HAAM Signal Agent
Turn website friction into an evidence-backed proposal.
The HAAM Signal Agent discovers important pages, collects measurable accessibility, performance, technical, and journey evidence, then turns the strongest findings into a proposal that a human reviews before it leaves the studio.
The problem
Most automated audits generate noise. Most sales outreach generates distrust.
A useful diagnostic system has to connect a visible issue to evidence, business relevance, a practical fix, and an honest scope of work. The Signal Agent is designed around that chain.
Evidence before interpretation
Deterministic tools record what happened before an AI model is allowed to explain why it may matter.
Priority before volume
The final scan emphasizes a few high-confidence issues instead of forwarding every warning produced by every scanner.
Review before outreach
Findings can be approved, kept pending, edited, or dismissed before the proposal and message are regenerated.
How it works
A controlled pipeline, not an autonomous consultant
The system separates collection, interpretation, and approval so each claim has a clear source and owner.
- 01
Discover
The agent selects up to five high-value public pages from navigation, links, titles, and the likely conversion journey.
- 02
Collect evidence
Playwright, axe-core, Lighthouse, screenshots, network logs, and deterministic checks record what actually happened.
- 03
Interpret carefully
Optional AI explains grounded findings, but it cannot create a claim without a matching page, quotation, screenshot, selector, or tool result.
- 04
Review and propose
A human approves, keeps pending, edits, or dismisses each finding before the system creates a proposal and outreach draft.
What it reviews
Six signal areas across the website journey
The scan combines objective checks with carefully constrained interpretation. Not every category produces a finding, and uncertain observations stay out of the final proposal.
Accessibility
Checks accessible names, form labels, headings, keyboard-sensitive controls, target sizes, overflow, and axe-core rule violations. Automated results are treated as evidence, not proof of full WCAG or EAA compliance.
Performance
Runs Lighthouse on the starting page and surfaces slow loading, blocking work, oversized resources, and other measurable performance signals that can weaken the user journey.
Technical quality
Finds broken links, failed requests, console errors, missing metadata, weak heading structure, and other implementation issues that are easy to verify and expensive to ignore.
Journey friction
Reviews important paths such as booking, checkout, contact, signup, and application flows to identify dead ends, unclear actions, and moments where users may be unable to continue.
Clarity and trust
Looks for grounded signs that the offer, audience, next step, pricing, limitations, or proof are difficult to understand. AI-assisted findings must point back to captured page evidence.
Localization context
Flags places where language, regional platform habits, trust cues, content hierarchy, or market-specific expectations may need a closer human review.
Evidence model
Every finding needs a trail
A finding can carry the affected URL, screenshot, selector, scanner result, observed behavior, confidence, severity, expected consequence, recommended action, and whether manual review is still required.
Example finding
Mobile navigation control has no accessible name
Evidence
- Tool
- axe-core
- Element
- button.mobile-menu
- Observed behavior
- The button is announced without a name.
Recommended action
Add a stable accessible name, expose the expanded state, and verify the interaction with keyboard and screen-reader testing.
This is an illustrative finding. A real proposal only includes evidence captured from the reviewed website.
What you receive
A smaller report that is ready to discuss
The output is designed to start a credible working conversation, not to impress someone with the length of a PDF.
- Three to five prioritized findings instead of a wall of generic warnings
- Desktop and mobile screenshots tied to the affected page
- Evidence, confidence, severity, business implication, and recommended fix
- A reviewable HAAM proposal matched to the strongest problems
- A concise outreach draft that can be edited before it is sent
- Structured scan data that can later support benchmarks and case studies
Trust boundaries
What the agent deliberately does not do
The product is useful because it limits its authority. Automation can collect and organize evidence, but responsibility for the claim and recommendation stays visible.
No automatic compliance claims
Automated checks cannot establish complete WCAG, EAA, privacy, or legal compliance. Manual assessment remains part of responsible delivery.
No invented business impact
The system does not claim an exact conversion loss, revenue increase, or implementation cost unless real client data supports it.
No automatic outreach
A false or exaggerated finding can damage trust. Every external claim and proposal remains reviewable before anyone receives it.
No unauthorized scanning
The tool is intended for websites you own, manage, or have permission to inspect. Private network targets are blocked by default.
Who it helps
Use it when the website feels wrong, but the backlog is still vague
A Signal Scan is most useful before committing to a large redesign or optimization programme. It helps the team decide where deeper work is justified.
Founders and product teams
Use a focused scan before a redesign, launch, fundraising push, accessibility review, or conversion sprint.
Marketing and ecommerce teams
Turn visible journey friction into a smaller, more credible backlog tied to acquisition, trust, and task completion.
Agencies and technical partners
Create an evidence layer before scoping design, development, remediation, or optimization work with a client.
Technology
Built from browser evidence and structured review
The MVP uses Next.js, Playwright, axe-core, Lighthouse, TypeScript, Zod, and optional OpenAI Structured Outputs. It can run without an AI key and is designed to move to isolated workers, persistent storage, accounts, quotas, and CRM integrations as usage grows.
Questions
How to use the Signal Agent responsibly
Is this a fully automatic website audit?
No. It automates evidence collection and proposal preparation, then keeps judgment with a human reviewer. That separation is the point of the product.
Does it work without an AI model?
Yes. Deterministic checks, screenshots, scoring, and proposal generation can run without an API key. AI adds grounded interpretation and improved proposal wording when enabled.
How many pages does a scan review?
The MVP reviews up to five high-value pages so it can stay focused on the main customer journey rather than producing an unmanageable crawl report.
Can it replace accessibility testing or user research?
No. It is useful for triage and evidence gathering. Manual accessibility assessment, assistive-technology testing, analytics, interviews, and usability research answer questions that automation cannot.
Next step
Start with one important journey
Share the website, the customer action that matters most, and what feels uncertain. HAAM can use the scan to recommend a focused audit, optimization sprint, accessibility review, or implementation scope.