Loading today’s stories
Build1 distinct publisher3 min readUpdated
Jarvis Pro settles routing, scope and metric definition before generation, on the premise that a fluent wrong answer is worse than none when a human will repeat it to a merchant.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Grab's engineering blog has described how it built Jarvis Pro, an AI assistant for the company's account managers, around a pipeline that classifies the request before it generates anything [1][7]. The design came out of a specific failure: an early prototype returned a polished answer that named the merchant, summarised the week and recommended pushing promotions before the next review, and it was wrong [2][3].
The merchant's order volume had fallen, but the sharper problem was operational, with more outlets paused and fulfilment slipping, so sending more demand into that setup would have made the merchant look worse [4]. That is the cost structure that justifies the architecture. The account manager is not the last reader of the answer; they carry it into a merchant conversation [1], and Grab's own framing is that when scope, access, diagnosis or metric definition go wrong, confidence becomes a liability [27].
The inputs are not dashboard queries. Account managers ask things like "I am meeting this merchant tomorrow. What should I tell them?" [5], and those questions hide decisions about scope, access, business diagnosis and metric definition [6]. A dashboard answers a bounded question, while a review question has to diagnose what the work actually is [21]. If the dashboard, the warehouse table and a local spreadsheet disagree, the first task is metric reconciliation, not recommendation [22].
So the launch router is a prompted LLM classification step constrained by a routing guide that enumerates supported task types, allowed context, answer shapes, and the conditions that should force a handoff or a refusal [11]. A route is not a label: it sets structure, metrics, tool path and guardrails, and only the context needed for that task is loaded [12][23].
Memory got the same narrowing. Here memory means saved user context such as country, role, portfolio scope and preferences [17], which helps when someone says "my portfolio" but becomes risky when saved defaults override explicit context, missing identity turns into a guess, or a request crosses user boundaries [18]. Memory resolution therefore sits before answer generation and is restricted to a small set of outcomes: proceed, ask for onboarding information, prefer explicit context, use saved defaults, look up portfolio scope, or refuse [19]. Authorization stays in backend permissions and row-level controls [20]. The model is not the gate.
The numbers are modest in scope and Grab labels them as such. In offline evaluation, routing matched the expected safe route for 99.4% of 351 realistic prompts drawn from labelled eval sets from the first half of 2026 [8], which works out to roughly two prompts off-route [24]. A focused portfolio and brand answer-quality suite moved from an average of 78.5 to 91.0 [9], a gain of 12.5 points or about 16% relative [25]. Grab states these come from offline launch-readiness evaluation only and are not a measure of production performance or business impact [10].
The worked example in the post is explicitly illustrative, using no live or non-public data, and shows a route that scopes to one merchant in Singapore, loads an approved metric catalog, and carries guardrails against private competitor metrics and unsupported projections [13]. In that fictional case net sales fall 7.8% week-on-week and orders 9.1% while average order value rises 1.4% [14], with pause rate up from 2.1% to 4.8%, more than double [15][26], and fulfilment down from 94.6% to 91.8% [15]. Grab says the route prevented three recurring mistakes: jumping to discounts, inventing competitor spend, and answering from a generic growth template [16].
Worth watching: whether the 99.4% offline routing figure holds against the live prompt distribution, what the production refusal and handoff rates look like, and whether a hand-maintained routing guide stays coherent as supported task types multiply.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
In an internal offline evaluation, routing matched the expected safe route for 99.4% of 351 realistic prompts drawn from labelled eval sets from the first half of 2026.
In a focused portfolio and brand answer-quality suite, the average score moved from 78.5 to 91.0.
In the illustrative answer extract, net sales fell 7.8% week-on-week and orders fell 9.1%, while average order value increased 1.4%.
In the illustrative answer extract, temporary pause rate rose from 2.1% to 4.8% and fulfilment fell from 94.6% to 91.8%.
Grab says the route prevented three common mistakes: jumping straight to discounts, inventing competitor spend, and answering from a generic growth template.
A dashboard answers a bounded question such as "show net sales for merchant X last week", whereas an account review question has to diagnose the work to be done.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Detailed but wholly first-party and partly synthetic
The single source is unusually specific for a vendor engineering post: a named design principle, the router's implementation as a prompted LLM classification step under a routing guide, an enumerated memory decision set, a stated authorization boundary, and two quantified offline suites with sample size for one of them. That specificity is offset by structural limits: one publisher with no independent verification, a worked example the post itself labels illustrative with no real data, no production measurements, and an accepted-route-accuracy definition that is cut off mid-sentence in the supplied text, so the headline 99.4% cannot be fully audited.
Single-org internal system at launch-readiness stage
The only adoption signal is Grab's own disclosure that it built Jarvis Pro for its account managers and evaluated a 'launch router' for launch readiness. That establishes one internal deployment context but nothing about breadth: no seat counts, query volumes, rollout phases, geographies beyond an illustrative Singapore example, or reuse of the pattern by any other organisation. Adoption is therefore real but minimal and unmeasured in production terms.
Close to aligned, with self-imposed caveats
The post is unusually disciplined for vendor content: it leads with its own failure, twice states the numbers are offline launch-readiness only and not business impact, and labels the worked example synthetic. The small positive gap reflects two overreach risks rather than framing excess: a 99.4% headline whose failure count is about two prompts is presented as a strong safety signal, and the 'route prevented three common mistakes' claim is asserted from an illustrative example without any measured before/after failure rate.
First-party engineering-brand publishing, no product being sold
The sole publisher is the subject: Grab writing about a system Grab built, on its own engineering blog, where the standard incentives are technical reputation and hiring. That is a real bias vector — the post selects which numbers to publish and controls the framing of its own success. It is moderated by the fact that Jarvis Pro is internal tooling rather than a commercial product being marketed, and by the post's voluntary disclaimers and its opening account of a wrong answer, which cut against pure promotion.
Confident on design, weak on outcomes
Confidence is high that the described architecture, guardrails and evaluation approach are as reported, because the account is specific, internally consistent and self-caveated. Confidence is low on anything about real-world effectiveness: single publisher, no external verification, synthetic worked example, truncated methodology text, and no production or business-impact data by the publisher's own admission.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 20, 2026