We built our holy grail in two weeks. Then we shelved it.

Towards the end of last year, we built something called inmydata studio. It is an AI-driven dashboard designer. You ask it, in plain English, to build you a sales dashboard. An agent engages with you, teases out the details it needs, and constructs the dashboard. Drill-downs are dynamic. The agent makes an intelligent guess at what might be useful to drill into, shows you that chart, then suggests alternatives and asks what you would like to see next.

We had spent 24 years building analytics software. Studio was the product we had always wanted to ship. The interface that finally got out of the way. The thing every analytics vendor in the market has been chasing, in one form or another, since BI became a category.

We built it in two weeks.

We are not promoting it. Our current business plans do not mention it, except as a source of lessons. Two years ago, if we had somehow conjured studio out of thin air, it would have been unthinkable not to have it front and centre of every marketing initiative we ran. Today it sits in production, used by a small number of customers, and we do not lead with it.

That decision is the most interesting thing about the project. Not because it was hard, although it was. Because it is the cleanest illustration I have of what is happening to the software market right now, and the bet that every founder and CTO running an established software business is being forced to make, whether they have noticed it yet or not.

What studio was, and what it took to build

For most of those 24 years, more than half of our development effort went into the UI. The drag-and-drop dashboard designer, the chart configuration, the drill paths, the formatting controls, the security model wired through every interaction. Anyone who has built an analytics product will recognise the shape of it. The UI was what we sold on. It was also, more importantly, our moat. The amount of work involved in replicating the functionality that interface represented was the reason a competitor could not show up tomorrow and eat our lunch.

When Claude Opus 4.5 was released in November 2025, we built studio in two weeks.

The thing that changed was not the intelligence of the model that powers studio at runtime. Studio actually runs on Haiku in production, and a smaller model is more than capable of the agent reasoning involved. What changed was the coding capability of the model we were building with. We had been iterating on the architecture for studio in our heads for years, the two-phase design and render split, the LangGraph agent flow, the prompt caching strategy, the multi-tenant schema isolation. Opus 4.5 was the first model good enough at coding that we could implement that plan, properly, in two weeks rather than eighteen months.

The two-week build was not a thin wrapper over a model API. The product that came out of it has multi-tenant schema-per-tenant isolation, AWS Cognito authentication, Stripe billing with overage tracking, team management, and a parallel data fetching engine that delivers sub-second dashboard renders. It is a full commercial SaaS product. It works because it sits on top of a platform we had been refining for 24 years, a database with a semantic layer that LLMs can understand, sub-second query performance, row-level security wired through every call, and the operational integrations to pull live data from ERPs and legacy systems.

We could not have built studio two years ago. The agentic coding capability did not exist, and the platform underneath it was not yet structured for AI consumption. We could not have built it without the platform either. The model alone would have given us a demo, not a product.

That is the first thing I want to be honest about. The UI we had spent 24 years building, the thing we had defended as our moat, was made worthless in two weeks. Not because we copied it. Because we built something materially better, an interface where you describe what you want and the agent constructs it, that made the dashboard designer we had been refining since 2003 look like a relic. The platform underneath it, the data infrastructure, the semantic layer, the security model, the integrations, the operational know-how, that was the actual moat. We had been pricing the wrong asset for 24 years.

Why we shelved it anyway

Here is where it gets harder. We could have led with studio. The product worked. Customers liked it. We had the credibility of two decades in analytics behind us, a sharp differentiating story, and a finished product that genuinely was what every BI vendor had been chasing.

Instead, we recognised something more uncomfortable. A better dashboard, even a perfect one, is still a dashboard. The way people work is about to shift, and a dashboard is an answer to a question that fewer and fewer people are going to be asking.

When an AI agent can reason over the data in your business systems, hold a conversation with you about your goals, and reason with you about the best way to achieve them, you do not need a static sales dashboard. You need something completely different. You need a conversational interface that knows your business, has access to your operational data, and can also access the things that have never been in a database, the judgment of your sales director, the lessons your regional manager learned the hard way, the reasons certain customers get certain treatment.

The dashboard is a frozen artefact of a moment when computers could not have a conversation. Now they can. The artefact is becoming optional.

If we led with studio, we would be selling the same kind of tool we had been selling for two decades, just with a better interface. We would be competing on the assumption that the dashboard is still the unit of analytics. That assumption is going to be wrong, and we did not want to spend the next five years discovering that in slow motion.

So we made a different call. We rebuilt the company around a Knowledge AI platform for agentic systems. The data infrastructure underneath studio became one half of it, exposing operational data via MCP servers so any agent can reason over it. The other half was the harder problem, capturing the tacit expertise that has never been in a database, through agent-led interviews, and serving it through the same retrieval layer. The dashboard becomes one of many possible surfaces an agent can produce, not the thing we sell.

That pivot is the bet. We are betting that the unit of value in business software is moving from the application to the substrate the agents work on top of, and that the substrate has to include the eighty percent of organisational knowledge that has never made it into a system.

The bet SAP is making, and the bet Salesforce is making

We are not the only ones placing this bet. Every major business application vendor is being forced to place one. Two of them have placed it in opposite directions in the last few months, and the contrast is the cleanest available illustration of what is at stake.

In late April 2026, SAP published an updated API policy. Section 2.2.2 prohibits the use of SAP APIs for “interaction or integration with (semi-)autonomous or generative AI systems that plan, select, or execute sequences of API calls.” In plain terms, no third-party AI agents on your SAP data unless SAP says so. The endorsed routes are SAP’s own Joule, the Business Data Cloud, and as of this week, Nvidia’s NemoClaw, following SAP’s $1.16 billion acquisition of German AI lab Prior Labs.

The framing from SAP is about stability, security, and protection of its domain expertise. Critics, including the German SAP user group DSAG, have read it as lock-in. Both readings are partially right, and both miss the bigger point.

The bigger point is what SAP is trying to preserve. SAP is trying to preserve the application layer. Their bet is that customers will accept a small number of endorsed agent stacks reaching into SAP, and that SAP itself remains the system of record and the system of work. The dashboard, the form, the workflow, the application UI, all of it stays where it is. Agents are an add-on, governed and channelled through SAP-approved routes. The application is still the product.

Salesforce has placed the opposite bet, and they have been remarkably explicit about it. Agentforce 3 added native MCP client support, allowing agents to connect to any MCP-compliant server without custom code, across a partner ecosystem that already includes AWS, Box, Cisco, Google Cloud, IBM, Notion, PayPal, Stripe, Teradata, and WRITER. Earlier this year, Salesforce launched Headless 360 and the Agentforce Experience Layer. Per VentureBeat’s reporting, this layer separates what an agent does from how it appears, rendering rich interactive components natively across Slack, mobile apps, Microsoft Teams, ChatGPT, Claude, Gemini, and any client supporting MCP apps. Slack is being re-engineered as the conversational interface for the entire agentic enterprise, with the rest of Salesforce being rebuilt to be Slack-first.

Honest readers will point out that Salesforce also tightened access to Slack APIs in 2025, restricting third-party tools from storing or indexing Slack data for AI training. That is true. But it is the exception that proves the framing. Salesforce did not restrict Slack APIs to keep agents out. They restricted them to channel agent value through their own platform. SAP wants the application to be untouchable except by approved agents. Salesforce wants the agent platform to be the thing customers cannot live without, and is willing to remake the rest of its product surface to make that true.

These are two different bets about what is worth preserving. SAP is preserving the application. Salesforce is preserving the substrate the agents work on top of.

I understand why SAP has made the bet they have. They have a vast installed base, an enormous amount of revenue tied to the existing application surface, and a real obligation to their customers to keep things stable. Locking the agent layer down protects all of that in the short term. Every quarter that passes without a third-party agent disrupting an SAP workflow is a quarter SAP keeps its margins intact.

But I think the bet is wrong. The SAP bet only works if the dashboard, the form, and the workflow remain the primary way users interact with their software. The moment conversation becomes the primary interaction, and an agent that can talk to your SAP data alongside your CRM, your warehouse system, your email and your Slack becomes more useful than the SAP UI itself, the walls SAP has built start to look less like protection and more like a cage. Customers will not tolerate being told that the most useful AI tool in their stack is the one their ERP vendor blesses. They will route around it, the way they have always routed around vendor-imposed limitations, and the routes will get faster every quarter.

The Salesforce bet is genuinely riskier in the short term. It involves cannibalising parts of the existing product surface in service of a future that may or may not arrive on the timeline they expect. But it is the bet that takes seriously the possibility that the user’s relationship with software is changing. The studio project told us, in the most direct way it could, that conversation is going to replace the application UI as the primary surface. Not entirely, not everywhere, but enough that building a business on the assumption that the application UI is the unit of value is the wrong bet. SAP is making that bet. We made the opposite bet the moment we decided not to lead with studio and to rebuild the company around a Knowledge AI platform for agentic systems instead. Studio is what proved to us the bet was the right one.

The lessons, if you are running a software business right now

A few things follow from this if you are in our position. None of them are comfortable.

The first is that the moat you think you have is probably not the moat you have. For 24 years we thought our moat was the UI. It was not. It was the platform underneath. If your team would say, if pressed, that your differentiator is the interface, the workflow, the user experience, the configurability, you should treat that as a hypothesis to test rather than a fact to defend. The right question is what would survive if a competent agent could replicate your interface in two weeks. Whatever survives is your real moat. Build on that.

The second is that the right response to the disruption is not to ship the same product with AI bolted on. We could have done this. Most analytics vendors are doing some version of this. It works, until the customer realises that what they actually wanted was not a smarter dashboard but a different kind of interaction altogether. The bolt-on strategy buys you a year or two of looking modern. It does not change where you sit in the value chain. If your bet is that the value chain is going to change, bolting AI on is the wrong move.

The third is that your existing product can become the demonstration of why your platform matters, rather than the thing you sell. Studio is not the thing we sell. It is the thing that proves we can build agentic interfaces over our platform faster than anyone without our platform can. That is a more useful role for it than being the headline product would have been. The reframe was painful. It was also right.

The fourth is the hardest. The decision to not lead with your best work, your most polished product, the thing you have aspired to for two decades, is the kind of decision that is easy to write about in a blog post and very hard to make in a board meeting. The pressure to monetise the work you have done is enormous, and entirely rational. The argument for not doing it has to be stronger than the argument for doing it, and the argument for doing it is right there in the spreadsheet. We made the call we made because the alternative was to spend five more years selling the right product to the wrong market. Other founders will reach different conclusions, and some of them will be right.

What I am certain of is that the bet is being placed whether you place it consciously or not. SAP has placed theirs. Salesforce has placed theirs. Every software business with more than a year of history is in the middle of placing theirs right now, in the form of which features get the next sprint, which roadmap items get killed, which conversations with customers get listened to and which get filed away.

We placed ours by shelving the best product we have ever built. The platform underneath it is what we are betting the company on. Two years from now we will know whether that was the right call. The thing I am sure of, today, is that the call had to be made.