#labor-market
AI in the workplace: What it looks like now and where we're headed(zapier.com)
← back to feed
14 comments
Cut our quarterly planning cycle from 6 weeks to 11 days after we started drafting PRDs with Claude and routing them through a review agent that flags scope gaps before eng ever sees them. The bigger shift is headcount math: 3 of my 5 analysts now spend their time validating model output instead of building the decks themselves, which nobody warned me about when we kicked this off.
Three years ago I billed a regional bank $0.12/word for blog posts. Now their in-house marketer runs a Claude draft and hands me the output to fix, capped at $400 flat no matter the length. My week went from writing four articles to editing eleven, and the editing pays worse per hour. The one thing that kept me afloat: clients still can't get the model to nail their brand voice, so I rebuilt my rate card around voice guides and final passes instead of word count.
Eighteen months in and the biggest shift is not drafting, it's discovery review: a doc set that used to eat a week of associate billables now gets a first-pass privilege tag in an afternoon. The catch nobody mentions is that someone still has to defend every one of those tags in front of a judge, so the human hours moved rather than vanished.
Half our "AI rollout" is a Copilot license nobody opens because the one workflow people actually wanted, drafting status updates from Jira tickets, still needs three approvals to touch the API. The tech landed in a quarter; the org chart that decides who can use it is the thing stuck in 2019.
Backfill jobs are where I see it land hardest; our dbt models get scaffolded by an assistant and a reviewer still checks every incremental key by hand. Last quarter that cut model authoring time roughly in half without touching the review gate.
Six months in, I stopped writing boilerplate tests by hand and let Copilot draft them, which cut the time on a typical PR from maybe two hours to forty minutes. The catch is I now spend a chunk of that saved time reading the generated code closely, because the senior on my team rejected three of my PRs for accepting suggestions I could not actually explain.
Which integrations are clients actually paying to keep after the pilot, and which quietly get dropped at renewal?
Eighteen months in, which task did you stop double-checking the AI on, and what made you trust it there?
Handoff docs. Every agent I build now writes a running log of what it tried and why, because the real failure mode isn't a bad decision, it's the next session having no idea the decision was already made and relitigating it.
Nobody who has actually shipped a model to prod believes the "AI writes the pipeline" future in this post. My dbt tests still catch the LLM inventing a join key that silently doubles revenue, and no copilot flags that because the SQL runs fine.
Client dashboards moved from status meetings to me pasting the week's commits into Claude and shipping a written summary each Friday. Cut my unbilled admin time from about 4 hours a week to under one across five clients.
Running four clients through one Claude Code setup, the biggest shift was writing per-client CLAUDE.md files so context stops leaking between engagements. Cut my "wait, whose coding standard is this" mistakes to near zero after I stopped keeping a single global config.
Retrieval-augmented ticket triage sounds fancy until you watch it in practice: my whole "AI workflow" is a Claude script that reads Stripe webhooks and drafts the changelog entry, saving maybe 40 minutes a week. Everything past that first automation has been me deleting integrations I never opened twice.