Columbia Business School Puts Claude Code on Its Faculty Workshop Series

Columbia Business School scheduled a faculty workshop series this summer that teaches professors to use Anthropic’s Claude Code for their own empirical research. The two-session workshop sits inside the school’s AI in Business Initiative and follows an earlier session led by Columbia’s Santiago Balseiro with colleagues from Google. Session 1 ran on June 24 from 2 to 4 p.m., with Session 2 set for July 27 from 11 a.m. to 4 p.m. Anthropic’s own survey of social scientists explains the timing: 81 percent of quantitative social scientists have tried AI chatbots for research, while only 20 percent regularly use coding agents like Claude Code in their workflow as of early 2026.

The two named leads work at opposite ends of the school: Jesse Schreger, a finance empiricist who studies sovereign debt and exchange rates, and Malek Ben Sliman, a marketing-side machine learning practitioner who built models for Sotheby’s. Both are now teaching faculty how to run their own empirical work through Claude Code, in a moment when only 20 percent of social scientists regularly use coding agents on their research. Reported below: the workshop’s curriculum, the Anthropic data on faculty coding-agent adoption, and the privacy and access rules Columbia has published for faculty users.

What Each Session Covers

The workshop is built around a split between a gentle on-ramp and a full working day. Session 1 ran on June 24 as a 90-minute introduction. Session 2 on July 27 is the hands-on day for attendees who bring their own data.

The first session’s stated topics, per the workshop’s full session-by-session outline, cover installing and configuring Claude Code, the different ways to work with it (terminal, IDE integration, plus a “GUID” reference), pricing and how Columbia’s access works, model selection, and the four operating modes: plan, accept edits, auto, and skip-permissions. It also addresses data sharing and privacy, and demonstrates integration with R, Stata, and Python alongside version control with git. The organizing team asks attendees to come to Session 1 with suggested “assignments” ready, meaning their own data, so Session 2 can move straight to execution. The session’s stated takeaway is to “walk out with Claude Code installed and integrated into a research workflow alongside existing tools (Stata, R, Python), with a clear sense of how to start using it for one’s own research.”

The second session is where the workshop’s promise gets concrete. Faculty are expected to execute their own research in real time and discuss how to extend Claude Code with MCPs, subagents, skills, plugins, and API calls. The session’s stated takeaway is to “leave the workshop knowing how to produce original academic research with an agent-based workflow,” with the goal that some attendees “will even have a new result to build on for a full academic paper.”

Who Is Doing the Teaching

Jesse Schreger is the Ann F. Kaplan professor of macroeconomics in the Economics Division at Columbia Business School, and his work is primarily on international finance, sovereign debt, and exchange rates. He is a faculty research fellow in the National Bureau of Economic Research’s International Finance and Macroeconomics Program. Malek Ben Sliman is a Lecturer in Discipline in the Marketing Department, with research interests in machine learning, computer vision, and NLP applied to art valuation, social networks, marketing analytics, and online retailing, per his Columbia faculty page. Both are now teaching the workshop inside the school’s AI in Business Initiative.

Ben Sliman previously built models at Sotheby’s to track artists’ prestige and predict auction prices in real time. More recently, he worked on the data science team at Oden Technologies, an Industry 4.0 startup focused on analytics for manufacturers. The pairing puts a finance empiricist who routinely works with messy macro data next to a practitioner who has shipped industry models. The workshop’s earlier incarnation was led by Santiago Balseiro, a research scientist at Google and the George E. Warren Professor of Business at Columbia, with “a couple of his colleagues at Google.” Recordings and slides from that earlier session are to follow, per the AI in Business Initiative.

What Anthropic’s Data Shows About Faculty AI Use

Anthropic has been measuring this ground from the vendor side. Its analysis of 74,000 educator conversations on Claude.ai, drawn from May and June and paired with qualitative work at Northeastern University, produced a snapshot of how higher education professionals actually use the tool. Anthropic filtered conversations associated with higher education email addresses down to educator-specific tasks like creating syllabi, grading assignments, and developing course materials.

The most common use was developing curricula, at 57 percent of the conversations Anthropic analyzed. Conducting academic research came second at 13 percent, and assessing student performance was third at 7 percent. Most tasks sat toward the augmentation end of the spectrum, with university teaching and classroom instruction at 77.4 percent augmentation and managing educational institution finances and fundraising at 65 percent toward full automation. Writing grant proposals was also heavily augmentation-weighted, at 70 percent. Grading was the outlier: faculty used AI for grading less often than for other tasks, but when they did, 48.9 percent of those conversations were automation-heavy.

Anthropic also surveyed 22 early-adopter faculty at Northeastern to add color to the conversation data. Faculty described AI as taking over tedious tasks while staying in the loop for work that requires significant context, creativity, or direct student interaction. The split between augmentation and automation in faculty work shows up consistently in the data. One faculty member in the report frames the change this way.

What was prohibitively expensive (time) to do [before] now becomes possible. Custom simulation, illustration, interactive experiment. Wow. Much more engaging for students.

This is from a Northeastern faculty member quoted in the report, on what they now build with Claude.

The Productivity Gap the Survey Spotted

Anthropic’s separate survey of 1,260 social scientists, fielded in February and March 2026, asked sharper questions about whether coding agents actually move research output. The headline: 81 percent of quantitative social scientists had tried using generative AI for research at all, but only 20 percent regularly use coding agents like Claude Code more than once a week.

Where the survey gets specific is the productivity question. Coding agent users posted around 75 percent more working papers than non-users in the same discipline and career stage, with around 10 percent more empirical projects started. The pattern does not extend to journal submissions: coding agent users were not sending more new papers to journals or resubmitting faster. The authors flag this as a descriptive comparison, not a causal one, since researchers who self-select into coding agent use are also more productive in other ways.

• 81% of quantitative social scientists had tried generative AI for research
• 20% regularly use coding agents like Claude Code
• 86% of coding-agent users report Claude Code use, 31% report Codex use
• Around 75 percent more working papers posted by coding agent users in the six months before the survey
• Tenured professors were less than half as likely as doctoral students and postdocs to adopt coding agents

The same survey asked what coding agents were actually being used for. Coding and editing prose came first: 97 percent of coding agent users and 77 percent of other AI users reported generating code with the tools, and editing prose was the second most common use. Only about a third of all AI users in the survey had used the tools to draft original prose at all, and that pattern held across disciplines except economics and management, where drafting prose was more common.

The split between coding agent users and other AI users also shows up in what they actually output. Coding agent users were more likely than other researchers to start new projects, post working papers, and submit grant proposals, per the survey’s adjusted estimates. They were not more likely to push papers through to journal submission or resubmission. Anthropic frames this as evidence that coding agents help projects get up and running, with the last-mile polish still largely a human job, and for a working framework on how human review fits into agent-led coding, Andrew Ng’s recent letter on three AI coding loops applies a similar human-in-the-loop argument to software engineering.

Who Has Adopted Claude Code, and Who Hasn’t

The survey’s other finding is the harder one. Adoption among social scientists is sharply uneven, and the pattern runs along axes faculty rarely discuss in meetings. The uptake map by discipline is the first place the disparity shows.

• 39% of economists use coding agents regularly
• 25% of political scientists
• Single digits for public health, education, and communications

Adoption is also skewed by career stage and institution. Roughly a quarter of doctoral students and postdocs use coding agents at least weekly. Among tenured professors, that rate falls by more than half. Researchers at the top 25 universities, defined by Nature Index Leading Institutions, were 40 percent more likely than others to use coding agents.

The gender gap is the number the report flagged most pointedly. Researchers with typically male names adopted coding agents at more than twice the rate of researchers with typically female names, with the gap classified by name. The gap persists across disciplines and career stages, including among researchers who have already tried using AI for research. The authors frame this as a productivity gap that is already starting to compound, since coding agent users post more working papers and submit more grant proposals in the same window. Coding agent adoption at this point, the survey suggests, is widening that funnel in one direction at a time.

What Columbia Tells Faculty About Data Privacy

The workshop’s first session flags privacy as a covered topic, and Columbia’s published guidance for Claude for Education is unusually specific. Claude for Education is available to approved Columbia faculty and staff, with the license costing $300 annually per user through a chart string. Anthropic does not use conversations or uploaded files to train its models, and all conversations are encrypted in transit and at rest. Claude is not approved for use with Sensitive Data, as defined by the Columbia University Data Classification Policy, per the university’s Claude for Education guidelines. Office hours for the program are on a summer break and resume in September 2026, per the same page.

For a faculty workshop, the practical implications are concrete: which projects run through Claude Code, which datasets travel into it, and which stay outside the system entirely. Faculty running a real research project through a coding agent without a live escalation path will need to plan around that gap. The point of running it on your own data, per the workshop’s framing, is to find what fails before the journal deadline does.

Where the Series Goes From Here

The workshop’s organizers said in the event page that they are planning “a series of additional workshops on using AI for academic research in the summer and fall.” The current two sessions will be recorded, and the team has said they may repeat the series in the fall if there is enough demand for those who missed it over the summer. Anthropic’s coding-agents research team is also preparing results from a randomized experiment providing researchers with Claude Code access, which would be the first causal read on whether the productivity gap the survey spotted survives a controlled study.

The next reference point is the curated set of case studies emerging across vendors, which include both research-stacked platforms and broader scientific workbenches. For one example of how the same agent pattern extends into adjacent research domains, Anthropic’s separate Claude Science workbench pushes the same loop into drug discovery. Anthropic’s coding-agents research team will publish experimental results later in the year. The workshop itself, before either of those result sets lands, is the closest the school will get to a live test of what faculty do with Claude Code.

The session design is to take a researcher from “installed but unused” to “running an original analysis on their own data” inside a two-week stretch. Whether faculty actually do that depends on the privacy posture, the data pipeline, and the time each attendee is willing to invest. The answer is what the fall repeat is being built to test.

Frequently Asked Questions

Who can attend the Columbia workshop?

Faculty and doctoral students, per the event page. Session 1 was framed for faculty who have not yet integrated Claude Code into their day-to-day. Session 2 is open to anyone interested in using AI for research.

How much does the workshop cost?

The event itself, listed under the AI in Business Initiative, does not list a fee. Columbia’s separate Claude for Education license costs $300 annually per user through a chart string.

Why is the workshop focused on Claude Code?

The session description names Claude Code and lists its specific features. Anthropic’s coding-agents survey found Claude Code is the most-used coding agent among social scientists at 86 percent, against 31 percent for Codex.

Does the workshop cover data privacy?

Yes. Session 1 explicitly addresses data sharing and privacy concerns. Columbia’s published guidance notes Claude is not approved for use with Sensitive Data under the university’s data classification policy, and that Anthropic does not use conversations or uploaded files to train its models.

Will the sessions be available afterward?

Organizers say the two-session workshop will be recorded and may be repeated in the fall if there is enough demand for those who missed it over the summer.