The Complete Agent Stack · Part 09 of 14 · Distribution layer

This guide is one part of a 14-part series on what a production data agent needs, from the data model it reads to the agent that keeps it accurate. Each part stands on its own, and they build on each other in order.

← Previous: Part 08: AI-Generated Dashboards · Next: Part 10: Claude MCP for Data Analysis →

Part 09 continues the distribution layer, picking up where Part 08 on AI-generated dashboards left off.

Key Takeaways

• Generated interfaces now compare with expert work: In a Google Research study of Generative UI, people overwhelmingly preferred model-generated interfaces to standard markdown answers, and rated them at least comparable to expert-crafted designs in 50% of cases.

• The polish step is usually manual: A review of 83 dashboards by Sarikaya and colleagues notes that BI users often screenshot dashboards into slide presentations to add context, which is the hand-built step an AI report builder replaces.

• Brand colours have to pass contrast rules: WCAG 2.2 sets a 4.5:1 contrast ratio for normal text and a 3:1 ratio for graphical objects such as chart lines and pie slices.

• Public links should carry snapshots: The W3C's guidance on capability URLs warns that shareable links leak through browser history, referrer headers and search crawlers, so a public report should never grant live data access on its own.

The 45-second version:

What Is an AI Report Builder?

An AI report builder takes analysis that already exists, such as a set of charts on a dashboard, and generates a finished report from a short design brief. The agent writes the title, pulls headline figures into summary cards, restyles each chart and lays the page out in the organisation's visual identity. The output is a single web page, usually HTML with interactive charts, that can be shared, embedded or sent as a link.

This is a different job from the AI reporting tools that answer questions on demand, which our guide to AI reporting tools covers. A report builder works on the presentation layer. The analysis has been done, and the task is to make it legible and credible to an audience outside the team that did it.

Research on generative interfaces explains why this has become practical. In Generative UI: LLMs are Effective UI Generators, a Google Research team describes models that generate the interface itself instead of a markdown "wall of text". With generation speed set aside, human raters strongly preferred the generated interfaces to markdown output, though expert-built pages still ranked highest. The accompanying Google Research post is candid about the limits: generation can take a minute or more, and outputs contain occasional inaccuracies. Both points shape how a report builder should be used.

Why Don't Dashboards Look Like Your Company?

Dashboards look like the tool that made them. Default fonts, default chart colours and the vendor's layout grid are fine for an analyst's own work. They become a problem when the same board goes to customers, executives or a board meeting, where the presentation is read as a statement about the organisation. Our guide to white-label analytics makes the point for embedded products: every visible element, including exports and shared links, should carry the host company's identity.

The usual fix is manual. Someone takes screenshots, drops them into a branded slide template and adds a title, a few headline numbers and some commentary. Sarikaya and colleagues observed exactly this in BI practice and read it as a sign that dashboards capture the data without the context people need to communicate it. A deck built this way looks right, but it is frozen, it has lost every interaction, and rebuilding it next month repeats the same work.

The alternative of a designer and a developer building a custom report page produces the best result and the slowest refresh cycle. Each new version needs the same people again.

What Does a Branded Data App Need?

A branded data app is more than a recoloured dashboard. Five properties separate a report people trust from one that only looks finished.

Brand inputs the agent can follow

The agent needs a logo and a clear design direction: colours, typography, layout and tone. The more precise the brief, the less the agent has to guess. Teams that already maintain design tokens have a head start. The W3C Design Tokens Community Group, whose specification reached its first stable version, describes tokens as the single source of truth for colours, typography, spacing and other design decisions, with support for brand themes. Pasting exact hex values and font names into the brief gives the agent the same source of truth.

A structure that tells the story

Segel and Heer's study of narrative visualisation identified seven genres, including magazine style, annotated chart, comic strip and slide show, and placed each on a spectrum between author-driven and reader-driven designs. One pattern they describe, the martini glass, follows a tight narrative path first and then opens up for free exploration. A branded report app fits that shape well. The title and headline stat cards carry the author's message, and the interactive charts below let each reader explore.

Numbers that come from the data

Headline figures carry the most weight and attract the most scrutiny, so they must be computed from the charts beneath them and never paraphrased by the model. Given the occasional inaccuracies Google's researchers report, the review step matters: before sharing, check each stat card against the chart it summarises.

Colour that stays readable

Brand palettes often include a pale accent that looks good on a logo and fails on a chart. Under WCAG 2.2 success criterion 1.4.3, normal text needs a contrast ratio of at least 4.5:1, large text 3:1, and logotypes are exempt. Under criterion 1.4.11, graphical objects needed to understand content, including individual lines on a line chart and the slices of a pie chart, need at least 3:1 against adjacent colours. A good brief asks the agent to keep brand colours for accents and use high-contrast variants for data.

A path to refresh

A branded report that cannot be regenerated becomes another static file. The app should be rebuilt from the underlying dashboard whenever the data or the analysis changes, so the polish is applied again in a single step instead of by hand.

How Should You Share a Branded Report?

A polished report is made to travel, which makes its sharing model the most important security decision in the workflow. The W3C Technical Architecture Group defines a capability URL as one where anyone who possesses the URL can access the resource, and stresses that URLs are not generally kept secret. Its guidance lists the leak routes: browser history and its synchronisation, referrer headers, third-party scripts, URL shorteners, shared screens and search crawlers. A QR code printed on a slide is the same link in another form.

Three rules follow from that guidance.

• Public means snapshot. A public link should contain the data as it was when the report was generated, with no live connection to the database behind it.

• Live access needs a second secret. If a link must show current data, require something beyond the URL, such as a passkey, and apply the viewer's own row-level security. Our guide to row-level security for AI agents covers the server-side rules that make this safe.

• Prefer named sharing inside the organisation. Sharing with the organisation or with specific people keeps access tied to accounts, which can be revoked, instead of to a link, which cannot be recalled once forwarded.

The W3C guidance also recommends expiring or revocable links and keeping third-party scripts off pages served through capability URLs, both of which apply directly to public reports.

Why Is Generated HTML a Security Question?

An AI report builder writes code, and that code runs in a viewer's browser. The OWASP Top 10 for LLM Applications lists Improper Output Handling as a distinct risk, including the case where an LLM generates JavaScript or Markdown that the browser interprets, resulting in cross-site scripting. OWASP recommends treating the model like any other untrusted user, encoding its output and applying a strict Content Security Policy.

For report apps, three precautions cover most of the risk. Serve generated pages from an isolated origin or inside a sandboxed frame; MDN's documentation of the iframe sandbox attribute explains that an empty value applies every restriction, and warns against combining allow-scripts with allow-same-origin for same-origin content. Apply a strict Content Security Policy that blocks unexpected network requests. Embed snapshot data in the page instead of credentials or query endpoints, so a compromised page has nothing live to reach.

How Upsolve Approaches Branded HTML Apps

Branded HTML Apps turns any Upsolve canvas into a branded report. On a canvas, the user clicks Make it an app, uploads a logo and describes the colours, layout and design they want. An agent rebuilds the canvas as a polished HTML app with a title, stat cards and restyled charts, and viewers can hover over values in the finished page. The original prompt stays in a side panel, so the user can keep prompting to change or remove anything.

In the demo, a sample user signed in to the embedded iframe version of Upsolve opens an existing canvas built from favourited charts. With the Upsolve logo uploaded and a pre-written design prompt, the agent produces a preview with a title, stat cards drawn from the charts' numbers, and restyled bar, line and donut charts. The user saves it as an app.

Sharing follows the canvas model from Part 08. An app can be shared with the whole organisation, with specific colleagues, or through a public link and QR code. In the demo, the user chooses snapshot data and the branded view without a passkey, and the public link opens for anyone. As Serguei Balanovich, Upsolve, explained:

"Public links are actually not allowed to be live links unless they are provided with a passkey. We don't want to give live access to underlying data to anybody who comes across it."

Refreshing works through the canvas. The user can rearrange it, add favourites, or open any chart's conversation history to drill further, then regenerate the app with the latest numbers. For the full feature list, see the Branded HTML Apps feature page.

Frequently Asked Questions

What is an AI report builder?

An AI report builder is a tool that turns existing analysis into a finished, presentable report from a short design brief. Given a dashboard or a set of charts, the agent writes a title, extracts headline figures into summary cards, restyles the charts and lays out the page in your brand's colours and typography. The output is typically an interactive HTML page that can be shared or embedded. It sits after the analysis step, so its job is presentation.

How do I make a dashboard match my brand?

Give the agent precise inputs: your logo, exact colour values, font names and a short description of the layout and tone you want. If your organisation maintains design tokens, use those values, because they are the agreed source of truth for brand decisions. Then review the result for contrast. WCAG 2.2 asks for 4.5:1 for normal text and 3:1 for chart elements, and pale brand accents often fail on data. Keep iterating with follow-up prompts until the layout fits.

Are AI-generated reports accurate?

They are as accurate as the data and checks behind them. The charts in a well-built report come from queries against governed data, so the main risk is in what the model writes around them, especially headline figures. Google's Generative UI researchers note occasional inaccuracies in generated outputs. Check every stat card against the chart it summarises, keep each chart linked to the conversation that produced it, and regenerate the report when the underlying analysis changes instead of editing figures by hand.

How do I share a report publicly without exposing live data?

Publish a snapshot. A public link should contain the data as it was when the report was generated, with no connection back to your database. W3C guidance on capability URLs warns that links leak through browser history, referrer headers and search crawlers, so assume any public link will eventually reach people you did not intend. If someone needs current figures, share with their account instead, or require a passkey and apply their own row-level security.

How do I keep a branded report up to date?

Treat the report as a view of the dashboard beneath it. Update the analysis on the canvas, by rearranging charts, adding new ones or continuing the conversation behind a chart, then regenerate the app so the branding is applied to the latest numbers in one step. Avoid hand-editing the generated page, because those edits are lost at the next regeneration and the report drifts away from its source data.

Is AI-generated HTML safe to embed in my product?

It can be, with the same precautions you would apply to any untrusted code. OWASP lists improper handling of LLM output, including generated JavaScript that causes cross-site scripting, as a top risk for LLM applications. Serve generated pages from an isolated origin or a sandboxed iframe, apply a strict Content Security Policy, and keep credentials and live query endpoints out of the page. Embedding snapshot data limits the damage if anything does go wrong.

Sources

1. Leviathan, Valevski, Kalman et al. (Google Research): Generative UI: LLMs are Effective UI Generators. arXiv. https://arxiv.org/abs/2604.09577

2. Google Research: Generative UI: A rich, custom, visual interactive user experience for any prompt. https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/

3. Sarikaya, Correll, Bartram, Tory and Fisher: What Do We Talk About When We Talk About Dashboards? IEEE Transactions on Visualization and Computer Graphics (InfoVis). https://tableau.com/sites/default/files/2023-01/DashboardsConspiracy_final.pdf

4. Segel and Heer: Narrative Visualization: Telling Stories with Data. IEEE Transactions on Visualization and Computer Graphics (InfoVis). https://homes.cs.washington.edu/~jheer/files/narrative.pdf

5. W3C Design Tokens Community Group: Design Tokens specification reaches first stable version. https://www.w3.org/community/design-tokens/2025/10/28/design-tokens-specification-reaches-first-stable-version/

6. W3C Web Accessibility Initiative: Understanding Success Criterion 1.4.3, Contrast (Minimum), WCAG 2.2. https://www.w3.org/WAI/WCAG22/Understanding/contrast-minimum.html

7. W3C Web Accessibility Initiative: Understanding Success Criterion 1.4.11, Non-text Contrast, WCAG 2.2. https://www.w3.org/WAI/WCAG22/Understanding/non-text-contrast.html

8. W3C Technical Architecture Group: Good Practices for Capability URLs. https://www.w3.org/2001/tag/doc/capability-urls/

9. OWASP Gen AI Security Project: LLM05 Improper Output Handling, OWASP Top 10 for LLM Applications. https://genai.owasp.org/llmrisk/llm052025-improper-output-handling/

10. MDN Web Docs: The iframe element. https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Elements/iframe

Turn any canvas into a branded app you can share. Start free at upsolve.ai with 2,000 free AI credits, or talk to our team about branded reports inside your product. Next in the series: Part 10: Claude MCP for Data Analysis.