry quickly moving on.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?