The Tools the Toolmakers Keep for Themselves

Business Insider ran a story this week that, on first read, looks like a small piece of corporate gossip. Google DeepMind engineers have been granted access to Anthropic's Claude for their coding work, while most of the rest of Google is restricted to the company's internal Gemini tools. A proposal to pull Claude access from DeepMind surfaced internally and was met with what the reporting describes as retention threats, engineers signalling they would leave if the better tool was taken away. Meanwhile, across Google more broadly, AI use is now wired into performance reviews. Engineers without access to the best tools are being measured against colleagues who have them.¹

You could read this as a procurement story, or an HR story, or a competitive-intelligence story about how much Google's most senior AI researchers prefer a rival's model to their own. Those readings are all available and all partial. I want to read it as something else: a very small, very clear window onto a pattern that is going to define a lot of what happens next. The people who build AI systems are already sorting themselves into tiers of access to those systems. And the tier you land in is starting to determine how your work is judged.

The most sophisticated users. Start with what the DeepMind engineers are doing when they reach for Claude. They are writing code, the most measured, most automatable, most "AI-ready" white-collar task on the planet right now. They are also, by any reasonable definition, among the most sophisticated users of AI coding tools in the world. If anyone should be satisfied with Gemini for coding work, it is the people who built large parts of it. They are not. They are reaching past their own product to a competitor's, and when someone suggested closing that door, they threatened to walk.

I find this clarifying. Not because it tells us Claude is better than Gemini at coding on some particular benchmark, that question moves month to month and isn't really the point. It's clarifying because it tells us something about how the people closest to this technology actually behave when the stakes are their own output. They do not accept the tool they are handed. They demand the best available tool, and they have the leverage to get it. The retention threat is the tell. It says: my productivity with the right tool is high enough, and my market value is high enough, that you will lose me before I will work with the second-best option.

My productivity with the right tool is high enough, and my market value is high enough, that you will lose me before I will work with the second-best option.

The review system. Now hold that next to the performance-review change. Across Google, AI use is being folded into how engineers are assessed. This is not unique to Google, Shopify, Microsoft, and a long tail of smaller firms have been moving the same way through 2025, with varying degrees of explicitness.² The logic is straightforward: if AI tools materially raise output, then refusing to use them is a performance problem, and managers need a way to register that. Fine. But the logic assumes the tools are roughly fungible, and that access is roughly equal. Neither assumption survives contact with what the DeepMind story shows us.

Inside one company, at one moment, we have:

A top tier of engineers who have fought for and won access to what they consider the best tool, and who have the leverage to keep that access.

A middle tier working with the company's own tools, which the top tier has evaluated and found wanting for their purposes.

A performance-review regime that treats AI-assisted output as a measure of the engineer, not of the engineer-plus-tool.

This is the bit I want to sit with. The standard framing of AI and work, in the coverage I read, goes something like: AI tools are becoming available, workers will adapt to them or not, and the ones who adapt will do better. It's a framing that treats the tool as a given and the worker as the variable. The reframing I want to try is that the tool is the variable, and the worker's leverage determines which tool they get. The DeepMind engineers have shown us this in miniature. The question is how the pattern generalises.

Why it generalises badly. Consider the call centre worker who is now expected to use an AI assistant to handle customer queries. She does not get to evaluate three models and pick the one she prefers. She gets the one her employer has licensed, configured the way her employer has configured it, with the guardrails her employer has installed, logging the interactions her employer wants logged. If that tool is worse than the one a competitor's employees use, she cannot credibly threaten to leave, the labour market for call centre work does not work that way. If her handle time is measured against a benchmark set by workers using better tools, she absorbs the difference. If AI-assisted productivity is folded into her review, she is being assessed partly on the quality of a tool she did not choose.

Consider the junior lawyer at a firm that has licensed one legal-AI product. She is competing, in her career, against lawyers at firms that licensed a better one. The difference will not show up as "tool quality." It will show up as her work product, her billable efficiency, her partner-track trajectory. The tool becomes invisible and she becomes the explanation.

Consider the teacher, the social worker, the radiologist, the journalist, the coder at a startup that can't afford frontier-model API bills. The pattern repeats: a tool is issued, output is measured, and the gap between what the best tool can do and what the issued tool can do is charged to the worker's account.

The DeepMind engineers have shown us that this gap is real and that sophisticated users know it's real. They know it well enough to stake their jobs on not being on the wrong side of it. What's striking is how quickly the same company that accommodates this knowledge at the top is building review systems at the bottom that pretend the knowledge doesn't exist.

I want to be careful here. I am not arguing that every worker should have access to every frontier model, or that performance reviews should ignore AI use, or that tools don't vary in appropriate ways for appropriate reasons. A call centre's AI needs different guardrails than a research lab's. That's fine and sensible. What I am arguing is that the differentiation of tool access is becoming a major axis of differentiation in work itself, and we are not yet talking about it that way. We are talking about "AI adoption" as though it were one thing that happens to a worker, when in fact it is many things, stratified by employer, by role, by seniority, by leverage, and the stratification is already visible inside the single company where the tools are made.

A second thread. There is a second thread in this story that I don't want to let go of. The DeepMind engineers' preference for Claude over Gemini is, among other things, a piece of evidence about Google's own product, leaking out through the behaviour of people who know it best. In a functioning market that evidence would matter. Inside a company, it's being managed as a retention problem rather than a product problem, which tells you something about how AI competition actually works at this level: the strongest signals are the ones firms have the most reason to suppress.

This matters beyond Google. A version of this is happening at every large firm building AI, and a version of it is happening at every large firm deploying AI. The people who know which tools are genuinely better are mostly inside the firms, mostly bound by confidentiality, and mostly incentivised not to say so publicly. The people who have to decide which tools to deploy to their workers, procurement officers, IT leads, operations managers, are working from vendor materials, benchmarks the vendors helped design, and whatever leaks out of stories like this one. The asymmetry is enormous, and it runs in the direction of the vendors.

So when a company rolls out AI-assisted work and ties it to reviews, what's actually happening is: an employer is making a sourcing decision under heavy information asymmetry, installing the result on a workforce that has no leverage over the choice, and then measuring the workforce against the output of that choice. The choice is made once, at the top. The consequences are paid continuously, at the bottom.

The choice is made once, at the top. The consequences are paid continuously, at the bottom.

I think this is the shape of a large part of the coming labour story around AI, and I think it's mostly being missed because the dominant framings keep treating "AI in the workplace" as a question of whether workers will adopt tools, rather than whether the tools workers are given are any good and who bears the cost when they're not. The DeepMind engineers have the answer to the first question: yes, obviously, aggressively, with leverage. They are also, by their behaviour, telling us the answer to the second: it matters enormously, the differences are material, and people who can choose will choose. It is only the people who can't choose who are going to be told the tool doesn't matter, that what matters is whether they used it, and how well they did with what they were given.

What's to be done about this is not, I think, the first question. The first question is to see it clearly. The standard story is that AI is a tool that workers use; the truer story is that AI is a set of increasingly stratified tools that workers are issued, with the stratification tracking existing hierarchies of power and leverage inside and between firms. The performance-review question, whether workers should be assessed on AI-assisted output, cannot be answered honestly without answering the tool question first. And the tool question cannot be answered honestly by a firm that is, at the same moment, accommodating its own top engineers' refusal to use the tool it's about to issue to everyone else.

I don't think Google did anything unusual here. I think they did the obvious thing, which is to let the people with leverage keep the tools they want and hand everyone else what the company preferred to standardise on. That's how firms work. What's worth noticing is that the obvious thing, done at scale across an industry, produces a labour system in which the quality of the tools you work with tracks the power you already had, and then gets folded into how you're measured, as though the tool were neutral and you were the variable.

That is the story I'd like us to start telling about AI at work. The DeepMind engineers, by fighting for Claude, have made it easier to see.

Footnotes

¶