Nigel Glenday told the Masterworks finance team that 80% of the company's machine tasks had to be done by an agent, offshored or outsourced, and let them choose which.
The usual CFO question is which AI tool to buy. Glenday's argument is that the tools have already converged on intelligence, and that the work left is encoding how the business actually functions so a model can act on it.
"They're so intelligent, but they're only as useful as the context I can serve up to them."
Glenday is CFO and COO of Masterworks, which securitizes individual paintings — 530 of them, $1.3 billion of assets, investors in 130 countries, and 300,000 partnership K-1s a year that no vendor sells software for.
The full interview is covered here so you can skip it. 46 minutes of audio, 16 minutes of reading.
Here are the 11 lessons that matter.
👤 Guest: Nigel Glenday, CFO and COO of Masterworks, the platform that securitizes individual artworks into SEC-registered offerings
🎙️ Host: CJ Gustafson, a self-described recovering tech CFO who hosts Run the Numbers
📰 Published: 14 September 2026 on YouTube (Run the Numbers with CJ Gustafson)
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 46 min | ✅ Time saved: 30 min
Key Takeaways
AI did not reduce anyone's workload, it raised what gets asked for
Glenday built a financial model in an hour and said the old response to that speed was "I need to see 30 of these on Monday"
The models have converged on intelligence, so the differentiator is context engineering
He calls the missing capability relevance realization: a model can hold every org chart and still not know what matters
He runs six to twelve Claude Code terminal sessions at once and built his own tooling to manage them
An "A team" directory to relaunch them, and an EA agent that passes messages between their inboxes
Masterworks files 300,000 K-1s a year because no vendor sells a product for it
The stack went from Alteryx workflows to Python, and a summer intern filed the taxes two months early
What he wants from a 22-year-old hire is curiosity plus base Python and SQL literacy
He picked an ERP on whether agents could talk to it, not on features
AI content is cheap to make and expensive to read, which is why he will not ship it unedited
1. More Output, Not Less Work
Gustafson opened with the image that prompted the invitation: arriving at a dinner to find Glenday at a high-top table with his laptop open and two Claude terminals running.
Glenday's position on AI and workload is the opposite of the popular one. "This whole view that AI is going to reduce people's work. I just I don't see it," he said
What it actually delivers is operating leverage, and leverage raises expectations. He said more output raises what a boss or a customer asks for next, so the analysts and auditors people worry about are going to do a hundred times more work
His own example is from the other side of the change. He built a financial model for a project in about an hour, and said that in his banking-analyst days a managing director who knew you could do that would not say to take the weekend off
"It's I need to see 30 of these on Monday," he said
Gustafson said the two of them had gone into a 20-minute demo of the setup at that dinner, which is what the rest of the interview walks through
2. The Art Market's Recovery
Glenday last appeared on the show to explain securitizing artwork, so Gustafson asked for a state of the art market before turning to AI.
The recovery has taken hold. He said the market had an extraordinary high around COVID, moderated, and began recovering at the end of last year through the auction season and into this spring
Year to date, he said, auction sales are the highest since 2022
What moves sentiment is estates rather than economics. Big collections come to market when people die and heirs face tax bills — he cited the Leonard Lauder collection at the end of last year, with a large Gustav Klimt painting
His most useful correction is about what a price fall in art actually is. There is no forced sale, so if the market is soft people simply do not sell: "sometimes what is perceived to be like a draw down in prices in the art market is really what you're seeing is just a draw down in quality"
The comparison he offers a finance audience is the IPO window — you need management teams that feel good, bankers with a read on volatility, and then you go when the stars align
The supply argument runs the same way as dry powder in private markets. He said people talk about money waiting on the sidelines, but if there is nothing of quality to buy that is not market degradation, just absent supply
"If you've got that prized Rothko and you're living with it and the market is slow or whatever, interest rates remove it, you just won't sell it," he said
The supply of a dead artist's work only shrinks. He said the stock becomes fixed and then dwindles as pieces are donated to museums, and mentioned a conversation with SoundCloud's CFO about the same effect in music catalogs after an artist dies
The example from the past season: a Jackson Pollock from the Si Newhouse estate coming up at auction, from a category where perhaps a dozen major works sit in private hands
3. Creativity With a Price
Gustafson put the awkward question directly: Masterworks is in the business of valuing human creativity at exactly the moment AI has made everyone ask what creativity is worth.
Glenday's answer is that cheap content raises the value of the expensive kind. In an age where content can be created at close to zero marginal cost, he said, the value of things that are human-curated and have a human touch only rises
The asset itself is experiential, not only financial. He said there is no argument that standing in front of these paintings is a completely different experience, and that Masterworks has done far more over the last two years on events and museum loans because people want to see the collection
His analogy for the trend is consumer nostalgia. Millennials going back to corded headphones and rejecting Bluetooth; his own vinyl collection, hung in his office as much for the cover art as the music; Taylor Swift fans buying physical CDs
The financial case has not changed: art has appreciated over time, the market is large and global, and wealth creators across cycles have come to it
4. Context, Not Model Choice
Asked how he picks between GPT-5, Claude and Gemini, Glenday said the question is close to irrelevant.
The models are converging, and the intelligence is not the constraint. "The difference between these models being kind of useful or not useful is not so much the intelligence, but it's the context you're able to engineer around it," he said
He learned it by accident, brain-dumping into the tools in 2022 and 2023 and finding that the more raw context he gave, the more useful the output — and how fault-tolerant they are
He said he has stopped spellchecking his prompts entirely: "I'll type and speak to these things like a complete Neanderthal"
He still expects the output back grammatically correct and in the right tone, which he called a very different trade-off
His definition of context is broader than documents. Asked whether he means org charts, process maps, account mappings or tribal knowledge, he said all of it — a business is made of knowledge artifacts, including the things that live in people's heads
Enterprise search over that material is a first step, not the answer. He said semantic search across documents and email is incredible, but that it is not the same as the model understanding how the pieces relate
The framing he prefers is knowledge modeling, and it is the spreadsheet habit applied to a whole company. He said he loves spreadsheets because they show visually how metrics wire up to each other, and that a model needs the same thing across the business: "So how do I get from this number to how CJ actually impacts this number and what are all the different steps along the way?"
5. Relevance Realization
The shortcoming he names has a name, and he took it from a book. He cited The Blind Spot, which he attributed to Adam Frank, for the concept of relevance realization — the thing LLMs are bad at
His explanation of why is that the models have no lived experience. They operate purely on language, often code, so they have a very hard time teasing out what is important in a business
The anecdote that makes it concrete is about a colleague the model volunteered. Building a workflow for the media company, the model told him to pass a step to Cali, who runs recruiting
"It actually remembered somebody on the team from like a prior chat and just brought her out of nowhere," he said — it knew the org existed, and not what anyone actually does day to day
His image for the imbalance is memory against intelligence: "If we had the same like memory to intelligence ratio as these tools, like we wouldn't be able to remember how to like brush our teeth in the morning."
"They start from a blank slate. These tools need a map. They need a way to understand the world around them," he said
Which is why the CFO is the right person to do this work. He said the finance chief has a 360-degree view of the enterprise and is deeply involved in systems building, so is placed to shape the framing of context for these models
6. Twelve Terminals
Gustafson asked what is actually open on his machine.
He works almost entirely in Claude Code terminal sessions rather than a web app, which he attributed partly to nostalgia for MS-DOS in the 1990s and mostly to speed — he finds the terminal faster and more performant
The reason is parallelism. A web app puts you in one window with a sidebar of chats; separate terminal windows let each session hold its own context and its own function
The count: "It's really like anywhere from like half a dozen to like a dozen of these separate Claude Code terminal sessions running," he said, and he has had to build tooling around them to keep it organized
The agents he runs are departmental rather than general: an executive-assistant agent, a budget agent, document review agents, modeling agents for building a new financial model, and outreach campaign agents for specific projects
He also keeps a text-based CRM instead of a real one for some work, on the grounds that a lightweight text file gives a model an efficient view of the state of a project
Gustafson said the interface Glenday showed him was Ghostty; Glenday explained it is a terminal emulator, written in a particularly performant language and customizable with tabs and other affordances
7. The A Team and the EA Agent
The problem he solved first was relaunching. Each session ran out of its own local folder holding that agent's context and reference files, so closing the laptop meant reopening a dozen folders by hand
His fix was "this simple little tool called the A team," a directory that pops each session open and renames the window so he can navigate between them
The second problem was that the agents could not talk to each other. He wanted the agent managing his tasks and calendar to be able to hand work across, so he built an EA agent as an orchestrator with inter-agent messaging
"They all have inboxes and it's like very simple," he said, while acknowledging that frameworks exist that do this more natively
What he is engineering out is the upload step. Asked why Claude Code rather than Claude projects, he said the word in the question was upload: "Like I don't want to like sit there and upload stuff"
That is also what made the old finance stack frustrating, where a workflow could run end to end except that someone had to upload a CSV, which breaks the chain
The next question he is asking of every internal tool is whether an agent can use it. He gave the example of a colleague who runs payments orchestration and built a Slack automation for customer-payment questions, and said the follow-on question is whether it is exposed by an API so an agent can consume it
8. 300,000 K-1s
Gustafson said he had seen a LinkedIn post from the Masterworks controller about processing something like 300,000 K-1s, and asked how anyone approaches that.
The company had no choice, because the product creates the problem. He said Masterworks now has around 530 paintings and $1.3 billion in total assets, each individually securitized, each going through the SEC process with its own financials, each a partnership with its own K-1s — and trading on top, so a painting can have many partners
No vendor sells the answer. There was never an option to go out and buy something that files 300,000 K-1s, he said, which is why the company's habit of writing code predates AI
The path ran through a visual tool before it ran through code. He said the work started with Alteryx workflows, familiar to people out of the Big Four accounting firms — pulling partner information out of the back end, transforming it, populating the K-1s and the returns — with a stack of Alteryx, Python and Thomson Reuters ONESOURCE
AI collapsed that. "So now it's completely Python and now we can have an intern do it," he said
A summer intern working with the controller filed the taxes two months early
The hiring pattern he credits is finance skills paired with technical ones. He said the team recruited people with finance, tax and accounting skill sets who were also broadly proficient technically — including people who left finance for software development and came back
9. Hiring the 22-Year-Old
Gustafson asked what he would look for in someone straight out of college.
First, curiosity. He said he has always indexed on high intellectual curiosity and the ability to learn quickly, and that AI tools are excellent for learning rapidly if nothing else
Second, the literacy to use tools built by engineers. "What I would say though coming to accounting they're fundamentally coding tools right they're engineering tools that are designed by engineers," he said, so base literacy in Python and SQL is fundamental
He learned it himself by translation rather than by training. Sitting near a data analyst who did everything in R, he picked up a book on R for Excel users and used it to map his mental model of a spreadsheet onto code: what a variable is, what a table is, what a time series is
Gustafson supplied the canonical example, Wes McKinney building the pandas library in Python for manipulating time-series data
Glenday's gloss is that the library exists because spreadsheets are the wrong tool for some jobs. Resampling monthly data to yearly and interpolating the difference means average formulas and dragging in a spreadsheet, and is one argument in a function in code; transposing data is one character
The apparent contradiction he finds funny is that AI is supposed to remove engineers. "I find it hilarious how we're saying that AI is going to get rid of engineers, but at the same time, we need way more people who are kind of like almost quasi engineers," he said
On the five-year split between finance judgment and technical skill, he would not trade one for the other. Human judgment is not going away, he said; he wants people working at "the bleeding edge of the company", on new products, new divisions and scale, and the technical skills are the tools of the trade
His historical frame: the spreadsheet was the killer app that sold the PC in the late 1970s, and tool proficiency has moved from the slide rule to the spreadsheet to coding
10. Choosing an ERP for Agents
Asked what he has actually bought, Glenday named the ERP migration to Campfire, which he said was not on anyone's bingo card given the company had migrated two years earlier.
The ultimatum came first, and the team's answer shaped the purchase. He asked how to get to 80% of machine tasks automated, and the feedback was that they did not need a fancy agent doing everything — they needed the frictions removed, and most of them were manual steps where something had to be uploaded
What they wanted, he said, was straight-through processing
It is the only ERP transition he has seen a team be excited about
His definition of AI-native is API surface, not features. "What you really mean is like can any AI agent talk to it, right? Does it have like a legible API surface? Does it have an MCP? Can I integrate it in my other like Python workflows?" he said
The payoff is that his own scripts can reach the financial data. If he builds a script with the semantic layer around it, he can grab that data without bringing in outside resources — "Don't have to hire a consultant for another like $100,000"
11. The Work Slop Problem
Gustafson's closing question was where Glenday tried AI and went back to doing it by hand.
Writing in his own voice is the one that does not work. "I care a lot about great writing," he said, and despite giving the tools many examples, anything that needs his voice is really hard to get out of them
He cited a Brett Taylor post complaining about having to strip em dashes out of his writing
The failure mode is verbosity. "They're really verbose," he said — half the time his response to an output is to ask for 80% fewer words
He read out a slide from another company's all-hands, shared by Alex Cohen, which he endorsed completely: the company does not ship slop, and needs to get better at not shipping purely AI-generated content without human taste and filtering
The line he picked out: "AI content is very cheap to create and incredibly expensive to consume. When you create something with AI and pass it to co-workers, you're putting the burden of reviewing the AI on them, not you"
His own formulation is subtractive. He said value is created by addition by subtraction: you can have AI create as much content as you want, and the work is figuring out what is important, surfacing the right context, and applying human taste and judgment
Bonus Insights
Glenday says he declares "tab bankruptcy" periodically, and described himself as a hotkey nerd; Gustafson said he was "today years old" when he learned the tab-search shortcut
Gustafson named the agent directory better than its author did, asking whether the A team hop out of a black van
Masterworks was founded in 2017, which Glenday used to make a point about legacy data: nobody in 2017 was organizing their systems so that a model could read them later, and that retrofit is the work everyone is doing now
The customer-service agent he is building is not about drafting emails, he said — the hard part is a knowledge base covering questions like when a payout lands after a painting sells, or how to send a few hundred dollars to an investor in Moldova
Gustafson's own example of the deterministic-over-clever principle was a scraper that emails him when CFOs he knows post open roles: written with AI, run as a scheduled job, and using no tokens at all
Glenday's bottom line is that the constraint on AI in a finance function is not model quality or budget but whether anyone has encoded how the business works — and that the companies getting value out of it are the ones whose people were already close enough to the systems to build what they needed.
Products, Companies & Tools Mentioned
Masterworks (Glenday's company; around 530 paintings, $1.3 billion in assets, investors in 130 countries and 300,000 K-1s a year)
Claude Code (The tool he lives in, running six to twelve terminal sessions at a time with separate context in each)
Ghostty (The terminal emulator behind the setup, chosen for performance and customization)
Campfire (The AI-native ERP Masterworks migrated to, picked on whether agents could reach it through an API)
Alteryx and Thomson Reuters ONESOURCE (The original K-1 stack, since replaced by Python)
pandas (Wes McKinney's Python library for time-series data, Gustafson's example of code beating spreadsheets)
Slack (Where his team's payments automation lives, and the surface he says has become the default for interacting with agents)
SoundCloud (Whose CFO told him about the catalog spike when a musician dies, the music version of the art market's death effect)
Books & Resources Mentioned
The Blind Spot – Adam Frank, as Glenday named it (The source of "relevance realization", the thing he says language models are worst at)
R for Excel Users (The book that taught him to map spreadsheet thinking onto code, variable by variable)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: