Farsight AI's co-founder would grade enterprise AI on how much output survives review

One premise carries the argument, and Dutta states it without benchmark scores or comparative model results in support ^[19]. He wrote that "models are advancing so quickly that raw intelligence is becoming less of a differentiator and more of a commodity" ^[5]. A leader asked to cut a model-evaluation budget this quarter is being asked to take the convergence claim on the author's reading of the market.

Acceptance rate is the one proposal in the column with a number attached. Dutta wrote that "work that's only 60% usable often has very little value" ^[14]. He also wrote: "If employees have to discard almost half the output and start over, the anticipated productivity gains quickly disappear" ^[15]. Sixty percent usable is forty percent discarded ^[16], which he glosses as almost half. Acceptance rate becomes a quarterly number only once a team fixes what it counts against and what score counts as passing.

The pre-deployment step is the part that binds. Dutta wrote: "Without clearly defined objectives, nearly any result can be interpreted as success or failure" ^[11]. He asks for one or two business metrics tied to the problem, established before AI enters the workflow ^[10]. A metric named in advance can be missed, and a team that names one accepts the possibility of failing on it. Most organizations start with time savings, Dutta says, because it is relatively easy to quantify ^[9].

Dutta is CEO and co-founder of Farsight AI ^[1], and the piece ran on Forbes under the Tech Council banner ^[17]. Acceptance rate is what his own firm measures in financial services rather than an agreed industry standard ^[13]. None of that settles whether the metric works. Acceptance rate can be tested inside a single workflow in a quarter, and a company that runs the test can publish a number that contradicts him.

Nothing in the framework requires believing models have converged. A procurement team can keep comparing models on cost, latency and quality while the finance case moves to business outcomes, operational impact and return on investment ^[7]. The two questions have different owners and different renewal dates. This quarter's decision is which number a sponsor agrees to be judged on. The decade's question is whether frontier model quality stays close enough that the choice stops mattering. Dutta's third measurement dimension bets that it will. He argues that as foundation models keep improving, an organization's own expertise becomes more valuable, and he asks buyers to measure how effectively AI uses that institutional knowledge ^[18].