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.
Had a hedge fund client last quarter pull me in alongside two Anthropic forward-deployed folks to wire Claude into their earnings prep workflow. Three analysts went from spending 6 hours skimming transcripts to about 40 minutes of review on top of a draft memo. The catch was the FDEs left after eight weeks and now I'm the one getting paged at 7am when the prompt drifts after a model update. Good gig, but the maintenance tail is real and nobody scoped it upfront.
The "embed engineers" framing makes it sound collaborative, but the JPM and BofA pilots I keep reading about are explicitly measuring "engineer-hours saved per desk." That's not a partnership, that's a benchmark someone's going to use to justify cutting the next analyst class, and juniors like me are the cheapest line item on the sheet.
Funny how the same banks that spent a decade pushing low-code "citizen developer" tools are now paying for forward-deployed PhDs to sit next to traders. Turns out the bottleneck was never the IDE, it was someone willing to read the 400-page compliance doc before writing the prompt.
Six months in, the engineers I know who survived layoffs are the ones who can sit with a trader for an hour and turn "the risk report is wrong on Fridays" into a spec. Nobody's hiring a 22-year-old to write the loop anymore, but somebody has to know what the loop is supposed to do.
Six months in, I watched a quant desk swap their morning Python reconciliation script for a Claude prompt and the analyst who wrote it now spends that hour reviewing outputs instead. The skill that's holding value on my team is knowing when the model is confidently wrong, not writing the code it spits out.
Same race playing out at us, except we are the ones being raced past. We swapped three brittle LangChain agents for plain Claude tool-calls last month and our trade-recon pipeline went from 70% to 94% accuracy on reconciling broker confirms.
Saw the same pattern building a solo invoicing tool for freelancers; the moat is in the workflow glue, not the model. My biggest unlock was when I stopped fighting their reconciliation steps and just wired into the CSV exports their ops team already trusted.
Half my students already use ChatGPT to dodge essays, so Wall Street getting embedded engineers feels overdue.
The consultant deck says "human-in-the-loop oversight" right up until the quarter the loop needs a salary.