Hey, welcome back to The Interline Podcast.
Right now, fashion has a laser focus on product outcomes, and for good reason. The headline research tells us that industry growth is slowing down, which means brands are competing with one another for relative rather than absolute advantage. If you listened to our recent Mailroom episode, you heard Hannah Crump from BoF Insights talk about the fact that, in a low-growth environment like the one we’re living in, you can grow your slice of the pie by subtracting it from somebody else’s, but the overall dish isn’t getting any bigger. That makes optimising per-product profitability and reducing product- and collection-level risk essential.
The data also shows that consumers are way less aligned with brands than they have been in a while, and tend to make selections based on specific products instead. That makes subjective attributes like style and fit, again on an individual product or small-collection basis, matter a lot. It makes more calculated, objective criteria like pricing, value, market positioning and market counter-positioning matter arguably even more when fashion companies make the choices that bring products to market.
None of these outcomes just happens. Each is the result of a set of nested decisions that cascade down from ‘This is what we want to make, given the market conditions we’ve been handed’ to the granular choices that influence how that thing gets made, how effectively it meets its objectives and how it delivers its outcomes.
Each of those decisions is also, especially right now, a target for potential automation. That’s particularly true if you believe AI can make these mission-critical choices better, or can at least make them based on a more complete reading of circumstances than a human being can.
Because of that, and depending on who you’re talking to, there’s a very blurry line between which of the many choices that turn a slot plan, a brief or an idea into a finished product should remain the preserve of people, and which could be, should be or might be augmented or fully automated by AI.
Sitting back and listening to this, your instinctive reaction is probably: you let a human set the creative direction and determine the success criteria, because that’s what really makes a brand. Then you let a series of AI agents make the optimum choices in the subtasks underneath all of that.
But is that actually the best approach? Is that the right read? Does that framing undermine the creative decision-making that goes into those more granular stages we’re in a rush to automate? Does it over-index on AI output and capabilities that professionals tell us they don’t really trust? Does it overestimate the complexity and the essentially human nature of setting a collection direction or building an assortment plan?
Your instinctive reaction is probably also coloured by the systems you use and the processes you operate in. AI is promising a lot in terms of automation, but the inertia it’s working against is massively strong too. It’s not just that companies are saying, ‘This is the way we’ve always done things, so why change?’ It’s a different kind of weight, because it comes from a much more uncertain place.
To get a real answer, you have to ask yourself some extremely profound questions, in the most literal sense, about not just why you do things the way you do, but whether you need to keep doing them at all, or whether we’re in a completely different era.
A lot of this discussion has already come up in our AI Report 2026, which I’m sure almost everyone listening to this show has read. If you haven’t, I’d suggest grabbing a copy, because there’s a lot of timely content, context and data in there.
As part of that report, we interviewed a bunch of tech executives. One of those interviews has kept coming back to the front of my mind whenever I’ve been asked to talk about AI automation, how far it might go and what decisions might get made for us. It’s the one we did with Rupert Schiessl, Chief Strategy & AI Officer at Bamboo Rose.
I’ve had a few short chats with Rupert, on and off the record, over the last couple of years. He’s got a refreshingly objective viewpoint on how far AI is likely to change fashion. As his title suggests, he’s also of the mindset that AI and strategy are basically inseparable today: every strategic plan is an AI plan in and of itself.
Bamboo Rose also uses the term ‘decision intelligence’ to refer to the way AI is applied across the extended product lifecycle in the technology ecosystem, making all those choices that turn a product concept into a finished thing. When it comes to how far AI should influence decision-making, and what that means for product outcomes, I think Rupert is as well qualified as anyone to have a pretty far-reaching and impactful conversation.
So let’s get into it.
NB. The transcript below has been lightly edited.
Rupert Schiessl, welcome to The Interline Podcast.
Thanks for having me.
Not at all. Looking forward to this one. You and I have spoken a couple of times in different formats, but this is the first sit-down interview we’ve done. I’ve been looking forward to it because I know you have some good perspectives on the role of AI in business strategy and across fashion.
I try to.
Good. I’m looking forward to tapping into them. We start these shows every time with a look at the guest’s day-to-day, and then we ask them to define something. Let’s start with your day-to-day.
Your title, Chief Strategy & AI Officer, transcends a couple of areas that are probably near or at the top of our listeners’ agendas: strategy and AI. In fact, I’m going to hazard a guess that, for most of our audience, strategy is primarily AI or AI-centric at the moment.
Tell me how those two things fit together for you, what it means to do strategy and to do AI, and how much of that you think you have in common with people listening who are juggling the same priorities.
Five years ago, a strategy job was about markets, positioning and M&A. Today, almost every strategy question comes down to what AI makes possible and what it does not.
I think SaaS software vendors like Bamboo Rose, where I’m working, are having their Kodak moment. If you remember, in the nineties Kodak got into heavy trouble when digital photography arrived on the market. They saw digital coming. They built the first photo sensor and so on. But what they couldn’t do was kill the business that was paying their bills.
For a software vendor today, it’s very similar. AI is not one project you run inside the strategy, where you can continue as before. We have to do it properly and rebuild our whole business around AI. This is strategy. Doing anything else would just be adding features to a model that’s quietly running out, I guess.
A lot of companies are making the mistake of automating their existing processes. Then they say, ‘We’ve adopted AI. We’re working so much faster now.’ That’s the deadly trap I think Kodak went into thirty years ago. It’s not about accelerating existing things. It’s about completely breaking those processes and redesigning them entirely.
That’s true for us too at Bamboo Rose. My job only makes sense if it’s about changing what the company is, not adding AI to how we already do it or what we already sell.
I think that’s good framing, and it’s going to come out as we go through the rest of this conversation. For the thing I want you to define, I’ve picked something that sounds absolutely basic: what constitutes a decision?
Everyone’s got an instinctive understanding of what that means. We’re used to choosing one thing over another, or setting an objective and working towards it. But decisions aren’t one big unit, particularly in fashion: product decisions, market decisions, category decisions and so on. You can break those down into pretty granular pieces, and each piece may have a different person or team responsible for it.
If you think about a hypothetical style, you can decide you’re going to make something. But between that decision and the eventual outcome, a product on a shelf, there’s a massive spectrum of interrelated decisions, from choosing materials and manufacturing partners to setting margin targets and approving colours.
When you think about the sheer complexity of what a decision can be, and all the sub-decisions underneath it, how does that frame the way you approach decision intelligence, which I think is the framing you use for AI? Do we automate the big decision and give people the smaller decisions underneath it? Do we do it the other way around? Is the answer the same for everyone? Walk me through the impact of AI on the decision-making process.
In the industry we’re working for, fashion and apparel, a decision is almost never one single act. Between ‘We’re making this style’ and a product on the shelf, there are hundreds and hundreds of small decisions. Most of them aren’t really decided. They’re sort of inherited, because many are copied from last season, from a template or from whatever the supplier proposed. They’re already in our customers’ history. The retailer knows them already.
Most of what looks like deciding in our industry is actually looking things up in past data and making sense of it. People are rebuilding context: what did we do last time? What does the supplier normally charge? Is this certificate still valid? I think that work is 90% of the effort for our users.
For me, the line isn’t really big decisions versus small ones. We think we have to automate data retrieval as much as we can in the existing history. Automate putting the context together so AI can understand it and work with it, and automate every decision that is really just basic rules.
Concretely, that means a lot of the small decisions shouldn’t reach a person at all. If the rule is already known and the action can be undone, the system should just do it and log it. Asking a human to validate something that was never really a decision is, I think, a waste of time.
I guess we think in two levels. We have decision automation, doing things automatically when it’s possible, and decision support, where AI prepares a decision to be validated by a human. The mistake many companies are making today is putting too much effort into decision support because it feels safe when you bring a human in. But you end up with an expensive AI assistant plus the time spent by the human, and you don’t have any savings at all.
I think that’s an interesting perspective. What you’ve described is something I don’t think I’d quite crystallised in my head until just this minute: a lot of what we think of as choices aren’t really choices when you weigh up all the variables that go into them.
If we go back to our hypothetical style and think about choosing a material or a manufacturing partner, there may be one or two partners with the specialisms and capabilities to make that particular product. They can do it in the timeline, at the cost, and they have the right sustainability certifications and credentials.
A lot of that is sifting through information and making what’s basically a binary choice, if it’s a choice at all. I think that’s what you’re driving at: you’re trying to give people support to make choices that only have one or two outcomes, and those are places where automation is the more natural fit.
Exactly. Our main work is building this meaning and context so people can make decisions, but also so AI agents can make decisions based on that context.
How far you can go depends on two things: how governed your data is, and whether the decision can still be undone. A company with clean supplier data can go much faster than a company whose data still sits in spreadsheets. Instead of asking yourself what you should automate, in most cases you have to ask first: what do we actually know, and how is it stored?
You just talked about the difference between decision support, which feels safe, and decision automation, which feels risky. If I look at the survey findings in the AI Report 2026, published a month or so before this conversation and containing a short interview with you, one of the most revealing things was that, although nine in ten fashion professionals use AI every day at home and at work, only about a quarter currently trust its output enough to base an important decision on it.
In that interview, you talk about trust coming from transparency. People don’t want to be served recommendations, or see prescriptive action being taken, if they can’t understand why. But I think most AI fits into that second box: it’s inscrutable, it’s probabilistic, and it gives you authoritative outputs without necessarily giving you visibility into its working.
What do you think happens next? There are a couple of ways you can build trust in AI’s output. One is to do it extremely slowly: you slow-walk it into processes and say, ‘Look, it hasn’t made any mistakes. It’s retrieved the information accurately.’ You do that progressively over time. The other is to build it quickly: ‘This is the correct answer, I can show you it’s correct, and we can now instil confidence in this.’ How do you think we build the trust that’s currently lacking?
I think building trust with AI is very basic. AI doesn’t take responsibility for anything. If you sign off a compliance file, for example, you own it, and you’ll get sued if it’s wrong. You can’t own something you can’t explain.
On explainability, I think it’s both showing a quality output and being able to explain how it was done. Both are useful, because being transparent gets people to try it, but a good track record gets them to keep using it. You need both before anyone lets it act for them.
It’s like when you go to your doctor. If they prescribe medication and get you cured, you’re happy because you’re cured, but you’d have preferred to understand why you were sick. If you get a very good explanation of what you have but don’t get cured, you won’t go back to the same doctor either.
We have to be a bit more careful about what transparency means inside a company. It doesn’t just mean explaining the model. As you said, the models we’re all using today, these LLMs, these large language models, are probabilistic. They love producing accuracy and confidence scores, and no human wants a probability. Probabilities are difficult for us humans to read.
Meaning comes from showing where the answer came from: here’s the answer, here are the records behind it, click and check them. Providing a chain you can check is much more convincing than a confidence score. A score asks you to trust it even more, because you don’t know how it was generated, whereas a source requires less trust.
I think that’s right. One of the findings from the report I wanted to talk to you about was how the industry seems to perceive maturity. I think maturity begets trust here as well.
I know you see decision intelligence as something that sits across the entire product journey. In a way, that contrasts with what the industry is telling us: it perceives AI to be very mature and capable at the beginning of the lifecycle, in concept development, market, trend and competitive analysis, and other analytical work. Then at the end, in marketing and content creation: video and image generation.
People seem to perceive it as least mature, least ready and least trustworthy the closer it gets to the real product: technical design, sourcing, production and so on. I think some of that is down to seeing AI being good at the extremes in other industries. In our personal lives, it’s easy to say, ‘AI is really good at taking in a bunch of unstructured, non-normalised data and giving me insights,’ or, ‘It’s really good at generating images that I think look nice, or close to photography.’
Do you see that same weighting of expectations around AI capabilities when you engage with fashion brands? Or, if not, give me an example of where you think AI is creating real value in that middle piece, between the early and end stages of the product journey.
Honestly, I don’t see it exactly the same way. I understand why people answer like that. I think what these surveys measure is visibility, not necessarily capability.
In design and marketing, you can very easily see the output: a generated image, a product description, a campaign. If it’s wrong, it’s very cheap. You just run it again. In sourcing, costing, compliance, this whole value chain we’re managing within Bamboo Rose, if it’s wrong, it costs real money. People are less ready to take risks there.
There’s also the fact that a designer or marketer gives the AI material it can already read: a brief, an image, a piece of text. It’s coming from outside. A sourcing manager or compliance officer has to give it the system’s history.
Until now, the data we have was organised for a screen to display it in software, not for a machine to read and understand it. A column name in a database often doesn’t mean anything on its own, so an AI agent cannot use it.
The model is not weaker when you get to sourcing or costing, and AI is not less relevant. It’s just not connected to the data in the right way to use it. Nobody has translated the business information into something a machine can actually read, simply because that translation is a lot of work. I know what I’m speaking about, because we’ve done that work with Bamboo Rose.
The next question is a two-part one, and I’ll split it off, because you’ve hinted at it there. We’ve talked about trust a few times. We’ve talked about model capabilities being essentially universal, with the other elements around them determining success.
The AI Report survey shows what I’ve ended up calling a non-virtuous loop. Fashion professionals don’t think AI will give them reliable answers, so they don’t connect it to their technology ecosystems. If you don’t trust something, why would you integrate it with your ERP or PLM? But your ERP, PLM and supply chain management solution are the sources of truth that would allow an AI application to deliver more reliable answers.
How do we get out of that spiral where people say, ‘I don’t trust this, so I’m not going to make it part of my technology landscape. But if I did, I would trust it more’?
I think that’s absolutely right, and that’s what we’re betting on too. We’ve worked for over two years to expose Bamboo Rose through MCP, so customers’ own agents can also query our database. We were quite early on that because we don’t believe the interfaces of the future are the software interfaces people use today. Some vendors will probably fight that, but I think it’s a losing position.
I’d go even further than you on the uncertainty. We don’t know whether ChatGPT or Claude will win. We don’t even know whether MCP will still be the standard in two years, because new protocols for agents talking to each other are appearing. Maybe the big platforms will build their own. We’ll see. Anyone who tells you today how this will end up is 100% guessing.
What we’ve done is build things that can be swapped, that are model-agnostic and protocol-agnostic. We assume whatever sits at the other end, outside what we’re doing, will change more than once in the upcoming months. We’re already seeing that. It’s changing all the time.
We’re also saying that the interface itself is becoming a commodity. Customers already want to generate their own UI on top of their own data in an afternoon or so. Some have started doing this with smaller datasets. If the screen is no longer where the value is, defending the screen is defending the wrong thing.
I agree with you on that. I’m on record saying I think talking to software is where things are in the future, and custom interfaces are part of that. It does beg the question of how the user base for technology changes, because there are two ways to slice this.
Bamboo Rose, or a similar product, historically sold licences and seats to a well-defined, well-scoped community of potential users who interact with it, input data, extract data and take actions on that basis. When you think about exposing it via MCP, or whatever becomes the codified standard, you’re exposing it to a completely different interface paradigm that is basically ubiquitous.
If the assumption is that everybody will have their Claude, ChatGPT or whatever interface, and that’s how we interact with things, it changes who you sell to and what the offer is. That’s fascinating to me, because it’s a fundamentally different approach to positioning and selling software.
Absolutely. That’s what I said at the beginning: our industry and the whole way we see software is changing. We should focus on the things we’re sure about. So many things are changing that we have to find out what’s not changing.
In our industry, one thing is very easy to find: people still have to get dressed. Clothing is one of the things you can’t digitise. Material has to be transformed. Products still have to be made and shipped. All the data around that is knowledge: how a product is built, costed, sourced, shipped and sold. That knowledge exists in a database and belongs to the brand. Nobody else has it. There’s a lot of value in it, and AI cannot change that. That’s what we’re betting on.
The model, the protocol, MCP or whatever, the screens and user interfaces: those are changing. We think they should be rented and designed to be replaceable, because they’ll get cheaper every quarter, even if AI vendors are still losing a lot of money. The tendency is probably that technology gets cheaper over time. It’s becoming a commodity.
The structured knowledge of how your business actually makes things and sells them is the asset. Nobody can sell you that. You cannot rent it from someone.
As the new software vendor we’re becoming in this economy, we have to create meaning on top of the data. Above all, we have to create context. A database field like ‘ITM_CSD_42’ doesn’t tell you anything. The words the business uses, like ‘costed BOM’, ‘first cost’ or ‘active suppliers’, are not in the database as of today. They’re only in the heads of the people who use it. We have to map them, maintain them and keep them up to date. That’s very important.
We also have to bring in governance and security so an AI agent can connect to the data. Once it understands the data, it has to connect securely. We have to make sure the right agents, for example an agent representing a merchandiser or a supplier, can do the right things and nothing else.
We have to bring in the business rules that already live in the software. That’s a very important part of software. It’s not something you can invent, because business rules are closely tied to how a business works, its culture and its values. In enterprise systems, those rules are integrated, and they must be transferred to the agents.
With our metadata layer, we’re building all this to make the data and the whole business model available to agents. That’s what we call ‘AI-native’ within Bamboo Rose: building this layer. We don’t want to compete on the models or the shiny stuff. We think we have to compete on how we organise our software, the data and its context so that it’s usable for AI agents.
Okay. So we don’t want to compete on the commodity models, and we don’t want to compete on the commodity interface. That leaves logic, rules, structure, governance and everything you’ve just described.
To poke at that a tiny bit, do you see that as a model that’s safe in an era where Anthropic or OpenAI could suddenly say, ‘The next vertical in our sights after financial services, legal and coding is retail, or fashion’? They could start baking some of that logic and those rules into a verticalised deployment of Claude or ChatGPT, with pre-established connectors.
What do you think is the defensible position of a software vendor with deep domain expertise if one of those two companies decides this is the next industry it wants to go after?
Nothing is safe today, in 2026. We’re going through a very deep shift. A big part of the intellectual work humanity has been doing for decades is now partly externalised to models, and done by models.
What was human labour is becoming capital, because it’s now done by machines, by chip vendors. It uses energy. This human work, our intellectual capacity, gets owned by big AI vendors like OpenAI and Anthropic, or maybe by the vendors of the hardware or energy behind them, because they’re also very strong and concentrated.
Again, we have to prepare for what’s stable and still there: the hard processes, the goods flowing, the logistics, the data and the knowledge behind them. That’s private knowledge owned by our customers, accumulated through years of operations. It has a specific value.
What we’re seeing, at least in the upcoming years, is general knowledge developing with these generally capable models. It replaces part of the intellectual work in our customers’ heads, but also partly in the rules of our software. Instead of hard-coding rules, you can now trust an AI to reason upon data. This generic software will be able to interact with the proprietary data we have to organise for our customers.
My next question is one I’d normally position as a risk to a technology company. To dig into the AI Report survey again, we’re entering a phase of shadow technology adoption. People use their company-provided tools, but they also bring their preferred tools, specifically AI tools, in from home. Based on the survey data, that happens a lot, and it often happens without permission.
In the early PC era, people brought personal computers to work because they felt those exceeded the capabilities of enterprise software at the time. Now, that creates a governance problem: if people do enterprise work with personal tools, it exists outside the ecosystem and databases we’ve talked about.
But given how you’ve spoken about the interface going away, and what is and isn’t a commodity, part of me wonders whether you see this as a threat, or as the way things are going to play out. Is the value in connecting to the tools we use at home and at work?
I think we have to start by accepting that you can’t win anything around this with a policy. The model someone uses at home costs €20 a month and gets better every few weeks. Nothing in enterprise procurement moves at that speed.
We’ve seen this before with social networks, ten or fifteen years ago. People got used to intuitive interfaces on Facebook, Twitter and Instagram, and wanted the same in their enterprise software. Banning it doesn’t remove it. It just moves it out of sight. For me, shadow AI is not people being undisciplined. It’s a signal that the company’s tools are too slow.
Our position is very specific. We think an LLM makes sense inside the product where things should run automatically: extracting data from a PDF, classifying data or executing a specific step. It doesn’t make sense when a user could just as easily ask their own tool, because they all have Claude and ChatGPT on their side. If a person can do the work alone by asking questions, they can already do it in ChatGPT or Claude. We’re not going to build a chat interface that’s better than those.
The other thing I’d say is that shadow AI is also free product research for us. What people build at home, on their own time, or increasingly as tools in their preferred AI assistants, like artifacts in Claude, tells you exactly where your software is failing or what’s missing. We treat that as a backlog in engineering and try to understand how we should and could integrate it.
The other challenge in reconciling the decision-making we’ve talked about is that people’s decisions are governed by their priorities. If we think about a complex brand workflow across the product journey, made up of all the different people we’ve talked about, there’s a lot of reconciliation in a typical product lifecycle.
One team or person is trying to stick to a margin target, someone else is trying to optimise lead time, somebody else is trying to hit sustainability targets. Do you think of AI as a way to align some of those micro-decisions to a common objective?
The part I think a lot of companies will get stuck on is that there isn’t necessarily a universal right choice for everything. We said there are instances where there’s only one correct choice in a narrow area. But sustainability, profitability, time and quality exist on a continuum. If you pull one in one direction, it pushes another in another direction.
That feels like the space where human judgement might win out: prioritisation, if you want to call it that. Do you agree that’s where human judgement will remain, or do you think that should be feasible to automate as well?
It’s a complex question. First, I think AI cannot solve a conflict about values. If margin, lead time and sustainability pull in totally different directions, it’s not a data problem. It’s a priority problem, and that belongs to the leadership. It’s a matter of values and company culture. Anyone telling you AI will decide on this, in 2026 at least, is lying to you.
I think this will become more important over time. Today, an agent does one task. As agents start working across functions, like merchandising and logistics, we’ll have to give them objectives and let them do the work alone. An objective is basically a value, or a set of values.
There’s the famous thought experiment where an AI is asked to produce as many paperclips as possible. To do it perfectly, it destroys everything else on the way, including humanity. That’s totally absurd, of course, but the mechanism behind it is completely ordinary and very realistic in a model. It’s just a mathematical objective, and nothing in that objective said what was more important than paperclips.
The company version is probably a bit more boring, but also much more likely. If you ask an agent to reduce landed cost, it will reduce landed cost. On the way, it could quietly trade away lead time, resilience, a supplier relationship or a sustainability commitment if nobody puts that into the brief.
That’s very complex. These are new tasks and things we have to learn within companies. When you give an objective to an agent, you absolutely have to make a choice about those priorities. Otherwise, the agent will make them for you. It will do it very consistently, with a lot of efficiency and at scale, without telling anybody. That’s very risky.
That’s an alignment question. But I’d say a lot of organisations aren’t especially aligned in how they value and prioritise human decision-making either. It’s a new problem from a model point of view. I don’t think it’s necessarily a new problem for a brand to say, ‘Yes, I recognise we’re going to miss some sustainability targets, but profitability and landed cost are everything for this particular style. Thanks for your input, sustainability team, but it’s been discarded in favour of the choice we’ve made over here.’ I get where you’re coming from.
Let’s think about an era where the database, governance, logic and rules are the value. The people able to work with those rules, influence them and enter data corresponding to them exist within a brand’s walls, but also across its extended supply chain and partner network. It’s not just your own people entering information into that big database.
If you have a PLM today, you probably have role-based licensing, with suppliers able to enter data. How do you see roles and permissions changing in the future? That sounds like a dry and boring question, but it feels quite important when the value is almost exclusively in the data, rather than anything you build on top of it.
It sounds like a boring question, but it’s a very important one. Before AI, permissions used to say what you were allowed to see. In an AI world, they must say what an agent can see, but also what it can do for you. That’s a very different level of risk, because we let the agent act.
There are three things changing at the same time. First, many more people get access to intelligence. That’s a very big change. Data that used to require a report from someone becomes available to anyone who can ask a question. AI brings a real democratisation of data to companies. It also means people will see numbers they’ve never seen before, and a lot of them simply aren’t prepared for that. Access without training can produce very bad decisions in some cases.
The second change is that the user is not always a person. When an agent asks something on behalf of a supplier, whose permissions apply? At Bamboo Rose, the decision we’ve made, and it was quite complex, is to attach every agent to a human in terms of responsibilities. If you connect with ChatGPT to Bamboo Rose and work through the ChatGPT interface, you attach through a human, or directly through a private key the agent gets. Every call and every task executed by the agent gets logged for traceability.
The third change we’re seeing is a strong increase in agents communicating with the outside world. That’s very new, but it’s a tendency that’s starting. In our software, we already handle roles for suppliers and partners. For the moment, human users are the big majority. They’re still connecting, importing data and documents. Suppliers bring in product specifications and so on.
If you extend that into an agentic world, we have to prepare to see our customers’ agents talking to their suppliers’ agents, with both sides’ permissions enforced. That’s where all this is going.
The position we have to defend is that permissions belong in the data layer, in our case the metadata layer, not in each application. If every new AI feature invents its own access rights, you’ll have five or ten different security models in two years and you won’t understand anything.
I think you’re right about a lot of that. People get obsessed with agentic commerce and tend to ignore the business-to-business agent relationships that are more likely to constitute the majority of traffic or throughput in that regard.
Penultimate thing. If I think back to the AI survey, companies tell us they expect spending on AI to go up in the next twelve months. A lot of that is apparently earmarked for spending more on talent or tokens to realise the value of things they’re already doing, rather than doing anything fundamentally new, but some is for newness as well.
You’ve made it very clear that AI is a fundamental shift. We shouldn’t look at it in the same way as traditional software. How does that influence the way you suggest companies approach ROI analysis? Traditional software metrics may not be well suited, or maybe they are. Give me an idea of how companies should look for value from their AI spend.
I think the traditional way of measuring software ROI was counting users and hours saved. That’s still very much true today, because those are the easiest numbers to collect and inflate. But neither tells you whether a decision is getting better.
One example of the KPIs we talk about with our customers is cost: the cost you’re already paying for something. If someone is retyping test results and certificate dates from a supplier PDF into the PLM interface, that’s cost, because it’s time-consuming. Someone chasing a vendor by email for a missing certificate and tracking expiry dates in a spreadsheet, or a merchandiser waiting three days for IT to run a query: all of that is cost.
It has headcount and hourly rates attached, so you know exactly how much it costs and how much you’ll save. That’s a very good KPI for measuring where you can save time. For PDF extraction, for example, we have technology that is automatically processing hundreds of thousands of documents for our customers each month, as we speak. That’s a lot of time saved for customers.
A second KPI that’s easy to measure is decisions getting taken earlier. That’s not very intuitive, but I think it’s important. The value is not necessarily a faster answer. It’s the answer arriving before the decision was locked. If you find a compliance gap before you award the vendor, instead of after, it’s worth a lot of savings.
That value will never be measurable in an hour. But you can measure how long decisions take by taking recurring decisions, vendor awards for example, and logging when the question comes up and when it’s signed off. Compare the gap before and after. Both dates are already in your systems. They’re stored in Bamboo Rose. Nobody has to fill in a timesheet. When you progressively reduce the difference between those dates, you’re creating value.
A last thing, and maybe it’s a strange one: there’s value related to AI that comes from questions nobody has ever asked before. When asking gets cheaper, people start to explore more.
We’ve built a technology called Assist, which helps our customers ask questions directly of our data and run queries using natural language. This is a good example of users running report requests that never got made before. When a team can ask its own questions, it stops calling IT, and IT stops building one-off reports. That produces a lot of value. You can easily measure it by the number of questions users ask in the tool, and by the number of IT tickets, which should be dropping for IT teams.
I think anyone listening will have experienced that in their professional and personal lives. You sit down in front of a text box, and everybody asks the same questions. In your personal life, it’s always, ‘Give me a morning briefing.’ It’s a cliché because you don’t need much imagination or experience to come up with it.
The more time people spend having natural-language conversations with LLMs that have access to vast repositories of accurate information across their organisation and the product journey, I think you’re right: people will start to ask deeper questions that don’t currently come to mind.
Final question. We’ve talked a lot about decisions. Arguably, the biggest decision we haven’t talked about is not whether to implement AI, because that seems pretty much a fait accompli. It’s more a case of when.
This is the fastest-moving field I’ve ever been involved in, analysed or observed. The state of the art advances incredibly quickly. New model releases have capabilities that transform what you can accomplish every month, week by week it seems at the moment.
If you put yourself in the shoes of a fashion brand, this can seem like such a moving target. If the frontier is forever moving and what something is capable of is forever changing, when do I jump in, and what’s the first step I should take? What would you say to a company that feels that way?
I think that’s perfectly rational. Anything we commit to today might be obsolete in less than a year. The last three years have shown us that. It’s very scary. We humans aren’t built for this uncertainty.
But the conclusion we draw, to wait until it settles, is the main mistake we have to avoid. It’s not going to settle down. There’s no stable version coming. The question is not when to start. The real question is what to build that survives being wrong.
Personally, I try to follow some guidelines to avoid getting lost. First, we absolutely have to separate what lasts from what doesn’t. We’ve talked about your data model, permissions and what your company knows about how it works. That lasts. You should invest heavily there. A specific model, tool or interface doesn’t last. It’s disposable. You should rent it and keep it as replaceable as possible.
Then we have to make small commitments. You don’t need an AI strategy. You need a few questions your team asks all the time and cannot answer quickly. Take the first one, fix it, and run it for a few weeks with real users and a result you can check. That’s the right approach.
Building everything so it’s easy to reverse is important in a fast-moving world. The cost of a wrong decision is not the decision itself. It’s how long you stay stuck with it. The quicker you can reverse your decisions, the more value you’ll keep.
That’s how we run this ourselves at Bamboo Rose. What lasts for us is the metadata layer, the permission model and so on. What doesn’t last is all the models around it. Every architecture decision we make undergoes the same test: if the model underneath changes next quarter, how much of this do we have to throw away?
I think the only really losing move in today’s world is to spend a year waiting for clarity that will never come.
I think that’s right. The other thing I’d add, which has been a big unlock for me personally and professionally, is using the agentic side of things, scheduled jobs and loops, as an opportunity to ask yourself about the schedule you do things on, and when external prompting and validation are helpful.
In a prosaic, personal way, every Friday a scheduled job runs and prompts me to go back through any emails I’ve missed. It’s the most basic use case you can imagine. The model can change around that, but for me as an individual, I’ve discovered it’s helpful to have that done on a Friday. It’s beneficial to how I operate.
You can extrapolate that boring, tiny example into how you think about workflows, calendars and everything else around a product journey. If we put some of this on autopilot, on a schedule, with jobs and loops that run regularly, what would they be and how would we scope them? Then, if you add new capabilities afterwards, it’s additive rather than destructive. At the very least, you’ve found something out about how you operate.
That’s absolutely right. The only thing you have to distinguish is what’s pure reasoning for the task you want to resolve, and where values, like personal or political values, come in.
In the first case, if it’s pure logical reasoning, a model is replaceable. You get a better model, one from another vendor or an open-source model, and you’ll get more or less the same results, because there’s a logical solution.
If you get to something that needs judgement, a model is no longer just replaceable. We were talking about sustainability, but in many bigger companies politics and the international geostrategic situation come in. If you take a US vendor, a European vendor or a Chinese open-source vendor, it will make a big difference. That’s something we have to take into account when we make these architecture choices.
All right. Rupert, I’ve enjoyed this chat a lot. Thanks for joining me.
Thank you.
That’s the end of my chat with Rupert. It was a long one, so I’m going to get out of your way quickly.
I think the most important thing to take away, and go and think about, is that software as we know it is up for interrogation, rebuilding and rearchitecting in the same way that processes are. People tend to think of the AI era as something happening to them, steered by technology companies. But remember that those companies are part of the same upheaval. The forward-thinking ones are figuring out how to get ahead of it, just like the forward-thinking brands.
Different topic next week. For now, thanks for listening, and I’ll speak to you again really soon.
