Anthropic just made a live dashboard about as easy to generate as an answer in a chat window. Building the chart has been the most visible cost of self-service analytics, but Claude Dashboards now pushes that cost close to zero. But keeping that dashboard accurate carries a second cost. Someone still has to maintain the definitions, filters and table choices that determine what each number measures. I expect Claude Dashboards to make that upkeep a much bigger share of what analytics costs.

(Note: Of the companies mentioned in this article, Amazon and Salesforce are clients of my firm, Moor Insights & Strategy. Anthropic is not a client.)

What Claude Dashboards Does

Anthropic put Claude Dashboards into beta on its paid plans earlier this month. A user types a question in everyday language, and Claude pulls the data from a connected source and builds the dashboard. Anthropic lists Amazon Redshift, BigQuery, ClickHouse, Databricks, Snowflake and Salesforce among those sources. It says the dashboard keeps updating as the data changes.

I like that a click on any number in the dashboard shows the query that produced it. Seeing the query is one of the fastest ways for an analyst to double-check a chart before trusting it. Anthropic pitches Claude Dashboards for quick questions and says it’s meant to sit next to the business intelligence tools a company already runs. For anything more involved, a user can send the dashboard from Claude to a partner analytics tool such as Hex, Omni or Sigma. Any deeper analysis then happens in that tool, not in Claude.

Anthropic’s Own Data Team Found That Accuracy Requires Upkeep

Some of the best evidence for where analytics costs are heading came out four months before the launch, from Anthropic itself. In a June post about its own analytics work , the company’s data team said that Claude automates 95% of Anthropic’s business analytics queries at about 95% accuracy overall. A big part of getting there was a set of plain-text guides the team calls “skills.” They tell Claude which sources to check first, how to handle ambiguous data and when an answer is complete. Without those guides, Claude never topped 21% accuracy on the team’s internal tests. Adding them pushed accuracy above 95%.

The guides also need steady upkeep. Anthropic’s data model, meaning how its data is organized into tables and fields, changes every day. A guide that nobody updates goes out of date within weeks. The team’s tests showed accuracy sliding from about 95% to about 65% in the space of a month without regular maintenance. After that, Anthropic started treating guide upkeep as engineering work. Now each guide sits next to the data model it describes, so the change that alters a model can update its guide in the same step. Roughly nine in 10 of the team’s data-model changes include a guide update.

The team also tried having an AI model write the definitions of metrics straight from raw tables and old queries. Those drafts scored worse on the team’s tests than a smaller set of definitions that people had built by hand. The drafts looked sound but kept the same ambiguities Anthropic had been trying to remove. Its advice now is to let Claude draft the overall documentation while a person stays in charge of each definition.

All of these results point to the fact that most of the difficulty in AI-driven self-service analytics comes from ambiguous data. Once a question maps to the right data, writing the query is the easy part. And while Claude Dashboards takes care of the query writing and chart building, working out what ambiguous data means still requires the input of people. It also requires the analytics and data management vendors that have spent years building tools to settle what a company’s numbers mean.

More Dashboards Mean More Logic To Keep Straight

Many well-run BI deployments rest on a semantic layer, which records what each business term means and how to calculate it. The charts people look at sit on top of that layer. Claude Dashboards builds those charts quickly, but I’d argue that trust in a number still comes from the layer underneath, and someone has to maintain it.

Cheaper dashboards should also mean many more of them. When anyone can build one by asking Claude a question, I’d expect plenty of them never to reach a data team. Each one carries its own choices about filters, date ranges and which table counts as the source. A dashboard the data team doesn’t know about can easily run on logic nobody is maintaining. Keeping every number tied to the semantic layer is manageable when a single team builds the dashboards. I think it gets much harder once a whole company is building its own.

Is Selling Dashboards Still A Business?

I hear people say dashboards are dead, and they mean one of two things by it. Some people expect charts to fade away once a chat window can answer questions directly. I think plenty of quick questions are moving into chat, sure, but a chart still shows a trend faster than a paragraph of text can. (A picture is worth…) When I say dashboards are dead, I mean they’re a market dead end. Companies will keep building and using them. But with an AI assistant now building one on request, selling dashboards on their own has become too commoditized to earn a living from.

Investors have sold off whole groups of stocks after Anthropic launches this year. Legal research, cybersecurity and design software companies each took a hit when Anthropic released a tool aimed at their kind of work. BI vendors could look like the next group in line.

I’ve thought the “SaaSpocalypse” was overblown from the start, though. AI coding tools are excellent, and they keep getting better. Someday a large company with compliance rules to meet may trust a vibe-coded app to run a mission-critical system. We’re not close to that yet. I think Claude Dashboards poses less danger to BI vendors than a selloff would imply, at least to the vendors that hold a company’s business definitions.

Analytics Vendors Are Opening Their Definitions To AI Assistants

Nearly every BI and data platform vendor I cover now lets outside AI assistants reach into its semantic layer. Most of those vendors use the Model Context Protocol, an open standard that lets AI assistants connect to outside tools and data. That way, a question typed into a general-purpose assistant can draw on the metrics definitions that a company has already approved inside the vendor’s product. The pitches I’ve seen promise that the access rules in a vendor’s own tools will still apply.

A vendor that opens its definitions to any assistant accepts that some users may never open its dashboards. I read the move as a bet that customers will keep paying for the place where each number’s definition is written down. I’m also seeing analytics vendors build AI agents that watch key metrics and explain why one moved. When those explanations draw on the semantic layer, a stale definition does more damage with every feature built on top of it.

This also applies to buyers, for whom each new assistant connection raises the stakes on whatever semantic layer they already run. Every assistant that pulls its numbers from that layer inherits the layer’s definitions, whether those are current or stale, and those definitions largely determine whether the assistant’s charts are right. So as AI model providers push further into analytics, I’d expect them to want more control over the semantic layer, too. In a July piece on why everyone in AI now sells context , I argued that companies should keep that control for themselves.

What To Decide Before Turning On Claude Dashboards

By default, Claude Dashboards is switched off under Anthropic’s Enterprise plan, which gives customers room to settle three decisions:

  • Pick what the assistant will query. Connect Claude first to sources that the data team controls. Good options include a semantic layer reached through a connector, or tables the team has already approved. A chart built on approved definitions starts from logic someone already checked, which greatly improves the odds that it’s actually reflecting what you think it is.
  • Fund the upkeep. Give each core metric an owner, and update the data model’s documentation whenever the model itself changes. That upkeep deserves its own line in the analytics budget, next to the software licenses.
  • Set a graduation rule. Decide when a chat-built dashboard will move into the company’s official BI tool. Two sensible triggers are (1) the moment it’s shared beyond the person who initially built it and (2) the moment someone uses it to make a spending decision.

My guess is that Claude Dashboards will put live charts in front of a lot of people who rarely open a BI tool. More people asking questions of their data directly is a healthy change for analytics. But I do think the companies that benefit most will be the ones that already pay people to look after what their numbers mean.