AI-generated dashboards can take the dashboard-drawing part of Looker, but not the whole BI seat, on the best public evidence as of 28 October 2026. That evidence is David Lequin's write-up of 14 October: three data platform engineers at a company of about 500 people built an internal platform for agent-written HTML dashboards, starting in February. In the week of 5 October it had 156 weekly users against Looker's 157, and a week later their analyst agent spent about $0.53 of tokens per dashboard. They did not cancel Looker. My recommendation: move dashboard drawing to an agent only when the warehouse already enforces permissions for each viewer and the dbt models already hold the metric definitions. Keep a BI tool for governed metrics, external viewers and Explore-style self-service.
Looker sells four jobs, and only one is drawing charts #
A Looker replacement has to carry four jobs, and Lequin lists them: "When we bought Looker we were buying 4 things: a UI to draw charts, a permission model, a semantic layer, and somewhere to host it." A coding agent takes over the first job outright. The other three do not disappear. They move to the warehouse, the transformation layer and a hosting service that someone has to run.
His team already owned those three: "Snowflake had the permissions all along, dbt has the semantics, and hosting some HTML files costs about $500 a month." A team in the same position can swap the chart-drawing layer. A team without them has to build all four.
The money at stake is also less certain than the title. The "$150k Looker" is not the company's invoice. Lequin writes "Let's say $150k a year, the average Looker licence, roughly 300 seats", and traces the figure to a Vendr average quoted by Coefficient, a vendor that sells a competing product. Coefficient's page puts that average at "50-100 users", not 300. Google's Looker pricing page lists every edition as "Call sales", so no public price exists to check either number against.
The stack that replaced Looker's dashboards is more than a bucket #
The stack behind the agent-written dashboards at Lequin's company started as files in a bucket and grew into a service. In February a teammate built, in one afternoon, "a small service that stored the files in S3, SSO in front of it, and (some) tests". By October the platform had 325 commits and about 74,000 lines of code, 20,000 of them tests, mostly written by Claude.
| Slot | What the team uses |
|---|---|
| Warehouse and permissions | Snowflake roles, queried with each viewer's own Snowflake OAuth token, the only login since April |
| Metric definitions | A dbt project of more than 2,000 models |
| Agent | Claude ,Claude Code and their own analyst agent, which moved toClaude Opus 5.5 in the week of 28 September |
| Publishing | One remote MCP connector for the whole company; anything an AI uploads lands as a draft until a person clicks Publish |
| Hosting | Amazon S3 , a service onKubernetes ,PostgreSQL for versions and grants |
| Isolation | Each dashboard in a sandboxed iframe on its own origin, no cookies, strict CSP, one-hour tokens for one dashboard and one viewer |
| Review | A metadata service that records the tables, scripts and network destinations each dashboard uses, and runs rules on them |
The live query API, added in March, is what Lequin calls the "best decision of the year". It also turns static HTML into an application: dashboards query Snowflake at view time as the person looking at them. The team also runs Apache Airflow, Airbyte and DataHub underneath; the post does not tie them to the dashboards.
What the team measured over eight months #
Lequin's team published two series as data inside the post's chart script: weekly users of Looker and of the new platform, and the analyst agent's cost per dashboard. Both are self-reported. The cost series comes from the agent's run logs; the post does not say how users were counted. The platform's event log starts on 16 March, so the user series covers about seven months, not eight.
| Week of | Looker weekly users | Platform weekly users |
|---|---|---|
| February 2026 (average of 4 weeks) | 230 | built 20 February, no log |
| 16 March 2026 | 223 | 23 |
| 22 June 2026 | 184 | 136 |
| 14 September 2026 | 147 | 143 |
| 5 October 2026 | 157 | 156 |
By my count of the series, Looker fell 32% from its February average, and the platform never passed it in any week. The post's "as many weekly users as Looker in our company, maybe even more" holds only for the first half. In Lequin's words, "Looker's usage leaked rather than crashed": "some teams still live in Looker, and some of its dashboards have no replacement yet."
The cost series covers ten weeks, 10 August to 12 October, and only the analyst agent. Cost per dashboard peaked at $2.59 in the week of 7 September and was $0.53 in the week of 12 October. Lequin credits the switch to Opus 5.5: "The agent does the same work; the tokens just got cheaper." His data shows two causes. Between the early September peak and the two October weeks, the price per million tokens fell from about $1.75 to between $0.93 and $1.30, and the tokens per dashboard fell from 1.46 million to between 0.43 and 0.57 million. Part of the token drop came on 21 September, a week before the switch.
Five ways to fill the BI slot, and what each commits the team to #
The options for a team weighing Looker against agent-written dashboards fall into five groups as of 28 October 2026. They differ less in price than in where the permission model and the metric definitions live, and in who has to run what.
| Option | Where permissions and metrics live | What the team takes on | Price, licence |
|---|---|---|---|
| Keep Looker | LookML model and Looker's own row-level controls | The contract | Quote only |
| Build your own, as Lequin did | Warehouse roles and dbt models; metric rules only warn | A service, a sandbox, a security review, the on-call | Engineer time, plus about $20k a year in AWS and tokens at his scale |
| Claude Dashboards | Not stated in the announcement | A beta feature inside the Claude plan | Paid Claude plans, beta since 8 October 2026 |
| Dashboards as code: dbt Charts ,Evidence ,Rill | dbt models or YAML metric files in git | A repository and a deploy, reviewed like code | Apache-2.0 or MIT cores; Evidence Studio $2,500 a month, Rill from $20 a seat |
| AI-native BI on a semantic layer: WrenAI ,Lightdash ,Omni ,Hex | The tool's modeling layer | A new vendor, or self-hosting an open core | WrenAI Cloud $179 or $559 a month, billed annually; Lightdash Cloud Pro $3,000 a month; Omni quote only; Hex from $36 an editor |
Claude Dashboards is the closest product to Lequin's build. Anthropic says the dashboards "work alongside your BI and analytics tools", not instead of them, and Lequin writes that a team starting today should "probably look at it first". WrenAI keeps its dashboards and row- and column-level security in the commercial tier; the Apache-2.0 core is the text-to-SQL engine with no UI.
Which BI jobs do agent-written dashboards not take over? #
Agent-written dashboards, in Lequin's build and in the products above, leave several BI jobs uncovered or only partly covered. Self-serve exploration is the biggest: Looker's Standard User licence includes Explore, SQL Runner and drill to row-level detail, while in the HTML model a person who wants a different cut asks the agent again. The post does not describe write-back, threshold alerts or analytics embedded for customers.
| BI job | In Lequin's build |
|---|---|
| Drawing and changing charts | Yes, by prompt |
| Exports to email and Slack, and schedules | Yes; email and Slack exports since April, scheduling later, the scheduler not described |
| Data permissions | Yes, through Snowflake roles and the viewer's own login; row-level only where Snowflake row access policies exist |
| Governed metrics | Advisory; reading raw tables raises a flag, not a block |
| Stale and duplicate dashboards | Partial; a rule flags dashboards nobody opened in 90 days |
| Explore-style self-service | No; re-prompting instead |
| Embedded analytics for customers | No; every viewer needs a Snowflake user |
| Write-back | Not addressed |
The Hacker News thread on the post (133 points and 80 comments on 28 October) pressed on the same gaps. One commenter, yfontana, wrote that BI needs "proper access control (with RLS/CLS)" and governance "so that the same metric doesn't get re-defined in 12 different places".
Where do the semantic layer and access control go? #
The semantic layer and access control do not come with an agent; each has to live somewhere the dashboard cannot bypass. Hosting-layer controls such as SSO, Google's IAP or Cloudflare Access decide who can open a page, not which rows that person sees. A static HTML file with data baked in has no row-level security at all.
Row-level control needs the dashboard to query the warehouse as the viewer, which is Lequin's choice, or to go through a server that enforces the rules: a BI tool, or a semantic layer such as Cube or the dbt Semantic Layer. The dbt docs say AI tools can connect "through the dbt MCP server, so their answers use your governed metrics instead of guessing at raw tables". That needs a paid dbt Starter or Enterprise account, and Snowflake's own row access policies need its Enterprise Edition.
Metric governance is the weaker half of Lequin's build. A dashboard that reads raw data instead of a dbt model gets a flag and still loads, by choice: "Every constraint you put between people and their dashboard sends them back to the Downloads folder." Two dashboards can then disagree on revenue, and only someone reading the flag will notice.
The cost comparison leaves out the engineers #
Lequin's bill puts building the platform at $8.3k over eight months ($6.4k of Claude tokens, $1.9k of AWS) against about $100k for eight months of a $150k Looker licence, and running it at about $20k a year. He names the gap himself: "the biggest cost, which is us: 3 engineers spending 8 months on this, on top of everything else a data platform team does", with no number on the hours.
Three more costs are missing. The cost series counts only the analyst agent, not the tokens of everyone else who prompts a dashboard. Claude seats sit outside the bill, though the company presumably pays for them anyway. Snowflake compute for each viewer's live queries is not mentioned. Lequin is also part of data-stack, a studio that sells AI agent engineering to companies.
AI-generated dashboards are a security project #
Agent-written HTML dashboards are untrusted code by design, and Lequin says so: "The whole product is basically XSS by design… We can't block the XSS, because the XSS is the product, so our job is to make sure it can't reach anything." His first version ran dashboard JavaScript on the same origin as the user's session, and "every employee could read every dashboard". It shipped because it sat behind VPN and SSO.
The current design answers with an iframe per dashboard on its own origin, no cookies, a strict CSP, short-lived tokens, and rules that block scripts from unknown packages and WebRTC connections. A team copying the idea has to copy that part and keep it reviewed. Claude Dashboards and the hosted tools move that work to a vendor.
Replacing Looker is a bad trade in five cases #
Replacing Looker with agent-written dashboards is a bad trade when the conditions Lequin's team had are missing. His post, the Hacker News thread and the vendors' own docs point to five cases where the BI tool or a semantic layer should stay:
- No platform team to run it. Lequin: "If you need to hire a team to build this, the maths is different."
- Viewers without warehouse logins , such as customers or partners.
- Certified numbers , such as finance or regulated reporting, where a warning is not enough.
- Heavy Explore users who slice data themselves and do not want to prompt.
- A large LookML model that is the real asset.Cube , which sells a semantic layer, puts it this way: "Replace the dashboards and lose the model, and you've traded down."
On Stackness, as of 28 October 2026, no public stack lists Snowflake, Metabase or Evidence, and Looker joined the catalog today (how I count). The numbers are small; most of the data tools in this post sit among the DevOps and infrastructure tools developers list.
What would change the recommendation #
The recommendation to keep a BI tool next to agent-written dashboards rests on one self-reported post and a three-week-old beta. If Claude Dashboards leaves beta with warehouse permissions and a published price, it becomes the first thing to try for a team without a platform group. If a second team publishes usage data, especially one that cancelled Looker outright, the case gets stronger. A count of how often agent-written dashboards disagreed with the governed numbers would matter most, and nobody has published one.
What I would do #
- If the warehouse already enforces permissions for each viewer and dbt holds the metrics, let an agent write the dashboards, publish through a draft-and-approve step, and keep Looker for the teams that still use it. Measure weekly users on both before cutting seats.
- If there is no platform team, try Claude Dashboards or a dashboards-as-code tool like dbt Charts or Evidence before building anything.
- If customers see the dashboards, or the numbers are certified, keep a BI tool or put a semantic layer in front of the agent, and treat agent-written pages as drafts.
- If the LookML model is large and trusted, renegotiate the Looker contract on seat count before replacing it.
New posts by email #
About one email a week with what we published. No account needed.
Which of these do you use? #
Nobody on Stackness runs Looker yet. Tick what you use and be the first.