Build1 publisher2 min readPublished
UST's reviewers rewrite or reject 40% of Codon's output before a customer sees it
The rate comes from UST's chief solutions officer, who called it healthy for the product's current stage. It sits inside the same pipeline UST credits with halving the time from a defined requirement to a deployable change.
The Engineer · Build desk
What happened
- UST chief solutions officer Kailash Attal told RuntimeWire in written answers that the company's experts rewrite or reject roughly 40% of the work its Codon workspace generates, a rate he called healthy for the product's current stage.
- UST launched Codon on September 1st as an AI workspace for building and changing applications on ServiceNow, SAP, Salesforce, Oracle Fusion and Workday from a plain-English requirement.
- UST says the workspace cut the time from a defined requirement to a reviewed, deployable result by about 50% against its own historical benchmarks for similar ServiceNow and Salesforce engagements.
- Every tool call is classified as a read or a write operation, and anything unknown or ambiguous fails closed, according to UST.
- Attal said a production pilot went live on September 9th and was moving toward a 50-user readiness gate; UST has not identified that customer.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- contradiction UST's two speed claims sit an order of magnitude apart: the launch release's single application, about a week against an estimated three to four months, works out at roughly 13x to 17x, while the average over more than 50 builds works out at 2x.
- cost Attal said UST absorbs excess AI and platform costs on managed or outcome-based engagements, so on those contracts the rework behind the 40% figure lands on UST's margin before it reaches the customer.
- constraint Write attribution only reaches individual developers where the target platform supports that identity model. On platforms without it, the trail records the change while the person who asked for it goes unnamed.
- decision An enterprise evaluating Codon is deciding whether to buy review capacity with a cost ceiling, because nothing reaches its approval process until a senior UST practitioner has passed it.
The rejected 40% still has to be built. A senior UST practitioner reviews each proposed change before it enters the customer's own approval and deployment process [3], and roughly 60% of generated work clears that review as produced [6]. UST prices the rest in advance. Codon previews expected AI consumption, implementation effort and total cost before a build begins, combining model token rates with estimates for expert review into a quote with an upper limit [20].
Estimating is where the rework shows up as money. RuntimeWire, which obtained the 40% figure in written answers from Attal, wrote that "Generated code becomes cheaper; governance, review and contractual accountability become the service being sold" [22].
UST's two speed numbers are not independent. RuntimeWire reports that the roughly 50% reduction covers a specific stage of delivery, uses UST's historical work as the baseline and includes expert intervention [13]. For the figure to transfer, a buyer's baseline has to resemble UST's own historical ServiceNow and Salesforce engagements [9], and its definition of done has to stop at a reviewed, deployable result. The builds measured were internal: more than 50 of them across more than 10 projects, by Attal's account [10].
Inside the workspace, a developer names the target platform, and Codon plans the work, routes it through the relevant Model Context Protocol server and platform software development kit, then assembles a proposed change [2]. Production changes pass through an approved-build phase, and destructive actions require explicit intent [16]. Activity is recorded in what UST describes as a tamper-evident, hash-chained audit trail [17]. Customer environments run single-tenant on AWS Fargate with default-deny outbound access, encryption in transit and at rest, and a warm standby in another region [18].
I would look at the terminal route first. Codon offers chat and terminal interfaces under the same policy and audit engine, with chat aimed at workflows, catalogs, configuration and documentation, and the terminal giving each developer a per-user shell for platform command-line tools and SDKs [19]. In my view a per-user shell inside the same policy engine is harder to build than a governed chat box, and it decides whether the audit trail covers what a developer actually did.
Codon's public materials describe the intended buyer: companies that cannot let a general-purpose coding agent make untracked changes to payroll, permissions, customer records or financial workflows [23].
What to watch
- Whether the September 9th pilot clears its 50-user readiness gate, and whether UST names the customer.
- Whether the 40% rewrite-or-reject rate falls as Codon matures, and whether UST publishes the figure again.
- Whether per-developer write attribution reaches target platforms whose identity model does not support it.