New builders, stable foundations
September 27, 2026 · OSASI Commentary
In consecutive weeks in July, Value Driven Data Science gave data professionals two pieces of advice. First, stop reading about AI and build something with it, because the leap is smaller than it looks. Second, do not build your future entirely on rented models, because today's prices are subsidised. Together they describe a shift in who gets to build: those with new insight and new methods, and the recipe to put the best of innovation on a new but stable foundation on an even keel.
How a brief is made
A member reads an article or listens to a podcast and sends the snippets, key thoughts or concerns tied to a passage or a timestamp. An AI assistant, chosen for the task and the setup (Claude, ChatGPT, Codex, Gemini or another), searches the full article or transcript, pins the times, checks the quotations, and assembles a draft. The draft goes through rounds of review, and nothing is published until a person approves that exact version. The cues on each post start the player at the passage discussed. Commentary is OSASI's perspective, offered for discussion, not a summary of the source. Each brief is co-authored by a person and an AI assistant. AI assistant for this brief: Claude (Anthropic).
Source podcast · Value Driven Data Science
This episode prompted the commentary below. It is one perspective, offered for discussion.
Value Driven Data Science · July 23, 2026
Episode 115: Evolving Your Data Career for the AI Era
Listening highlights · Dr Genevieve Hayes with Nicholas Kelly of Delivering Data Analytics.
Tap a time above to start the player there. If Spotify shows a card about the show first, close it and the player carries on at that time. Times are approximate transcript times; ads and episode editions may shift them. This plays the full episode, not an extracted clip.
Open episode in Spotify · Episode page · Timestamped transcript
Who is talking
Dr Genevieve Hayes hosts Value Driven Data Science. Her guest in both episodes is Nicholas Kelly, co-founder and chief AI architect of the consultancy Delivering Data Analytics and author of The AI-Driven Data Team. Episode 115, on July 23, asks how data professionals should evolve their careers for the AI era. The Value Boost a week later asks what they need to know before, in its title's words, "the AI free ride ends".
Episode 115: build something
Kelly's change of direction began two years ago, when a client told him: "ChatGPT can do 50% of what you do right now." He chose to get ahead of it, and one of the first things his firm did was buy a GPU rack, "crazy very expensive", because "to have a seat at the table, that was one of the things we said we needed to do."
Hayes places the moment in a longer history. Her analysts once hand-coded their reports in SAS, until finance teams asked why they were paying a large SAS bill when R and Python were free. She expects the same question of today's dashboard licences. Kelly agrees: "Certainly in the case of Tableau, that's their struggle", with Power BI already bundled into Microsoft licences and cloud inference bills adding to the pressure.
Halfway through a training day for a BI team, Kelly abandoned his slides: "I need to practice what I'm preaching here." From a wireframe one of the team had drawn, he built a working app in about an hour and a half, on a laptop with nothing installed. His case is that a data professional already does two hard things, working with data and working with the business, and that adding AI-assisted building is a smaller step than teaching a developer both, "and I'm saying this speaking as a computer scientist". Hayes sets the conditions: the fundamentals of statistics and databases, and "data security, privacy, governance, all those things, because otherwise it's just gonna be a ticking time bomb."
Then comes the correction. Hayes remembers Tableau's promise that anyone could build a dashboard; after "a little go with it", the work went back to the data team. Kelly agrees: "I think initially people will think, 'Oh, this is easy. I can build an app.' And then they'll realize actually no", because testing, change management and the people work still have to be done. His advice for a first project is to make it personal, "something you actually care about."
What OSASI hears
Where the show says data analysts, read appraisers too. An appraiser working to data science principles is doing the same job with a different subject. The advice to try it first is the advice OSASI would give: until you have dug in, you cannot know what the fuss is about, let alone weigh the pros and cons.
The episode's arc matters as much as its advice. It starts near anyone can do this and lands at you can, but it is still hard. Clients and end users could build their own tools, but as the tools grow more capable, so does everything around them: security, governance, privacy, and the widening scope of what the tool itself does. AI can carry a team from one platform to the next, from a subscription service to Tableau, to Power BI, to software built in-house. The specialties remain. What shifts is who runs the tool. The experts with the domain knowledge now run it too, which frees those more interested in putting the insight to work. Hayes expects some will give it a go and hand the work back. Some will try it and decide it is still not for them. They will only know by trying.
Episode 116: before the free ride ends
The follow-up takes up the question the first episode touched on, running AI on your own machine, and puts a price on it. "We are currently living through a period of subsidized intelligence," Hayes opens. OpenAI, she says, is projected to lose $14 billion this year, and both it and Anthropic are expected to go public within a year, after which "it's only a matter of time before the cost of using AI models starts going up." Kelly would "almost categorize it as a catastrophic mistake" to build entirely on frontier models: trusting a fixed monthly cost "you're going to hope is gonna persist at that price, and it 100% is not gonna do that." Cost is the main reason but not the only one. He expects regulated industries to require more work to stay on premises.
His answer is hybrid, not all or nothing: "You don't necessarily need to run inference locally." Hayes sums it up as owned hardware running an open model for development and for "highly sensitive applications, so HIPAA data, financial data, things like that", with frontier models in production where privacy permits. The open models are a version or two behind the frontier, Kelly says, "but that's not gonna matter in six months." For experimenting, he puts a small AI workstation at about $5,000; for production, anywhere from $50,000 to $200,000.
Follow-up · Value Driven Data Science, a week later
The same guest returns to take up the question the first episode touched on: running AI on your own machine.
Value Driven Data Science · July 30, 2026
Episode 116: [Value Boost] What Data Scientists Need to Know Before the AI Free Ride Ends
Listening highlights · Dr Genevieve Hayes with Nicholas Kelly, a fourteen-minute Value Boost.
Tap a time above to start the player there. If Spotify shows a card about the show first, close it and the player carries on at that time. Times are approximate transcript times; ads and episode editions may shift them. This plays the full episode, not an extracted clip.
Open episode in Spotify · Episode page · Timestamped transcript
The haves and the have-nots
The subsidy is today's question. Behind it is a larger one: who will have access to premium compute once the subsidy ends? If the best models, and the hardware to run them, are priced for those who can afford them, the division between haves and have-nots that runs through incomes could creep into the white-collar professions as well, ours included: the well-funded firm with frontier models and its own machines on one side, the sole practitioner with neither on the other.
OSASI's answer is member-driven diversity. We want to build on collective intelligence, not just on opportunity for the fortunate: shared capacity, shared methods and shared learning that no single appraiser could afford alone.
The competition ahead is real. Truly evidence-based real estate appraisal analysis will crush the old, deficient methodologies and expose the weak. The new way will be hard work, even with the technology, because the same technology will demand transparency, whether through the conscience being built into the AI models themselves or through the critical oversight that will be demanded of those who use AI.
That conscience is already visible. Anyone who has done in-depth analysis with a frontier model will know that it tends to raise flags when it sees them, calling out conflicts or inconsistencies in your own work without so much as the slightest prompt. Getting past that takes a specific instruction to overlook or ignore what the model found. That will not be a good look, and keeping such instructions out of the workfile could be an ethics issue in its own right. Trust in AI is a two-way street.
Nor will it be possible to avoid that scrutiny by keeping AI to the "mundane" tasks. As its use grows, and with virtually every analyst using it, the question will no longer be whether an appraiser used AI but how. Use, and disclosure of that use, will be expected.
What this means for OSASI
Appraisal files hold exactly the kind of data Hayes calls highly sensitive: financial details, client identities, the inside of people's homes, all held in confidence under the profession's own rules. The hybrid split the follow-up describes maps onto that directly. Confidential files, and the learning that comes before trusting any tool with them, belong on capacity the firm controls. Work that carries no client data can use the frontier models while they remain cheap.
The previous brief asked how an appraisal firm budgets AI for 2027 when nobody can price it yet. Kelly adds the other side of that ledger: a machine bought before prices rise is a cost a firm can see and plan for. And his advice for a first project, something personal and low-stakes, points to where appraisers should start: somewhere to try the tools on owned hardware before a client file is involved. That is part of what the sandbox open-weights AI expression of interest is for: owned capacity shared among members, not reserved for the firms that can afford their own.