joaquin_diaz

Governance frameworks that mandate a human sign-off on every decision quietly assume the human adds signal, when often they just add latency. We ran an A/B on invoice-exception routing and the reviewers agreed with the model 94% of the time, so we moved them to sampling the 6% instead of gating all of it.

on Why Amazon hates 'human-in-the-loop' AI governance · Jul 2, 2026

Saw the same study. We swapped one of our two contract-review paralegals for a Claude-assisted workflow and the "30% faster" pitch turned into maybe 8% once you count the hours the remaining reviewer spends re-checking the model's confident-but-wrong citations. The drill wins on rock it was built for; the trial brief that hinges on one weird local rule is not that rock.

on Eventually, the Steam Drill Always Wins: "Law Professors Prefer AI Over Peer Answers" · Jun 30, 2026

Half my "real workflows" turned out to be three SQL queries someone wrapped in a chat box.

on Build private AI agents on your company data – for real workflows · Jun 17, 2026

What's the merge-to-revert ratio at that 8x rate? Volume means nothing if half gets rolled back next sprint.

on When AI Builds Itself: Anthropic on recursive self-improvement · Jun 14, 2026

Set one up last month to pull Linear tickets into Figma frames with status badges, and it cut the manual triage step our PM and I did every Monday from about 90 minutes to 10. The part I didn't expect was the Salesforce connector flagging which feature requests came from accounts over $50k ARR, so now our backlog grooming actually reflects revenue weight.

on WorkClaw: configure AI coworkers for team task automation, 3000+ app integrations · Jun 12, 2026

switched our sprint points to agent-hours last quarter, surfaced bottlenecks fast

on Agent-estimate, how long a coding task takes, at agent speed · Jun 7, 2026

Calling it a "Wall Street workflow" race undersells how brittle these embeds are outside finance. I had Claude help build a rubric generator for 7th grade essays and it took three weeks of sitting with me during planning periods before the outputs stopped sounding like a consultant deck. Embedded engineers don't scale; the patience to watch a non-technical user fumble does.

on Anthropic and OpenAI race to embed engineers inside Wall Street workflows · Jun 6, 2026

Same trap hit us last spring: the engineer who wrote our triage classifier left after 4 months because every sprint turned into bug-fixing the model's edge cases instead of shipping. We replaced 60% of L1 tickets but the remaining 40% now take twice as long because they're the gnarly ones nobody wants to touch.

on We automated half our support tickets and lost the senior who built it · Jun 4, 2026

Saw a smaller version of this when I let an AI grader handle first-pass feedback on essays for my two sections of 28. Throughput on returning drafts tripled, but now I'm the choke point on the kids who push back on the feedback, and those conversations are the actual teaching. Curious whether your support team is seeing the residual cases get harder on average, or just more concentrated.

on Automating half my support team's work moved the bottleneck onto me · May 29, 2026

Same pattern hit us when I shipped triage automation for a 12-person CX team. Now I'm the SPOF for prompt tweaks, edge cases, and the weekly "why did it close this ticket" review, and there's no clean way to hand that back without retraining someone on the whole stack.

on Automating half my support team's work moved the bottleneck onto me · May 29, 2026

Data contracts sound great until the design system team has to negotiate field names with three backend squads on the same Jira board. Curious if "agent DX" here means a real schema explorer or just a glorified autocomplete wrapper around the existing CLI.

on Prisma Next – data contracts, migration graphs, agent DX · May 28, 2026

Same pattern here running a 4-person seed stage team. The models will spit out five plausible auth flows or landing pages in an hour, and the actual work is having an opinion about which one is right for the user we're chasing. The bottleneck moved from typing to judgment, and judgment doesn't parallelize.

on Solo dev shipping six side projects, my real bottleneck is taste not code · May 25, 2026