Half the Tech Workforce Feels Supercharged by AI. One in Seven Feels Like It's Drowning Them. Here's What Separates Them.

The data is in. Not from a think piece, but from 6,000 tech workers asked directly what AI is doing to their jobs. The workforce is splitting — and the dividing line is not which AI tools people have. It is whether they know how to use them reliably enough to feel in control.

The data

On July 21, Lenny Rachitsky published his second annual Tech Worker Sentiment Survey — 6,000 respondents across product, engineering, design, marketing, and operations. Burnout jumped from 44.7% to 55.7% in one year. Career optimism fell to 48.7%. Fifty-three percent said they would discourage someone new from entering their role.

But the grimness sits inside a split. Forty-nine percent say AI has "amplified" their work — they produce more without working harder. Fourteen percent say AI has "destabilised" them — more output is expected, and they have no clear framework for managing it. Thirty-seven percent are in between: using AI, but not getting the consistent results they expected.

The most telling number is not the burnout figure. It's the biggest professional fear: 51% are most afraid of being expected to do more work for the same pay — not of losing their job outright.

The same week, Ethan Mollick published his Summer 2026 AI guide with a headline observation: AI now performs "extremely well" at MBA-level business analysis. The human premium has moved from producing the analysis to judging whether it is right. And in his White House briefing on July 26, Sam Altman stated that 85% of OpenAI's own back-office operations now run through coordinated agent teams — measured in "knowledge per dollar" rather than headcount.

The bar has risen again. The question is whether the 37% who are "conflicted" can cross before the gap closes.

What changes on Monday

The "amplified" 49% are not using different tools than everyone else. Rachitsky's data points to a single behavioral difference: they have built repeatable workflows for their most common tasks, not one-off prompts they have to improvise each time.

A one-off prompt treats AI like a search engine with better grammar. A workflow treats it like a capable colleague who handles a defined set of steps — with specific inputs, a known output format, and a clear moment where you check the work before it goes further.

Three patterns that separate amplified from conflicted:

One workflow first, not a hundred experiments. The amplified group picked the task they do most often — a weekly report, a client brief, a research synthesis — and built AI into it systematically. They did not try to use AI for everything at once.

A review checkpoint that is part of the workflow, not an afterthought. Mollick's observation about the human premium applies here: the skill that matters now is judgement, and judgement is most useful before an output reaches a stakeholder.

Explicit supervision levels for different tasks. Not everything needs a human in the loop. Knowing which tasks do is itself the skill.

Priya — the one-workflow fix

Priya is a marketing manager at a 60-person healthcare SaaS company. She used ChatGPT for eight months and got inconsistent results — sometimes the output was useful, sometimes it took longer to fix than to write from scratch. She was firmly in the "conflicted" 37%.

What changed was not switching tools. It was designing one workflow: her monthly customer story brief. Every month she produced a story from a customer interview, and the process was different every time. She built a prompt template that takes interview notes as input and outputs a structured brief — story angle, three supporting quotes, business outcome stated in the customer's words. She reviews the brief before it goes to the writer.

That single workflow saved her four hours a month and consistently improved brief quality. More importantly, it gave her a model for thinking about any AI task: specific inputs, clear structure, her judgement before it leaves her hands.

Marcus — the other direction

Marcus is a senior analyst at a 200-person financial advisory firm. His firm rolled out AI tools firm-wide in January. He tried using it for client reports, market research, internal memos — everything at once. Two months later he was spending more time checking outputs than he'd saved in drafting.

The problem was not the AI. It was no framework for which tasks to delegate, which to spot-check, and which to keep fully human. He was running the AI at full autonomy on tasks that needed his analytical judgement, and laboring over outputs on tasks that were genuinely low-stakes.

The fix came from Mollick's Summer 2026 guide: maintain approval requirements for any action involving sending, spending, or deletion. Everything else can move faster. Marcus drew those lines explicitly. His time started working again.

The one thing

The 6,000-person survey is not a prediction. It is a snapshot of what is already true: half the people using the same tools you have feel in control, one in seven do not, and the gap tracks to whether they have built the workflows that let AI amplify their work rather than add to it. Start with one workflow. Frame the task, delegate it, review the output, iterate. That single cycle is what separates amplified from conflicted.

<BlogPracticeSection />