Token Leaderboards Are The New Pageview Board, And Your Next Budget Will Fund Them

Victor Riparbelli, chief executive and co-founder of the enterprise AI video company Synthesia, argues in a Forbes Technology Council post that companies measuring AI adoption by tokens consumed have imported the media industry's pageview mistake into the enterprise ^[1][2][3]. The stake is not vocabulary but budget: a usage counter is the easiest number to put in a board pack, and therefore the fastest way to fund work nobody checked.

His analogy is concrete. A decade ago newsrooms including BuzzFeed and Business Insider ran real-time leaderboards ranking journalists by the pageviews their stories generated, which he says raised reporter anxiety and pushed editors toward clickbait and virality ^[4][5]. As search and social traffic fell, those businesses moved to paid subscribers, event attendees and sponsorship revenue ^[6]. He reaches for Charlie Munger: show me the incentive and I will show you the outcome ^[7].

The enterprise version, per Riparbelli, was "tokenmaxxing", which peaked earlier this year among software engineers and other roles, with companies recognising employees by how many AI tokens they consumed ^[3]. He calls this the vanity metric moment for AI and says it lasted only a few weeks before most COOs judged it unsustainable ^[8][9]. That timeline is an assertion, not a measurement, and it comes from a vendor with an interest in the market having matured. Seat activation counts and prompt volumes, which live in every AI dashboard sold today, are the same species of number and are not going away in weeks.

The proposed replacement is a three-layer scorecard: efficiency as time and cost per unit of work, quality as demonstrated behaviour change, and compliance as the ability to audit why an employee used AI on a given output and whether guardrails held ^[10]. The efficiency examples are vendor-shaped: rolling out a sales training programme in three days instead of six weeks, and cutting cost per training video from $5,000 to under $500 ^[11]. Those are reductions of roughly 93% in elapsed time and at least 90% in unit cost ^[12][13]. The named case is KONE, the elevator company and a Synthesia client, which he says saved a full day per course with pre-learning videos while producing content 30% faster ^[14]. Treat it as supplier-reported, because it is.

The durable material is elsewhere. He suggests customer support already has usable measures in tickets resolved, cost-to-serve and QA scores, and that corporate training should move past completion rates to skill development over time ^[15]. Three principles follow: reward verified outcomes rather than volume, sample for quality rather than quantity, and measure at the workflow level rather than the model level ^[16]. The last is the one worth arguing in a planning meeting, because a model accuracy score means nothing if the workflow around it produces no better result ^[17]. The legal example is sharper still: a team drafting contracts with AI should QA-check outputs, since one bad clause can cost millions and volume is irrelevant ^[18]. Synthesia's own legal team runs an AI avatar for contract management tasks including scheduling and discussions with outside counsel, with guardrails in training, testing and oversight ^[19].

Watch what enters next year's targets. If token counts, seat activations or "problems solved with AI assistance" arrive as self-reported figures with no sampling regime, the incentive is already set ^[16]. The honest test is whether finance can name the unit of work, the baseline cost per unit, and who audits a sample of outputs before the number reaches the board.