Pigment's prompt-built forms inherit the row-level permissions of the model beneath them

At a workshop with Uber, according to Pigment's head of product and AI strategy Ben Previeux, participants drew the application they wanted on a whiteboard, photographed the sketch and sent it to the Modeler Agent, which converted the image into a Frame the group could refine before asking the agent to build the back end ^[5]. Previeux puts that at about an hour for something that might otherwise have taken several weeks ^[5]. Read "several weeks" as three at 40 hours and the comparison is 120 hours against one, a 120-fold gap sitting on a counterfactual nobody in the room clocked ^[15].

What is being sold is the generation. What someone does on Monday is the verification. "We ask the user to verify because there is no 100% guarantee that something built by AI will work," Previeux said ^[7]. The genuinely useful piece for the person who has to answer for a number on Friday is narrower than the prompt box: the reviewer can see which Pigment objects supply each individual metric ^[6].

Pigment's differentiation claim is that standalone generative tools can draw a dashboard or an app without reaching live, governed enterprise data ^[2]. What carries that in practice is a security primitive rather than a model: the Frame inherits existing access rights, so a country manager opening a shared revenue application sees only the countries they are authorised for ^[8]. The same mechanic cuts the other way. A Frame faithfully redistributes whatever the access model already says, which means quota or region rights that are wrong in Pigment are now wrong inside a form that thousands of people type into.

On outcomes the record is thin. Pigment names Uber, Glean, LinkedIn, Carta, Unilever, Supercell and Vinci Concessions as customers already using Frames ^[9]. One of those seven is quoted describing a result: Glean's VP of finance and business operations Michael Miao says a manually created business-critical analysis became a live interactive experience where hiring assumptions can be adjusted and the effect on sales capacity calculated immediately ^[10][17]. There is no count of Frames published, no measure of whether a team opens one a second time, and no failure rate at the publishing gate ^[18]. Time-to-value here rests on a single vendor anecdote, and depth of use is not measured at all.

Two questions decide whether this is worth rolling out where you work. First, is the row-level access model in Pigment already correct, because Frames honours it rather than corrects it ^[8]. Second, who is named as the reviewer of agent-written formulas, and does that person's week contain the hours the publish check needs ^[6]. Pricing sits under both: Frames is included in the base platform at no extra charge while model usage is billed through credits ^[13], so building is the cheap part and iterating is the metered part, and a team that prompts five versions of a hiring form pays for five. Whoever owns the access model owns the rollout, and that is rarely the person who wrote the prompt.