The number came out this week: GPT-5.4 scored 75% on multi-step desktop productivity tasks.
The human baseline is 72.4%.
AI is above the line.
I want to be precise about what that means — because the hot takes are already pulling in two directions. One camp says this proves the AI-kills-jobs narrative. The other says benchmarks are synthetic and don't reflect real work. Both are missing the point.
What the Number Actually Tells You
A benchmark score isn't a performance guarantee. It's a signal about capability trajectory.
GPT-5.4 didn't score 75% on your specific workflow, in your specific system, with your specific data. It scored 75% on a standardized test of multi-step desktop tasks. That's a meaningful result — but it's a floor, not a ceiling, and it's an average across task types with very uneven distribution.
Some tasks in that benchmark are probably at 92%. Others are probably at 40%. Knowing the average tells you the capability is real. It doesn't tell you which of your workflows it's ready to take over.
That's the distinction that matters for anyone making hiring or building decisions right now.
The Roles Going Up in Value
When AI crosses the human line on productivity tasks, specific roles become more valuable — not less.
The people who understand how to design workflows that AI can reliably execute are now the most important people in your organization. Not because they're managing the AI. Because they understand the gap between what AI can do in a benchmark environment and what it actually does inside your messy, partially-documented, inconsistently-maintained systems.
That gap is real. I wrote about it a few days ago. The Oracle Flood put 20-30K enterprise IT specialists on the market — people who understand how large systems actually behave, not just how they're supposed to behave. The Snowflake-OpenAI $200M deal announced this week is predicated on embedding autonomous agents into enterprise data workflows. Who builds that? Who debugs it when it breaks at 2 AM? Who tells leadership why the agent made the decision it made?
Not the AI.
The Roles Being Restructured
Some functions are genuinely on a compressed timeline.
Multi-step desktop productivity tasks — think: research synthesis, data formatting, report generation, scheduling coordination, document drafting — these are the first to be restructured, not eliminated. The human doing five of them in a day becomes the human overseeing the AI doing twenty and reviewing the outputs that fall below confidence threshold.
That's a different role. It requires different judgment. It commands different pay.
The companies that figure out this restructuring deliberately — "here's what the human does, here's what the AI does, here's how we verify quality" — will run faster and leaner than their competitors within 12 months. The companies that ignore it will keep paying for work AI could already do. The companies that panic-cut without building the oversight layer will have cheap outputs and no one to catch the failures.
What to Do With This Information
Three things.
First, stop treating the benchmark as a binary signal. It's not "AI can do this, so we don't need humans." It's "AI can do this reliably enough that the human's job is changing."
Second, if you're hiring right now, get explicit about what the human is actually accountable for in each role. If the accountability is "do the thing" and the thing is a desktop productivity task, you're hiring for the wrong thing. If the accountability is "verify, decide, and own the outcome" — that role is getting more important.
Third, the AI-literate workforce is not large. The people who can design AI workflows, manage AI outputs, and translate capability into business outcomes are commanding premiums because they're actually scarce. If you find one, don't lowball them because the market isn't fully priced in yet. It is.
VC5 Consulting works with companies figuring out what their teams should actually look like as AI capabilities accelerate. If you're thinking through the restructuring — what to hold, what to change, who to hire — let's talk.