Hey, welcome back to The Interline Podcast. If you’re listening to this, the odds are that you’re spending money on AI. At the time I’m recording, 55% of American businesses are, and those figures are probably reflective of what’s going on in the EU and the UK as well. Then let’s add in the fact that The Interline’s audience is made up of fashion professionals who have a vested interest in staying near the cutting edge, and you can see why that number would be selective.

Now, to be clear, you might not like AI, or you might not be seeing a return on your spend. That’s a different story, and it seems clear at this point that cultural attitudes towards AI and business spending on it are on divergent tracks. That’s actually one of the main themes of our 2026 AI Report, so if you haven’t already, I’d encourage you to grab your free copy of that for additional context for this conversation.

But while paying for AI might be a binary on/off state, what you’re paying for, what results you’re getting, which capabilities are the most deployed across the huge surface of image generation, coding, general knowledge, Cowork and so on, and which vendors you’re buying it all from, is a much more complicated mix. And it’s also a mix that has been hard to find numbers to substantiate. There’s all manner of consumer surveys about how people feel about AI on the street and how they use it, and our AI survey from the AI report I just mentioned is focused on extending some of that sentiment analysis and maturity benchmarking straight into fashion.

But when you’re looking at a potential transformation on the scale that AI advocates are promising, you need bigger than consumer, you need bigger than fashion, you need the broadest and most technically robust statistics you can find. For us here at The Interline, and for a lot of media, the AI Index from Ramp has become the go-to source of that kind of economic data, and of expert analysis of it. That data set is behind a lot of the articles you’ve read about AI adoption, about AI spend per employee, and about the rapid ascendance of Anthropic to lead the enterprise market after OpenAI had dominated for so long. That index is put together by Ramp’s Economics Lab, which is a four-person team embedded in a much larger financial technology company.

That team includes an economist, a director of applied science, a head of analytics and data science, and a lead economist. So, not the kind of team you usually find working for a fintech company, in other words. Today’s guest is that lead economist, Ara Kharazian. He used to lead economic research at Square, and you might have seen him in the New York Times, on NPR, or in the Financial Times. He’s become, in a pretty major way, one of the most visible faces of the economic side of AI. And as companies across fashion continue to frame AI less as an open-ended possibility horizon and a money sink, and more as a capital expenditure that needs to deliver a quantifiable return, more and more brands are interested in the commercial and the behavioural side of things that Ara and his team cover.

So I asked him to come on the show when we were putting the finishing touches to our AI Report for the year, so our readers and listeners can get the most complete picture we can offer of AI in fashion — both from directly inside the industry, from our survey and our analysis, and from a much higher and broader level, from the economic research that Ara and his team have been doing. So, here we go.

NB. The transcript below has been lightly edited.

Okay, Ara Kharazian, welcome to The Interline Podcast.

Thank you, Ben, for having me.

Not at all — I’m excited about this one. We’ve referenced data that you have had a very direct hand in producing multiple times, I think, over the last year or so, so I’ve been looking forward to this discussion. Now, we start these shows pretty universally by doing two things: we ask the guest what their day-to-day life looks like, and we ask them to define something.

Your day-to-day is really interesting to me in a couple of ways. The first is that you work for a tech company that does spend management — accounts payable, employee cards, expenses and so on. None of which, no offence, is outwardly that exciting. It’s not that thrilling unless you work in finance or HR. But you’re also a fully fledged economist of the kind you’d normally find somewhere else entirely. I would expect people like you to be working for banks, or advising on monetary policy at a think tank, or something like that.

So that’s one piece that I find interesting. The second is that, weirdly, we’ve had a number of guests recently who are also in the media business in a way, despite you not thinking that from their job title, because you produce written analysis that people then bookmark and come back to and put a lot of stock in. I’ve heard people say that your research, and the AI Index in particular, accounts for something like a third of all of Ramp’s non-platform traffic. You’re routinely cited on the big news. There’s a Claude connector I have live in Claude Cowork and Claude Code where I can go and pull live data from the Ramp AI Index.

There’s just a ton about your day that I’m interested in. So tell me how you got here — that’s one thing. How do you end up as an economist working for a tech company? And tell me what you actually do when you sit down in the morning.

The joke that we used to make about this job is that when I was hired, it was because rates were way too low, and the tech company could shell out to hire this kind of researcher who’s not even focused on the product. But once you think about our product, and you think about the dataset that we might have access to, I think the value of this job becomes very clear. So Ramp is a corporate card and bill pay product. It’s an AI finance platform more broadly, but if you think about the core dataset we might have access to, it’s all coming from the business spend that’s occurring on our platform. So that’s also the dataset that I work with.

Think about the universe of business spend — essentially anything except for payroll. Now think about other parts of the economy that already have access to great datasets. You see consumer credit card companies routinely producing research about consumer economic trends using their unique datasets. You see all these housing platforms on the internet reporting on the price of a house in this or that city. So in our economy, there are generally a lot of great private sector datasets that are used to drive important decision making.

You buy a house; you go applying for a job, you want to check the salaries online, you can do that with something like Glassdoor. There is no dataset, at least no public dataset, that tracks where and how businesses spend. And you can think of that as the main value proposition of the research that we do at Ramp. We collect this data, and everything we publish is aggregated and anonymised, so there are no privacy concerns with whatever we release. But we’re producing research about how fast-moving businesses spend, ideally with the goal of helping other businesses make similar decisions, and make those decisions with a better dataset.

So lately, in the age of AI, that means: how do you buy AI for the first time? This is a new and nascent technology for which there isn’t a great playbook for how to integrate it into your very complex business. And when you go online, it’s also the kind of sector that attracts a lot of pundits who may have a profit incentive for the kind of advice they’re trying to give you. So one place we’ve tried to start is: we see how all these businesses are paying for AI to start. Can we publish those kinds of trends to then help other businesses get on those same adoption curves, and then drive similar growth through AI adoption?

Who are they buying AI from? What’s the difference between OpenAI and Anthropic? And then, what does it really mean to use AI effectively? Is it just a chatbot for everyone in the workplace, or is it using something more advanced, like a coding agent? So that’s the general focus of my work: how do we publish this kind of data about how businesses are operating today and how they’re changing, and can we ideally make some recommendations so that other firms can adopt those same changes and grow faster.

There’s a point in there that I 100% agree with. I think in fashion as well, there’s been a long history of companies wrestling with the idea of best practice. This is very idiosyncratic to the industry, but it’s a market made up of brands that have spent a long time — in some cases centuries — operating as islands and doing things their own way. Technology responds better, in some cases, to doing things in a best-practice way. There’s more of an argument to be able to say, well, everybody is doing this, and this is why they are doing it, and it is the most effective way to do things, so you can benefit from that.

So I think the data you produce is incredibly valuable, as we’ve already talked about. And I think for fashion in particular, this shift to AI is an important juncture point, because it represents that difference between the way the industry has historically thought about its own processes and its tools, and what is now an onrushing technology wave — and treating that as an opportunity to say, well, if everybody is doing things a certain way, that’s probably the best practice. Fashion has always had that kind of push-and-pull relationship with best practice that I don’t know that other industries have necessarily had.

Well, I agree with you. You often see the same trends in tech, though, too, where large software companies — when talking about B2B SaaS, right — become very entrenched in a specific sector or for a specific business process. Even if it’s not necessarily the most up-to-date tool for a task, there does tend to be a lot of vendor lock-in and there does tend to be a lot of stickiness. And ultimately, particularly when you’re thinking about these technological adoptions in terms of firms’ operations, smaller firms are much more nimble. They are much more able to quickly move between different software vendors, or adopt new technologies fairly quickly, and to onboard all their employees with new technologies.

That is ultimately what enables the same startup culture that has driven the tech sector for a few decades now: large companies are rarely safe when they stop innovating. But we shouldn’t have this impression in our heads that because the tech sector created these technologies, they are immune to the kinds of changes that may affect the market share of even the largest players.

That’s an interesting point. And you just mentioned vendor lock-in there — that was actually the term I was going to ask you to define. There’s a thing that it means to me as somebody who’s been around for a couple of turns of technology in fashion, sixteen years or so. The way that traditional enterprise tech has been rolled out over the previous decade or so has been defined by long custom implementations with a lot of services days — i.e. consultant and advisory time — baked into it, supported afterwards by maintenance contracts.

But very heavily along the lines of: we will customise and bespoke things for you, with the goal of building up ARR as much as we can as a technology vendor. And to do that, we will architect all of this in a proprietary data model. We will have to build bespoke integrations and portals for your suppliers, and so on. You get the idea.

That’s what vendor lock-in has meant to me. That’s how it gets its hooks into organisations, at a cost level as well. That model I’ve just described is very expensive. Companies in fashion spend way too much on building systems of record to ever move away from them, even if over time you actually end up with less data in there than you thought you would have — a bunch of it being abstracted out because people build workarounds and things. Long story short: when you chuck a million dollars at making something work, you’re suddenly very incentivised to use it as the platform to make everything else work from there.

Now, the opposite side to what I’ve just talked about there would be somebody saying, well, sure, but the technology was really important, so we had no other choice. The technology was that vital to the growth of the company, and the outward market and the world were changing. We had to do it that way. That’s what gets me to AI here, because I think, like I said before, if this is the most pivotal tech-is-changing-what-it-means-to-be-a-company moment, then it feels like technology companies should be champing at the bit to run that playbook again, to lock up their share of it. Right?

And that’s where my mind goes when we talk about vendor lock-in. You said, I think it was in one we opened last month or so, that unlike traditional software, AI doesn’t really have vendor lock-in, and that the most advanced AI shoppers you see in your data run something like eight AI vendors at once. I’m going to hazard a guess that the big AI labs — your Google, OpenAI, Anthropic and so on — would like to get that down to one or two. What is it that you think makes that unlikely? What is it that you actually think makes AI different when it comes to vendor lock-in, compared to what I think of it as in traditional software?

Well, think about the way that you use AI in your day-to-day life, right? All these traditional features of a product or a piece of software that a vendor can build to keep you on the platform don’t really exist in AI. It’s not like Salesforce, where it’s an extremely complex technology that involves building integrations through all of your workflows and training all of your employees to use Salesforce exclusively to track customer interactions. AI is not like that, in that it’s actually pretty easy to switch between different models — if not advantageous to be using multiple models at the same time.

Not to say that there isn’t an advantage to having high market share and holding it, or that being early in the marketplace may entrench a firm as one of the defaults. That really is kind of the story of OpenAI and Anthropic: OpenAI was, for a long time, this consumer default, in that when you thought about AI, you were thinking about ChatGPT. And because AI is this new and nascent technology that businesses are buying for the first time, as well as consumers — if you’re a business buying AI for the first time, you’re probably going to buy OpenAI. At least that was the case from 2023 through most of 2025.

At least in our data in the Ramp AI Index, we found OpenAI’s adoption rates shortly after launch, at least in the enterprise, shot up to about 23 or 24% as of May 2023. So around a quarter of businesses in the United States were already paying for OpenAI. Most businesses had never heard of Anthropic, and probably weren’t aware of Google’s offerings in that area. And then it hovered around there and grew up to about 40%, until Anthropic started to make significant inroads into that market. Anthropic — popular with technology companies and particularly early AI adopters, in a much smaller way — started to grow its market share by several percentage points month over month, which you don’t see in most sectors.

And then toward the end of 2025 and the beginning of 2026, it skyrocketed in terms of adoption rates, ultimately passing OpenAI in May 2026 and becoming the most popular AI company, at least amongst businesses. So you have a story here of a company, OpenAI, that was the first mover, became the consumer default, and in many ways became the business default. No one really expected that to be toppled by another AI lab. Maybe we would have expected a large, heavily resourced firm like Google to take some market share.

But it ended up being Anthropic, which was a small, nimble firm, was able to develop products that were especially geared toward enterprise, and was particularly successful in activating those early adopters and evangelists of AI at a firm, to then spread AI to more firms and then within the firms they were already at. And then, finally, becoming the number one player in the market.

In your experience, are most companies running both now? Because I’m guessing it’s not a zero-sum game in most cases. It’s not a case of, okay, we were using Codex and now we’re switching exclusively to using Claude Code, or what have you. I’m guessing a good number of companies, or a good number of people, have those two icons side by side in their dock.

Ara Kharazian: Most companies that are using one at this point, yes — most companies that are using one are also probably using the other. But remember that, in the same way that every firm is on its own AI adoption curve, and that there are some firms that are more advanced than others, you see those same dynamics within a firm too. So the engineers may have been using Anthropic for a long time; that’s typically how we saw Anthropic proliferate through our dataset. When it started at companies, it was primarily being used by the engineering teams. But then over time, other teams start adopting these tools and technologies too.

So the marketing team, the sales team, the product team, shortly after engineering, is generally what we see in terms of the flow of adoption. And so other teams may start using multiple models. And that’s generally what drives the rising costs of AI across a firm: a firm buys the technology, but it’s extremely variable as a tool that is being used. Each person, as they adopt AI, is going to drive more spend on their own part. And then, as AI proliferates through a firm and more employees at the firm, particularly in non-technical roles, start to adopt AI, then you see even more increase in spend across the firm.

So it’s both a function of the intensity of an individual’s usage, and of the number of teams within a firm that are using it.

Yeah, and we see that firsthand. Our repeat monthly billings are with Anthropic, but we have token budgets and token spend across OpenRouter and also across OpenAI. So yeah, I understand that. And these things are used for different functions.

I think if I was one of those vendors and I was chasing lock-in and I wanted to try and run that traditional technology playbook, I would be leaning into connectors to the extended enterprise, and skills for specific disciplines and specific industries. And I think we see that in Anthropic, right? And also OpenAI. Anthropic comes to mind as having a specific offer for the financial industry, for instance, or having a specific offer for the legal industry, which encompasses skills, plug-ins and connectors to libraries like LexisNexis, or what have you, if you’re in law. That, to me, feels like the way to get people into your ecosystem and keep them there: to be able to say, well, we have connectors that the other place does not.

And remember, that means that much of the development of AI as an industry is going to come not from making the models themselves better, but from making them perform better in production for an employee, by giving those models access to datasets they don’t already have. Being able to connect to your enterprise workflows is ultimately what’s going to drive the productivity gains that we hope to see from AI.

Yeah, I think that’s right. And there are some of those bits there that you can rely on most people having — every company is either a Google Workspace or a Microsoft house, with some variation in the EU for companies that use Proton or what have you. But then it gets much more heterogeneous from that. You can’t assume that everybody is a Salesforce house; there are a lot of different CRMs out there.

You can’t assume that everybody is on Shopify, although it is the dominant e-commerce platform. So I think you’re right. I think that level of how well it works for your enterprise is definitely more important than frontier model capabilities.

And that is also constantly changing. I think that’s ultimately going to be great for the businesses that are adopting AI, in the same way that it’s great for consumers, in that this is an increasingly and extremely competitive industry, where — similarly to the month-over-month changes we see in market share — we also see the companies trying to sell AI employ competitive tactics to try to get the others’ customers. So it’s not just about releasing better models. We routinely see these kinds of price cuts that are enacted so that you can make sure you’re offering the most price-effective model for any given customer. And that is going to favour firms that are model agnostic. The businesses that benefit most from this are the ones that do not employ vendor lock-in, because they don’t need to.

And you don’t generally want to, given that there are so frequently going to be new model improvements and price cuts in this market that allow you to take advantage of the best model at the lowest cost. Now, lately, that’s meant employing what you just mentioned: OpenRouter. But the vast majority of firms on our platform, even the tech-forward firms, aren’t even using OpenRouter. So, for listeners: OpenRouter is an LLM inference platform. It’s essentially one of the platforms businesses can use to get access to models outside of OpenAI and Anthropic.

So you can use OpenAI and Anthropic, but you can also have access to some of the open source models, and some of the models offered through — if you’ve seen the buzz about China — Chinese models. They often are cheaper. But on our platform, less than 10% of companies are even using those. And so OpenAI and Anthropic are relatively less inclined to respond to those competitive pressures from open source models and Chinese competitors. But as that share goes up, you’re going to see their market share be threatened as well.

And then OpenAI and Anthropic will start enacting some more price cuts in response. Ultimately, that’s going to benefit the firms that are using multiple models.

That’s a good answer. I will say, and I’m sure a number of listeners feel this, that the usage and pricing page of the Claude desktop app is the most confusing thing I think I’ve ever seen in my time in business. Sometimes I look at it and I think, how did we get here? How did we get to the stage where I have five different sets of limits, and different sets of usage credits, and incentives that are being pushed in my face for 50% additional usage by a set period? There’s some arcane stuff going on with model pricing across those big companies.

But as you said, that’s because they’re fighting fiercely for share of spend.

How much visibility do you have into the composition of AI spend? We just talked about, I think, the total amount of it. So if I think about fashion, and if you ran a pure token in/token out analysis — and again, for listeners, tokens are roughly representative, if we think about text, of words; it doesn’t quite map out like that, something like 0.75 of a word represents a token or something.

But if you ran a pure token in/out analysis on fashion, most of the usage in our industry would be visual. It would be image and video generation. In part, that’s just because we’re a visually geared industry. I interviewed the CEO of Photoroom a couple of months ago and he put it this way: fashion sells images, and then it has to fulfil the promises of those images with product. But I think it’s also a matter of AI maturity, or at least the perception of AI maturity.

We ran a survey in the springtime, the results of which will be out by the time this episode is out. It’s a much narrower data pool than the one you’re dealing with. But in it, people tell us that they believe image generation — both for in-house use at the beginning, so you come up with a design, you want to present it to colleagues, you want to put mood boards and things together, and then downstream in campaign photography and product detail page e-commerce imagery — people perceive that to be way more proven and reliable than, say, retrieval-augmented generation across an extended tech estate that includes ERP and CRM and PLM and all these other things.

So how much visibility do you get into the kinds of AI that companies are spending on? Can you see what the capabilities are? Or do you just see it at the pure token spend and subscription spend level?

No, we do. And that’s been the focus of our research going forward. When we started this work, we were focused on adoption rates, because AI was so early in its implementation at most firms that the only question we really could effectively answer was: does this firm use AI at all? Now, as the sector has developed, we can start measuring the intensity of that adoption and use that to separate firms in terms of their place on the adoption curve. So the most costly share of AI spend is API usage. That’s generally going to correspond to usage of coding agents, but it’s also the use of AI to label large corpuses of information. So you can imagine the way Ramp’s product uses it: you take a receipt, you upload it to the platform for your corporate card reimbursement, and Ramp’s AI will go through that receipt, label it, and file that expense for you. That’s the kind of product experience that we use day to day in many AI-native products. You don’t even necessarily think of it as using AI as a consumer. That’s going to be one of the biggest categories, followed by coding agents.

And then some of the other really advanced forms of AI adoption, particularly in consumer companies, are AI as it is used for customer service operations and sales. That can take a lot of different forms — everything from a customer service chatbot, which at this point has become quite advanced, and many of us have probably spoken to an AI and didn’t even know it. There are voice agents that are quite advanced now too, that companies might use to try to close your inquiries before they send you to a human agent.

Image and video generation is increasing in size, but it is a very small share of actual AI spend. And I think you can imagine why. It’s not particularly scalable in the same way that using AI for a customer service operation is, where you’re deploying it across the entire organisation, handling thousands of calls a day. Whereas image and video generation, particularly for advertising, tends to be much more ad hoc and team based.

Yeah, no, that seems right. One thing I just want to pin down about the data that you have available to you: it’s US only, correct?

We’re a UK publication with heavy EU readership, although the US is our biggest market, so I just want to understand the geographical limits of it. But I also want to quiz you a little bit on the enterprise nature of it, and the fast-growing-company nature of it, because fast-growing company could mean different things to different people. Just run me through how you delimit the scope of the information that you have — geographically, and in terms of company type.

It’s a question that we get often: is our dataset representative of the typical firm in the United States? And it’s not. The businesses that tend to use our platform tend to be quite tech forward, no matter how you cut it. Not to say that they’re always in the tech sector — actually, most businesses that use Ramp are not in the tech sector. But Ramp is itself an AI finance platform, such that the manufacturing and construction and consumer firms that use our platform also tend to be quite tech forward and likely to adopt new technologies.

Now, that is actually somewhat of an advantage for the research that we are trying to produce. AI is a technology that is going to be invested in by tech-forward firms, so it benefits us to have access to this dataset that looks at what they’re spending money on specifically. In our work, we will control our analyses, when it makes sense, to show our results relative to the typical firm. We recently wrote a paper examining AI’s impact on the job market. We used data from Ramp to group out firms that use AI versus firms that don’t.

And then, within those firms that use AI, which ones are using AI particularly intensely versus in a very light way — maybe like ChatGPT subscriptions for everybody. And then we got workforce data from a company called Revelio Labs, and we joined these two datasets. So we could see: these firms use AI, these firms didn’t. What happened to jobs over time? And it’s there that we got these very compelling results that showed that, despite a lot of the talk going around about tech layoffs and people being fired because of AI, firms that use AI heavily are actually growing faster.

They hire 10% more over the two years following AI adoption. But it’s concentrated. It’s essentially only happening amongst the firms that are using AI intensely. So what that means is that most firms that are using AI today aren’t really seeing all that many benefits at all. Because for most firms using AI, they’re not using image and video generation, or AI for sales and customer service.

They’re not using the coding agents. The vast majority of firms that are using AI are still on fairly simple ChatGPT or Claude subscriptions that are given to all the workers of the company, but are not necessarily connected to all the tools. If you’ve used those very basic forms of the technology, you can see it’s kind of nice to have. But it’s very hard to see how that might be so productivity-enhancing that it’s going to vault the company into the next generation of productivity growth. And so a lot of what we study is: what’s it going to take to get more firms not just on an AI adoption curve, but on a high-intensity adoption curve?

Now, as for the rest of our dataset, what I’m actually very excited for is that Ramp has now launched internationally — this summer it launched at least in the UK and the EU.

Yes, I saw you’re available in the UK. We don’t use Ramp, but we are paid by partners who do.

In a couple of years, I’m hoping we’ll get a long enough period of time in our dataset that we can start reporting results that compare the performance of firms in the US versus firms in the EU, particularly in terms of AI adoption and its impact on labour markets.

Perfect. And we see the same thing with our survey, to be fair. We’ve always said there’s a degree of preselection that happens whenever we talk to the market, or have the market talk to us, which is that people who read The Interline — a technology publication for fashion — are already people who are fairly deeply into this space. So we recognise that. But we also see it the same way as you: you’re better off looking at the high-intensity folks and the more committed companies than you are at the people who are just coasting.

Just to go back to the vendor lock-in side of things very briefly. Do you see consolidation coming in the market? Because it’s a strange thing now. You basically have two major suppliers, I would say, for most companies. Google is a weird one because, technically speaking, we do pay Google for AI, as I’m sure a lot of companies do, but it’s baked into Workspace and we would not be paying it separately. You can’t not. So it’s in there for that reason. But on a selective basis, you basically have OpenAI and Anthropic.

You hinted before at the idea that — you think about Chinese models, you think about other companies — there’s maybe an expansion of the market coming. And then, is there a contraction? Do you think there’s a long-term market where there’s a wider cohort of AI companies that can actually survive out there? Or do you think this is destined to be a two-horse race for the foreseeable future?

I don’t think it’s going to be a two-horse race. I think there are going to be many more competitors. And that’s because I think the firms that use AI most effectively — the strategy in terms of buying AI that’s going to win out — is going to be the model-agnostic strategy. It behooves you as a firm to experiment with many models to figure out which one works for what task, but it also behooves you because you’re going to want to use the most performant model for a certain task at the best price. And there are already some technologies entering the market that allow you to do that. OpenRouter is actually one of them.

Right? They allow you to automatically and intelligently direct tasks to the model that makes sense. This is a really hard market to keep track of, too. If you’re using OpenAI and Anthropic yourself — you’re just using Codex or Claude Code — let’s remember that they don’t make it particularly easy for you to get the highest ROI from your model usage, and they have no incentive to. They make the most money when you use the most tokens.

And so if you’re using OpenAI or Anthropic for your projects, you kind of have to be extremely informed on AI models as they develop, and on the market, so that you can make sure that — oh, do I need to pay for the most expensive model for this task, or can I send it to the cheaper one? And that’s just not going to be natural for most people to do. I don’t think most people are capable of doing that in such a fast-moving market. And I don’t think OpenAI and Anthropic today have the incentive to create a product themselves that will intelligently route you to the most appropriate model at the most appropriate time.

So I think there will be increasing numbers of companies that develop that kind of technology, that businesses will buy, that route you and your tasks to the most appropriate model without you having to do anything. And that’s going to support a future that is much more model agnostic, for which there are many more models besides OpenAI and Anthropic — and that may include open source models too. And, by the way, other American model companies. Google.

And they do as well. There are a lot of companies out there that have their own frontier LLMs. But then, when you widen the net, as you said, there are a lot of companies out there that have smaller specialised models for vectorisation and embedding and things like that, and text to speech, and what have you.

And then think about the international complications of this. If you’re a company based outside the US — we just saw, with the Fable ban, the Anthropic model ban, that the United States has the ability to impose export controls that within minutes cut off your company’s access to a certain AI model, simply because you may not be a company based in the United States, even if you are otherwise based in a US ally country. Now, if you are a European company that is building using AI, you might be very hesitant to adopt OpenAI and Anthropic and build your business on their tools, because you might have the technology pulled out from under you without any notice. So what do you do if you’re one of those companies?

You probably are going to explore providers outside OpenAI and Anthropic, if not outside the US entirely, and distribute your efforts across many firms so that you are not tied to one.

So let’s think about that practically speaking. The contemporary example of that would be the OpenAI unreleased model — I was going to use the word hack; it’s the wrong word — the OpenAI unreleased model finding its way into production databases at Hugging Face. And Hugging Face not being able to use US-developed closed-weights models because of guardrails, and having to turn to open-weights and Chinese models in order to get around that.

It feels like we’re barrelling towards a bit of a hyper-balkanised future of AI, from what you’re describing there and what they’re looking at. What should just a rational actor, just a regular company in this space, practically be doing to do that kind of risk spreading and diversification? Because we’ve said there are alternative platforms like OpenRouter and so on. From a pure what-do-they-install-on-their-employees’-computers point of view, what is it that actually gives them a way to spread that risk?

Well, it’s hard to say today, because there aren’t great options, at least on the frontier, that don’t rely on OpenAI and Anthropic. So I think today, I wouldn’t tell a firm not to use OpenAI and Anthropic. You should be using OpenAI and Anthropic. You should be using the frontier models. You should have your employees experiment as much as they can with the best models that are available.

And also, if you are in charge of procurement, particularly for a company outside the United States, you should be looking to implement the kinds of platforms that would set you up to be model agnostic fairly quickly. And so that’s going to be platforms like OpenRouter today.

And if we think about the Chinese model side of things, Moonshot’s Kimi K3 is the current Chinese model du jour, as it were, and it’s the one that seems to be getting a lot of attention. Now, I feel like we’ve done this dance before. If everybody thinks back to the initial DeepSeek R1 release, all of a sudden there was a big drop out of the bottom of tech stocks and things, and everybody was saying, well, okay, this fully democratises the capabilities. It’s the end of the business model for OpenAI and Anthropic.

That did not pan out with that initial DeepSeek release. Do you think this turn is different now? What is it that you think American companies and international companies can get from models developed overseas that makes them more attractive? And do you think it’s actually going to move the needle this time?

Well, it’s similar in that, like last year when DeepSeek got its first viral hype moment, it exposed the main feature of OpenAI and Anthropic’s dominance in the United States that isn’t working for businesses: that they are themselves not incentivised to build highly performant models that are also cost effective. And they have to date not built a product infrastructure that allows their customers to use their technologies more cost effectively. So the last time it happened, last year, when DeepSeek had this big hype moment, even though the vast majority of firms weren’t switching off OpenAI and Anthropic, it did spook them. And they responded by putting out some cheaper, highly performant models. Those ended up being significantly better than what was otherwise out there from DeepSeek, and so we didn’t hear about DeepSeek for a while.

Now it’s happened again, where it’s not just DeepSeek — there are many more Chinese models that are highly performant. And I want to say that it’s as easy as the American model companies enacting some new price cuts and releasing some cheaper models, but it’s going to require much more than cheaper models. It is going to require a rethink from OpenAI and Anthropic — assuming that they are affected enough competitively to respond in this way — to start offering products that allow firms to better measure the ROI and monitor their spend. Today, that is not offered. The fact that there’s no automatic model switching offered in Codex or Claude Code, in the sort of way that companies are asking for —

It just does not exist in any sophisticated way. And so there are changes around the product that businesses are clearly asking for, but that OpenAI and Anthropic are not able to meet. Now, I think the mistake would be thinking that the threat is only going to come from Chinese model companies. It’s not. It’s also going to come from Google.

It’s also going to come from other AI companies that are starting up in the US that are trying to offer alternatives to the two top model companies.

Now, one of the statistics from the AI Index that we’ve used a couple of times, and that I personally find the most interesting, is token spend per employee. Forgive me, I haven’t looked at whether it’s been updated in the last week or so, but the last time I did, I think the upper bound for what companies would spend on an individual employee on a token basis was something like $7,500 per employee per month. That’s the absolute high end. The top 10%, I think, was a bit more reasonable, at around $600 a person.

But the median firm — your average company, which we’ve been talking about — spends about $11 per employee on the token side of things. I know regional pricing is different, but that’s about, or slightly less than, a single chatbot subscription. It’s cheapest here in the UK, so that’s two cups of coffee. It’s like a 600-fold difference. How do we look at this?

Do we think about it in the friendly way, which is that the company spending $7,000 a month per employee can’t possibly be getting that much value and is just lighting cash on fire? Or do we look at it the other way and say that the companies only spending $11 per employee per month, just giving somebody a cheap seat to ChatGPT or whatever, are missing the boat, and the market is going to outpace them? Because I look at this from our experience: it’s very easy to burn $200 on automating some admin and then look back and say, was that necessary? Did I need to do that?

But then you also look at the world’s biggest retailers, Europe’s biggest retailers — Zalando — going all in on image generation as the future of content creation. That is not cheap. It’s potentially cheaper than traditional photography, but it’s not cheap. So I’m left a bit conflicted over whether the biggest companies are just trying things here, or whether they know something that the rest of the market does not: that you need to become that high-intensity, high-frequency company to really start to get the value and get ahead.

I think about it from my perspective as a researcher. When we were considering this metric to track spend per employee, we kind of had to break it out when we looked at the distribution of spend. Because although most of our metrics are reported at the median, you start to see how that falls apart for AI, because there are such large differences between firms in the top 1%, or even the top 10%, versus the median firm. So our goal with this kind of research is not even to suggest, hey, here’s how much you should be spending on AI, but at least to give some benchmark for the wide range of spend a firm can have. Look, it’s all going to come down to the actual ROI for a firm. We went through a couple of different phases in the tech sector as far as what good AI adoption looks like.

I think, actually, one of the unfortunate phases that we’re still in is that firms still do not know how to talk about the value of AI. Not that they don’t know how to talk about the value — they’re really not sure how to measure it within the firm itself. That’s why we end up seeing all of these different metrics thrown around: firms measuring themselves by how much they spend on AI, or measuring it by the percentage of workers at their firm that use AI. That’s not the worst one.

I’ve seen, similarly, the share of code at our company that is written by AI. And none of these metrics give you any sense of the quality of work that is being done, or whether the work is being done in a cost-effective manner, or whether any of it was worth it. So I will say the kinds of metrics that I’ve heard that are quite impressive as far as measuring AI’s impact — and I think are generally the way most firms should go — are the ones that are tied to outcomes that are felt by customers. I talked to a tech CFO, involved in the consumer space, who set an actual goal, a minimum threshold, for the share of customer service inquiries that need to be addressed by AI completely, with no human touch, by the end of the year. I thought that was great.

You’re not saying, oh, we need this many customer service agents using AI. You’re saying, we need to have a certain level of customer service operations that can just be done by AI completely, because that’s going to allow our human agents to then focus on more inquiries. And then you can think about the other metrics that you would define. Maybe even for your human agents, they can use AI to complete customer service calls faster because they can look up information and answers. So you can define a benchmark based on that too — the average amount of time it takes for a human agent to close a call, and see if that comes down over time as AI adoption proliferates through a firm.

These are much better metrics that are actually tied to the impact AI can have on a firm and can have on your customers, and that I think resonate more with the market and resonate more with customers especially, rather than us hearing all the time about the promise AI can have because, oh, more code is written, or we have more data centre jobs available.

They are harder metrics, but they’re also metrics that bring me to my very near final question. You mentioned earlier that the kinds of companies you capture in your data are fast growing, and that they are also growing in terms of headcount despite being high-intensity adopters of AI. If I think about call centres — the sense of what you’ve just been talking about — if that company succeeds in that metric by the end of the year and has a fixed amount, maybe a majority, maybe not, of its customer service interactions handled fully autonomously without a human touch, that has to lead to net job loss in that segment at the very least, because presumably those queries are handled by people today and they will not be handled by people in the future.

And the thing for me that I keep coming back to is that fashion, by and large, is not made up of fast-growing, VC-backed, well-capitalised companies. It’s made up of companies that trade on pretty thin margins, companies that are in the SME market in particular, a bad collection or two away from going out of business.

These are the kinds of companies that, if they scale their AI spending in the way that we’re talking about here, and they become these high-intensity folks and they set these concrete metrics, that spend comes from somewhere else. There’s no magic money tree to fund these things. Do you think there’s maybe a different world that you’re looking at with these fast-growth, well-capitalised companies versus your typical brand or retailer when it comes to the outlook of the relationship between AI and jobs?

AI is going to affect my job. It’s going to affect other sectors too — mine, yours, and sectors that we don’t even know to consider yet as far as AI’s impact on them. I definitely fall on the side that people are overestimating AI’s impact on the job market as far as job loss. First, because we haven’t actually seen any broad job loss from AI at all. And so these predictions are not really based on anything quantitative or observable.

And the second is that I really am not compelled by any of these suggestions that AI is going to develop differently from any other technology that has also automated work in the past. Now, AI will definitely lead to some industries declining in jobs. We may not know to what extent, and we may not know what those industries are. But my instinct is that it’s probably going to happen in industries that we aren’t thinking about and don’t even expect. There tends to be, for example, a lot of focus on white collar work and how AI is going to automate the work of software engineers and lawyers and accountants, and therefore we’ll have fewer lawyers and accountants and software engineers.

Now, we have the result from our dataset that shows there’s actually a lot of job growth in white collar work, especially at firms that adopt AI intensely. But then you can think of the similar impacts that other technological innovations have had in decades past on those sectors. Computers in general made the work of a lawyer much easier and much faster. Go to a law firm and they historically had, inside the actual office, libraries that lawyers had to go to physically, to open books, to find case law, to refer to them in their legal briefs.

And now we have computers and web searches that allow you to build those briefs, theoretically, much faster. And you can think about the hours and hours of time that are saved per day by not having to look up case law in a book. There’s no suggestion out there that lawyers, suddenly, because of the advent of computers and LexisNexis, were able to start going home at 4pm every day, right? Because their competitors also had access to that same technology.

Very true.

And therefore they also started building their reports much faster, and therefore found ways to make their legal briefs even stronger or better, and to do other human-oriented work that technology could not do. And that’s why they all still go home at, like, midnight.

Which is also why I’m not a lawyer, to be clear. And I think the other thing as well: putting a computer in the legal office, okay, fine, so you replaced the job of the archivist or the librarian, and you created an army of paralegals and other roles that had the computer as a given — you needed that technology revolution to create those other jobs. And I think if I frame it through a fashion lens, a lot of focus is on the creative work. So, particularly, campaign photography and things like that, where up until now you needed a photographer, you needed a stylist, you needed a make-up artist, you needed somebody on lighting and so on to put together a campaign shoot. Now, you don’t need any of those people.

You can do it all with AI. But those jobs initially were also created by the advent of photography. They were particularly created by the advent of digital photography, which made all of this scalable and so on. So I see where you’re coming from. I think it does impact jobs, but I buy the idea that a lot of those jobs were created by technology in the first place.

And I think it’s maybe a natural evolution from that point of view. Let’s end, because we are over time, on prediction. If I asked you to guess what you would see in the AI Index if we did this interview a year from now, what would your best guess be? Do you think AI, at least as a paid product, is going to be more or less universal? Do you think the big labs are still going to dominate?

Do you think that medium brand we’ve talked about is going to pick up and become more high intensity? Do you expect to see spend per employee go up? Just give me your potted overview of what you see happening.

I think we’re definitely going to see AI adoption rise to virtually 100% of firms within the next five to ten years — probably closer to the five side. The average firm will become more intense in its AI adoption. I also think, unfortunately, that firms that are less inclined to adopt AI at all will probably suffer from market competition that is able to grow faster with AI. And those firms may end up going out of business.

And that’s going to happen — not to be a doomsday prediction, but that’s generally what technological innovation does, even in sectors where you wouldn’t expect there to be massive productivity gains for that firm. As an example, we can think about doctors’ offices again, or about advertising here and there. The firms that never adopted even the basic technology of a computer or a Google Maps listing are probably the firms that went out of business, even though the technology itself was not that beneficial to the growth of the industry itself.

So we’re going to see something very similar with AI: that AI is enabling some firms to grow faster even if it is not enabling them to produce more themselves. It just ends up being this growth engine in ways that we do not expect.

What ways do you expect the high-intensity firms to change in that five-year period? What does a high-intensity firm look like — and I know this is difficult to say across business models and things — but if you stay at that frontier as a company and you commit to this, what do you see being the benefits that you look back on in five years’ time and say, that was advantageous for us and that has changed the type of business that we are?

I think a lot of the growth is going to come from adoption outside of the main models. If there’s one headline point I would point to in our research over the last two years, it’s that you don’t really see that many productivity gains from AI as we know it, as most consumers at least understand it — and that the free or relatively cheap subscriptions to ChatGPT and Claude are not particularly productivity enhancing. Some of the most exciting areas of development for AI are happening outside of the hype cycles. I use customer service operations as an example often, because it may be a little bit dry as far as ways in which AI can improve our lives and improve a company’s operations, but it is also one of the areas where a customer will actually feel the change. Helping customer service operations at a company move faster, helping customer service agents close out more inquiries and help more customers, and ultimately leading to a better experience for more customers.

I like to focus on versions of AI where you can clearly see the ROI, both for the company and for the customers that it serves. I look for uses of the technology that are able to prove their value there.

Good. Well, I think everybody is in the same boat. That’s probably the focus of our AI Report, which should be up around the time people listen to this as well.

Ara Kharazian, I’ve taken up a little bit more of your time than I intended, but this has been a great conversation. I appreciate you joining.

Thank you so much for having me.

that’s the end of my conversation with Ara. I got a lot out of this one, and I hope that came across. Now, I do tend to agree with his projection there that we’ll see total AI adoption within five years or so. You can infer that from basic chart reading, in the sense that if we got to 55% uptake in three years, another five should really capture the rest. What I still think isn’t clear is what the adoption is going to look like, especially in fashion.

Our AI survey this year focused very much on AI maturity by sector and on how people feel about the technology. I’m starting to think we should run a separate vendor-focused data collection exercise as well, so we can understand what capabilities people are spending on and which providers they’re using for it. Because frontier LLM spending is obviously very concentrated on ChatGPT and Claude, but I think the picture would be far more blended if we did a sector-specific slice of the profession. I also want to emphasise that just because AI has the potential to create jobs that didn’t exist pre-AI, like Ara and I just talked about, there’s no guarantee that those are net new jobs, or that they’re going to be applicable or available to the people who work in traditional fashion disciplines today. It’s not always the case that you can upskill to be ready to fill a role that replaces your current role.

What is obvious, though, is that AI adoption across every sector, and in fashion specifically, has a lot more twists and turns to make over the next few years, and I’d really like to have Ara back on at some point in the future so we can compare benchmarks again across the broad suite of data and across our fashion-specific one.

For now, thanks for listening. If you’re a fan of these weekly interview shows and you haven’t checked out our other show yet, The Edit, go do that — because that’s where Grace and I spend twenty-five minutes or less each week talking about the news headlines. It’s a breezier, more accessible, more fun format designed to fit into your commute, but without sacrificing the expertise that The Interline is built on.

Whether you end up listening to both shows or not, I’ll talk to you again really soon.

Further Reading:

Ben is a senior fashion technology storyteller, analyst, and moderator with more than a decade's experience in retail.