A friend of mine runs a 7-person product agency. In late 2025 he messaged me: "We're probably saving 40 hours a month now. But three things broke that we didn't expect."
I asked him to walk me through it. Here's what they did, how they did it, and — importantly — what went sideways.
I'm sharing this because most AI-at-work content is either cheerleading or fear. Neither helps you figure out what to actually do.
The Setup
Seven people: 2 designers, 2 developers, 1 strategist, 1 ops person, and my friend who runs the whole thing. They do product strategy, UX design, and early-stage build work for startups. A project typically runs 6–12 weeks.
They were doing well, growing, and also drowning in process overhead. Three workflows in particular were eating time:
- Client meeting notes — Summarizing calls, distributing action items, keeping clients informed
- Proposal writing — New business pitches took 6–10 hours of senior time per pitch
- QA reporting — Developers writing bug reports and summarizing test runs by hand
These weren't broken workflows. They were working fine. They were just slow and expensive.
What They Changed
Workflow 1: Meeting Notes
They started routing all client calls through a transcription service (they use Fireflies), then piping the transcripts into Claude with a prompt that extracts decisions made, open questions, and action items with owners.
The output goes into Notion automatically. The ops person reviews and sends the client summary — a task that used to take 45 minutes now takes about 5.
Result: ~40 minutes saved per client call. They have 3–4 client touchpoints per week.
Workflow 2: Proposal Writing
Proposals at this agency follow a recognizable structure: situation analysis, recommended approach, team and process, timeline, investment. My friend built a prompt template that pulls from a Notion database of past projects and outputs a first draft.
A senior strategist still reviews and personalizes everything — especially the situation analysis, which requires real understanding of the client. But the skeleton work that used to take 4–6 hours now takes about 90 minutes.
Result: Senior time on proposals dropped by ~60–65%. They're pitching more because the cost of pitching dropped.
Workflow 3: QA Reporting
The developers started pasting test results and error logs into Claude and asking for structured bug reports in their standard format. They also use it to generate first drafts of test case lists when starting a new feature.
Result: Uneven. More on this in a moment.
What Actually Broke
1. The junior devs stopped learning how to debug
This is the one my friend feels worst about. When a junior developer pastes an error into Claude and gets a clear explanation, they understand the error. But they don't develop the debugging intuition that comes from sitting with confusing output for a while and working through it.
One of his junior developers shipped a bug fix correctly but couldn't explain why the original error happened. \"He knew the answer. He didn't understand the problem.\"
They've since added a rule: for bugs that are good learning opportunities (ambiguous, architectural, novel), the junior has to form their own hypothesis first before using AI to check it.
2. The proposal quality variance got worse
Proposals got faster but more uneven. A good first draft made it easier to ship quickly. A mediocre first draft made it easier to ship something mediocre — because the revision felt like a smaller lift than it was.
The baseline raised. The ceiling lowered. They fixed this by adding a mandatory senior edit on the framing and situation analysis sections specifically, treating those as non-delegatable.
3. Client meeting summaries lost texture
The AI summaries were accurate and thorough. They were also flat. What got lost were the softer signals — the thing the client said hesitantly, the concern that got raised and then moved past, the vibe of the conversation.
The ops person started adding a \"Read Between the Lines\" section at the bottom of each summary — two or three sentences written by hand about what wasn't in the transcript but was in the room. That solved it.
What I Take From This
AI didn't break their workflows. It accelerated whatever was already there — including the weaknesses.
The junior dev issue was always a risk. Speed just made it surface faster. The proposal variance was latent in their review process. The summary flatness exposed how much they were relying on implicit knowledge that wasn't in documents.
Every workflow they touched got faster. Every workflow they touched also exposed something they hadn't been explicit about before.
That's not a reason not to use AI. It's a reason to use it with your eyes open. The things that break will tell you something true about your process.
I write about AI at work at denismoroz.ai. The newsletter is where I share this kind of case study in more depth.
