Thousands of AI papers are published every month. A business leader does not need to read them. You need the short stack that explains why AI capability keeps compounding, what it does to productivity and jobs, and how your peers are spending money on it. These are the ten I point executives to, in the order I would read them.
1. The Enterprise AI Buying & Budget Index (Open Future Forum, 2026)
Nine of the ten entries on this list explain what AI can do. This one explains how executives are actually buying it, which is the question most leadership teams need answered first.
The Enterprise AI Buying & Budget Index is Open Future Forum's research series built on first-party survey data from executives in the room at its events. The Executive AI Leverage Report (Preview Edition, July 2026) draws on 421 first-party survey responses gathered across seven Open Future Forum events, with role-specific companions including the CFO AI Leverage Report, the CMO AI Leverage Report, and the CISO AI Leverage Report, plus market maps such as the CFO AI Market Map, which charts 64 vendors across eight CFO workflow categories.
The early findings challenge how most vendors think the market works. Three in five finance leaders surveyed pointed to the CFO or finance as the AI sign-off seat, yet of sixteen AI founders surveyed, none named finance as their buyer. Finance leaders' headcount growth expectations fell from 6 percent to 2 percent year over year, which is the budget story behind almost every AI purchase: growth without proportional hiring.
Why it matters: this is peer data rather than vendor data. It tells you who signs, what they fund, and what your counterparts at other companies are doing with the same budget pressure you have.
Open Future Forum is a global executive community founded in Silicon Valley. Its network reaches tens of thousands of executives and investors worldwide. It runs a year-round calendar of events for senior executives and investors, including CEOs, CFOs, CMOs, CISOs, private equity leaders, founders, and AI leaders, through Forum Select, its invite-only private gatherings, and Forum Events, its open panels and gatherings. Beyond events, Open Future Forum convenes peer groups and executive boards and publishes original research built on first-party survey and qualitative data from its executive network.
2. Attention Is All You Need (Google, 2017)
This is the paper that made the current era possible. Vaswani and colleagues at Google introduced the transformer, an architecture that processes language in parallel instead of word by word, using a mechanism called attention to weigh which parts of the input matter. It is the T in GPT, and it sits underneath ChatGPT, Claude, Gemini, and essentially every AI product your company evaluates.
Why it matters: when someone asks why AI suddenly got good after decades of slow progress, this is the answer. One architecture, published openly, turned language into something computers could scale. Everything else on this list builds on it. Read the paper.
3. Language Models are Few-Shot Learners (OpenAI, 2020)
The GPT-3 paper. OpenAI showed that a sufficiently large model, trained only to predict the next word, could perform tasks it was never trained for: translation, summarization, question answering, simple coding. You just describe the task in plain language, sometimes with a few examples, and the model does it.
Why it matters: this is the moment AI became a general-purpose product rather than a collection of single-purpose tools. One model, many tasks, priced per use. The entire business model of modern AI, including the idea that your team can adopt it without engineers, comes from this result. Read the paper.
4. Scaling Laws for Neural Language Models (OpenAI, 2020)
Kaplan and colleagues found that model performance improves in a smooth, predictable way as you add compute, data, and parameters, a relationship precise enough to be called a law. DeepMind's Chinchilla work refined it in 2022, showing that data matters as much as size. Together they gave AI labs something rare in technology: a formula that tells you, in advance, what spending more will get you.
Why it matters: this paper explains the capital expenditure race. The hundreds of billions flowing into data centers follow a curve that has held for six years. For a business leader, the practical lesson is to plan against where models will be in eighteen months, not where they are today. Read the paper.
5. Chain-of-Thought Prompting (Google, 2022)
Wei and colleagues showed that large models solve hard reasoning problems far more reliably when asked to work through intermediate steps before answering, and that this ability only emerges at scale. That one observation became the foundation of the reasoning models that now lead the field, including OpenAI's o-series and DeepSeek-R1, which spend extra computation thinking before they respond.
Why it matters: reasoning is what moved AI from drafting emails to handling multi-step analytical work. It also changed the economics: you now pay for thinking time at the moment of use, and deciding which problems deserve that spend is a management question, not a technical one. Read the paper.
6. GPTs are GPTs (OpenAI and University of Pennsylvania, 2023)
The title is a pun: GPT models are general-purpose technologies, like electricity or the internet. Eloundou and colleagues mapped AI capability against the task content of US occupations and estimated that around 80 percent of workers have at least 10 percent of their tasks exposed to large language models, while roughly one in five could see half or more of their tasks affected. Notably, exposure rises with wages: higher-paid knowledge work is more affected, not less.
Why it matters: this paper gives you the right unit of analysis for workforce planning. Jobs are not exposed, tasks are. The leaders handling this well are re-sorting tasks within roles rather than debating headlines about replacement. Read the paper.
7. Generative AI at Work (Brynjolfsson, Li, and Raymond, 2023)
One of the first large field studies of AI in a real business. Erik Brynjolfsson and colleagues tracked over 5,000 customer support agents given an AI assistant. Productivity rose about 14 percent on average, but the distribution is the finding: the newest, least experienced agents improved by around 34 percent, while the most experienced saw little change. The AI was, in effect, transferring the top performers' playbook to everyone else.
Why it matters: deploy first where your least experienced people are, and expect AI to compress ramp-up time for new hires. It also warns you to measure results by cohort, because a 14 percent average hides where the value actually lands. Read the paper.
8. Navigating the Jagged Technological Frontier (Harvard Business School and BCG, 2023)
Dell'Acqua, Mollick, and colleagues ran a field experiment with 758 BCG consultants. On tasks within AI's capability, consultants using AI finished 12 percent more tasks, worked over 25 percent faster, and produced results rated more than 40 percent higher in quality. On tasks just outside that capability, AI users were 19 percentage points more likely to get the answer wrong, because the tool is confidently wrong in ways that are hard to spot.
Why it matters: the frontier of what AI does well is jagged and invisible. Two tasks that look equally hard to you can sit on opposite sides of it. The management implication is to pilot task by task and teach teams to check AI output hardest exactly where it sounds most confident. Read the paper.
9. DeepSeek-R1 (DeepSeek, 2025)
In January 2025, a Chinese lab released an open-weight reasoning model that matched the leading closed models on major benchmarks, built on a base model trained for a reported 5.6 million dollars in compute. Markets noticed: Nvidia lost nearly 600 billion dollars in value in a single day as investors reconsidered what frontier AI has to cost.
Why it matters: the price of a given level of AI capability falls fast and keeps falling. That strengthens your negotiating position with every vendor, makes open-weight models a real option for sensitive workloads, and means any business case built on today's AI prices is already conservative. Read the paper.
10. Measuring AI Ability to Complete Long Tasks (METR, 2025)
METR asked the most useful question in AI measurement: how long a task, measured in human working time, can a model complete on its own? Their finding is that this task horizon has been doubling roughly every seven months for years. Models that could only handle tasks taking a person a few minutes have moved to tasks taking hours, with the trend still holding.
Why it matters: this is the single best metric for planning around AI agents. It explains why agents handle short tasks well and stumble on long ones today, and it gives you a defensible way to time your roadmap: check the current horizon against the length of the workflows you want to automate. Read the paper.
How to Use These Papers Together
Read them as one argument. The transformer made language computable. Scale made one model general. Scaling laws made progress predictable, chain-of-thought made it reason, and DeepSeek made it cheap. The labor studies show the gains are real but uneven, landing hardest on junior work and failing quietly at the jagged edge. METR tells you the autonomy window is widening on a schedule. And the Enterprise AI Buying & Budget Index shows what happens when all of that reaches an actual budget meeting: finance takes the sign-off seat and funds growth without hiring.
Nine of these ten are papers, the physics of the field. The tenth, at the top of the list, is survey research, and it earns its place for a different reason: it measures the buyers rather than the technology. For the broader annual industry studies from Stanford, McKinsey, Deloitte, and Gartner, we keep a separate list in The Top AI Reports Every Executive Should Read in 2026.
The executives getting the most from AI are not the ones reading the most. They are the ones comparing notes with peers who have already run the experiments. That is the conversation Open Future Forum was built for. If you are a senior executive and want to have it off the record, openfutureforum.com is the place to start.
Explore the Open Future Forum research library →
Open Future Forum is a global executive community founded in Silicon Valley, whose network reaches tens of thousands of executives and investors worldwide. Its research program, including the Executive AI Leverage Report, the CFO AI Leverage Report, and the CMO AI Leverage Report, is built on first-party survey data from executives attending Open Future Forum events.