Stop counting your software as an asset

On Thursday morning I connected Claude to our demo data and typed one line. I want to try Claude Dashboards and Motion on the Inmystore sales data. No brief, no questions, no hints about what to look for.

The first sentence it wrote back was the one I would have written. Sales through period 4 were up 5.2 per cent on last year. From period 5 to period 9 they were down 62 per cent. So the turn at period 5 is what needs explaining.

I know what happened at period 5. Inmystore is our test dataset, and in May we changed the way the data is generated. The model found the seam, worked out what it looked like, and wrote the explainer.

That is our dashboard designer gone, as a product, in one morning. I think it is good news. To explain why, I need to go through what Anthropic shipped, how the analytics industry is responding, and what it means for anyone whose company’s most prized asset is a piece of software.

What Anthropic shipped

On 8 October Anthropic released Claude Dashboards and Claude Motion in beta. Dashboards is a template for a Claude artifact. You connect a warehouse or any app Claude can reach, ask a question in plain language, and Claude writes one query per chart, runs it, and lays out the page. Every chart shows the query behind it and when its data was last refreshed. The dashboard stays live.

Motion is the same idea applied to film. Claude writes code that animates text, charts and figures, so you can change a word or a number, and exports the result as a video. There is no video model involved.

OpenAI shipped a near identical data agent for ChatGPT about a month earlier. Anthropic positions its version for quick, exploratory questions and lists the BI tools you can hand a dashboard on to. The two largest model vendors now both treat a dashboard as something you ask for, and both are careful to say they are not replacing anyone.

The industry’s answer

The response from the business intelligence world settled into one sentence within a week. Great analyst, terrible source of truth.

The argument goes like this. Claude writes a fresh query for each chart, so two charts can define revenue differently. There is no semantic layer, no certified data source, no row-level security, no audit trail, no scheduled refresh. So it is exploration, and governed reporting stays where it is. One head of data called it the end of governance as we know it if left unchecked.

I spent twenty years in this industry. That argument is half right, and the half that is right does not help the vendors at all.

They are right that a window onto ungoverned data is worth nothing. Where they are wrong is in where the governance belongs. It never belonged in the chart. It belongs in the layer underneath, the one the agent reads from. Once it lives there, the window on top is free, and anyone can draw it.

That is what my test showed. Claude never wrote SQL against raw tables. It called our platform through an MCP server. The platform already carried the metadata that told the model what each field meant, the security that decided what this user could see, and the speed to summarise millions of rows while the conversation waited. The semantic layer was there. It just was not in the dashboard.

The window was never the point

We built our own agent driven dashboard designer about a year ago. It took two weeks. We said at the time, and I said again on stage at ScotSoft and PUG Challenge this autumn, that it was no moat. Anything a small company can build in two weeks, anyone can build in two weeks, and now a model can build it in a conversation.

A dashboard is a window. You look through it, you form a judgement, and you take an action. The point was never the window. It was the action. That is why we moved away from selling analytics and towards building systems that act. A living knowledge base gives the agent judgement. The data platform gives it governed operational data. Then we give it the tools to do something.

So when Claude drew our dashboard for free, it took the part we valued least. The platform underneath became more useful the same morning, because it now serves a window we did not have to build.

The platform is not safe either

Here is the part I find harder to say. The platform took us nearly five years to build. It has three properties that make it worth connecting an agent to. A metadata layer so the model reads the data correctly. A governance layer so access is secure and auditable. Speed, because a conversation cannot wait for a query.

We have just rebuilt it from the ground up. Storage agnostic, per-user security, read only by construction, served through an MCP server. The rebuild took under three months and goes live in the next few weeks.

Five years to three months is the sound of the engineering moat draining. The expertise it takes to build a data platform is still real, but the models are absorbing it fast. What they could not absorb was knowing which three properties mattered, and what correct looked like for each. That knowledge is what let a small team do in a quarter what took half a decade the first time.

What is still worth counting

The analytics industry is the case study because it is the one I can measure on my own products. Three layers, three speeds. The window went in a morning. The platform went from five years to three months. The thing that let us do both is the only thing that did not get cheaper.

The same is true of whatever you sell. The CRM, the ERP, the scheduling system, the vertical SaaS that took a decade to get right. The screens will be drawn by a model. The plumbing will be rebuilt by one. If your balance sheet says the software is the asset, it is counting the part that is losing value fastest.

What is left is worth more than it has ever been. Domain knowledge, because the model can build anything but cannot know what to build. Experience, because the model can produce a correct answer but cannot tell you which of five correct answers your customer actually needs. And agility, which is what domain knowledge and experience give you when the build cost falls to almost nothing. We did not rebuild the platform in three months because we type fast. We did it because we knew exactly what it had to be.

So put the governance under the agent. Let the model have the window. Then take a hard look at what you are counting as an asset, and move the number to the people who know why the software exists.