The company, not the chatbot, should own what AI learns

On October 27, Salesforce published a survey of 250 US executives with a finding that names a problem most companies feel but rarely measure: work is getting trapped inside one person's AI tool¹.

Salesforce calls this single-player AI. Rob Seaman, who runs Slack, put it plainly a few weeks earlier: when the work happens in a private chat, “the knowledge stays in the tab”². Asked to choose, 72% of the executives said they would rather have AI that works across the team with shared context than AI that each person uses on their own¹.

We agree with the diagnosis. This piece follows it further, through the independent research on why AI pilots stall, how companies buy models, and where agents fail, because we think the cause is more specific than “AI is not shared”, and the fix asks for more than one product category usually offers.

The most cited study of enterprise AI last year was MIT NANDA's The GenAI Divide. It found that about 95% of organizations were getting no measurable return from their generative AI pilots, and it was direct about why: “The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time”³.

The same study found the tools themselves are not the problem. Workers at more than 90% of the companies surveyed used personal AI tools regularly, while only 40% of the companies had bought an official subscription³. Microsoft and LinkedIn found the same pattern a year earlier: 78% of people using AI at work were bringing their own tools⁴. People like these tools for quick work. MIT's respondents preferred AI for drafting emails (70%) and basic analysis (65%), but for complex or long-term work they preferred a human by nine to one³. One lawyer interviewed for the study explained the line they draw:

That is the single-player problem from the other side. Context lives in people's heads and in their private chats, so every session starts over, and nothing one person teaches the tool reaches the next person. Atlassian's survey of 12,000 knowledge workers found teams already lose a quarter of their time searching for answers⁵. Salesforce's executives estimate another 3.6 hours a week go to switching between AI and the other tools people work in¹. AI that forgets adds to that bill instead of paying it down.

The obvious answer is to give the AI a memory, and the large providers are doing exactly that. Projects, saved memories and connected files make a single provider's assistant much better for a team over time. MIT's report predicts where that leads: “Organizations investing in AI systems that learn from their data, workflows, and feedback are creating switching costs that compound monthly”³. A financial services CIO in the same study said it more bluntly: “Once we've invested time in training a system to understand our workflows, the switching costs become prohibitive”³.

Those switching costs are not hypothetical. In Menlo Ventures' survey of 150 technical decision-makers, only 11% of teams had changed model providers in the past year⁶. Andreessen Horowitz's survey of 100 CIOs found the same pull from a different direction: the rise of agentic workflows is making it harder to switch between models⁷.

Look at what actually moves when a company tries to leave. Enterprise exports exist, and they are built for compliance: OpenAI's Compliance API, for example, delivers time-stamped logs of conversations, files and memories to auditors and security tools⁸. That is a record of what happened. It is not a working memory another model can pick up and use. The tools built for switching are thinner still. When Anthropic added a way to bring memory into Claude from ChatGPT, Gemini or Copilot this year, it worked by giving the user a prompt to paste into the old assistant, which writes out what it has stored as a block of text to paste into Claude⁹. What crosses over is whatever one assistant can summarize in one reply: not the projects, not the sources behind a fact, not who decided what and why.

Regulators have started treating this kind of lock-in as a real cost. The EU Data Act's switching rules for cloud and software services took effect in September 2025, require providers to let customers move their data and digital assets to another provider, and end switching charges entirely from January 2027¹⁰. Whatever the law ends up requiring of AI assistants, the direction is clear: what a company's tools learn about the company should leave with the company.

Lock-in would matter less if one provider stayed the best at everything. None does. At the end of 2023 OpenAI held about half of enterprise model API usage; by mid 2025 Anthropic led with 32% to OpenAI's 25%, with Google at 20%⁶. Stanford's AI Index measured the same churn at the top: on the Chatbot Arena leaderboard the gap between the first and tenth model shrank from 11.9% to 5.4% in a year, the gap between the top two from 4.9% to 0.7%, and the gap between the best closed and best open-weight model from 8.04% to 1.70%¹¹. Companies have noticed. 37% of the CIOs a16z surveyed already used five or more models, up from 29% a year before, and the main reason they gave was that different models are better at different jobs⁷.

That makes choosing the model a job of its own, and research says it is a job worth doing well. Stanford's FrugalGPT showed a learned cascade of models could match the best single model with up to 98% lower cost, or beat it by 4% at the same cost¹². LMSYS's RouteLLM, trained on human preference data from Chatbot Arena, cut costs by more than half in some settings without lowering the quality of answers¹³. Nobody should have to read a leaderboard every week to know which model to use on Tuesday.

The next wave makes all of this more pressing. In McKinsey's 2025 survey, 62% of organizations were at least experimenting with AI agents, but only 23% were scaling one anywhere in the business¹⁴. Salesforce found 49% of executives deploying agents with some autonomy, and only 15% with AI embedded across multiple systems¹. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls¹⁵.

An agent is single-player AI with more reach. If it cannot see what the company knows, it either asks a person for context every time or acts without it. MIT's report makes the same link: agentic systems are the class of AI that can close the learning gap, because they keep persistent memory instead of needing full context each time³. The agent and the memory are one design problem, not two.

Put the research side by side and the requirements fall out of it. A company needs the best model for each task, without betting the company on one provider. It needs a memory of its work that grows from everyone's use, that it owns and can take with it. And it needs agents that work from that memory, safely. Any one of these alone recreates the problem: every model without shared memory is still single-player; a memory inside one provider is lock-in; agents without either are the projects Gartner expects to be canceled.

This is why we built Guile the way we did.

The result is the multiplayer AI the Salesforce executives asked for, without making the company's memory a feature of one vendor's product. That includes us: an export that works is how a company knows its knowledge is its own.

Surveys of executives are estimates, not measurements, and several of the sources here come from companies that sell AI, Salesforce and Guile included. Salesforce's survey covers 250 US executives and its hours are their estimates¹. MIT calls its own findings preliminary, drawn from 52 interviews, 153 survey responses and a review of public deployments³. Routing results come from benchmarks and will not transfer one to one to any company's own work^12,13. We will keep testing these claims against what we see in practice, and publish what we learn here.