Quinn Slack estimates that only around 5% of professional developers use code AI, despite widespread claims about adoption and productivity.
2
Most current code AI usage is autocomplete, while features such as chat, edits, and agents have weaker frequency and verification properties.
3
AI product builders should use their own product every day, build manual workflows before adding automation, and reduce hype when pitching the product.
Summary
Quinn Slack argues that the code AI market is much smaller than public adoption figures suggest. A GitHub study reported that 92% of developers use code AI at work, but its sample was limited to 500 people in the United States and used a broad definition of use. Slack estimates that around 5% of professional developers use code AI, with most usage limited to autocomplete. He connects this gap to the revenue expectations behind foundation models, infrastructure companies, and AI applications. Drawing on Sourcegraph's experience building Cody, Slack explains why autocomplete works so well, why chat and agents face harder product problems, and why hype misleads both customers and builders. His advice is practical: use the product daily, describe its value without saying AI, start with manual and explicit workflows, make errors easy to fix, and set expectations lower than feels necessary.
Public adoption figures make code AI look more widely used than it is
Slack opens with the claim that most developers do not use code AI. He questions a GitHub study reporting that 92% of developers use code AI tools at work because it covered 500 people in the United States and used a broad definition of use. GitHub also reported 1.3 million paid Copilot subscribers in January 2024, but Slack asks how many actually use the product. Another figure showed that 935,000 developers received a suggestion during the study period, without necessarily accepting one. Against an estimated 26 million professional developers worldwide, Slack's best estimate is that around 5% use code AI.
Most code AI usage is still limited to autocomplete
Slack says the vast majority of current code AI use is ghost-text autocomplete. He accepts that autocomplete is useful, but says developers who have tried more of the technology know it can write files, answer questions, and fix bugs. This means the market is seeing only a small part of what code AI can do. Slack tells the audience that people in the room are the unusual users, since 19 out of 20 professional developers would not be using code AI under his 5% estimate. Builders therefore need to change the minds of developers who have little reason to try these tools.
The AI economy depends on recurring revenue from real enterprise use
Slack describes a path from foundation model companies and infrastructure providers to experimental use, paid individual accounts, workplace adoption, and recurring enterprise contracts. He says the first-year contract matters less than whether a company such as JP Morgan renews and pays more in later years. Slack estimates total recurring revenue from code AI at around $300 million in annual recurring revenue. Using Sourcegraph's experience that less than 10% of revenue goes to AI inference, he estimates that roughly $30 million would flow back to model and infrastructure companies. He compares the total with Salesforce's $36 billion annual revenue to show how early the market remains.
Product builders should use their own product every day
Slack says hype fooled Sourcegraph, its customers, and the audience. His strongest advice is that a team building a product must use it every day. Customers repeatedly asked Sourcegraph for fine-tuned models, partly because the feature appeared in competitor messaging. Slack thinks they were really asking for the product to work well. He also says a builder should be able to explain why the product works without using the word AI. If the team cannot do that, he doubts the product is ready. He pushes the point further by asking why a product that makes developers 20 or 30% more productive is sold as software instead of being used to run a software outsourcing business.
Autocomplete works because it is frequent and easy to check
Sourcegraph evaluates features by how often they are used and how quickly a developer can decide whether the result is correct. Autocomplete occupies the favorable corner of that framework because it appears on every keystroke and can be checked in milliseconds, or by a quick glance. Slack calls this an unusually strong product fit and says builders were spoiled because the first important code AI feature happened to have these properties. Chat is used less often and produces longer responses that are harder to verify. Inline edits have appeal, but Slack warns that the successful form of code AI may differ from the features available today.
Agents and other features need to reach a simple, frequent workflow
Slack says agentic code AI is probably the future, but it is not ready. He asks how many people have used a code AI agent to merge a pull request in the previous week, and only a small part of the audience responds. Sourcegraph is searching for the next major code AI modality, with ideas such as next-edit suggestions and chat-oriented programming. Slack describes chat-oriented programming as keeping one ongoing conversation open while coding, a practice he has seen among both younger developers and older programmers. He expects developers to discover uses that builders have not yet considered.
Manual and explicit workflows should come before automation
Slack says builders should first make the explicit version of a workflow work. For chat, that means allowing developers to manually mention the context they want before trying to insert context automatically. For an agent, he recommends putting it in the editor so the developer can correct a wrong result there rather than moving to a separate interface such as Code Spaces. The product should make manual correction easy. Slack expects teams to spend enough time making the basic workflow good that they may not add the automated layer for a while.
Lowering expectations is part of building and selling the product
Slack advises builders to dehype their products aggressively because customers will increase the hype in their own minds. He cites Sam Altman's earlier warning that ChatGPT was still not very good despite having tens of millions of daily active users. Slack says a less validated product should communicate its limits even more strongly. He closes by saying foundation model companies, infrastructure companies, and application companies depend on one another. Their shared outcome depends on developers using the products often enough for the applications to make substantial revenue.
"If you can't describe why it works and why it's great without using the word AI then you're probably not going to make it."12:46
Who should watch
You are building an AI coding product and need a clearer test for whether developers will use it repeatedly.
Your adoption numbers look strong because they count trials, subscriptions, or impressions, but you do not know whether developers accept and rely on the output.
You are designing chat, edit, or agent workflows and want practical advice on verification, manual correction, and product expectations.
