Case study 03 · KaDeep · B2B SaaS · Developer tools · Compliance
KaDeep Accessibility — What do I fix first?
Teams do not need another 300-row dump. They need three things, in order, with time estimates.
- Role
- Product Manager / Accessibility Lead — I owned problem definition, ICP and jobs-to-be-done, product narrative, MVP scope, scoring and prioritization logic, compliance framing (WCAG / ADA / Section 508 / EN 301 549), packaging, and go-to-market with engineering.
- Scope
- Problem framing through packaging — scan, score, plan, AI fix, re-scan
90-second brief
How to read this
Teams run a scan, get a 300-row dump, and close the tab. Legal and procurement want evidence, not intentions. Detection was not the gap. The first move was.
Stamp compliant / not compliant
Died
A legal claim we cannot stand behind. Buyers asked for it. I almost shipped it. That would have been the product-killing decision.
Another scanner / raw axe dump
Died
Same freeze. Fast signal, no site picture, no legal lens, no backlog. We would be a bookmark, not a habit.
Translation layer — never stamp
Shipped
Keep axe-core as the engine of record. Invent the score, the matrices, the P1/P2 plan, and the PDF. This is what shipped.
- Decision
- axe-core for trust. KaDeep Extended for WCAG 2.2. Product language is this scan’s pass/fail — not a certificate. AI writes patches. Humans review.
- Result I will defend
- xactload.com: 89/100 Good, 0 Critical / 0 Serious, one heading-order fix (~10 min), 14-page PDF. kadeep.ai: 49/100, 149 violations, 9 pages, 55 components — and a two-sprint score plan, labeled a projection.
- What I got wrong
- I almost shipped the stamp. I also let “49 → 100 in two sprints” sound like a promise. It is a score plan, not a certificate.
- Who loses if this is wrong
- Legal and procurement, if we imply a certificate. The engineer, if the plan is still a dump. The user who needs the site, if the team closes the tab again.
Metric spine
- North star
- Time-to-first-actionable-plan
- URL in → a P1 list with hours and projected score, under two minutes. If we lose to Lighthouse on time-to-insight, we lose the wedge.
- Guardrail
- Overclaim
- “Compliant” vs “this scan’s pass/fail against WCAG 2.1 AA.” Same violations. Completely different legal exposure.
- Not instrumented
- Free → Pro conversion
- Early-stage. I will not invent a rate. The packaging hypothesis is: free answers risk, paid sells the plan and the PDF.
- Not instrumented
- Re-scan within 14 days
- Would prove the loop is a habit. Not measured on a published cohort yet.
- Not instrumented
- Hours saved vs consultant
- xactload.com is a score snapshot, not a before/after time study. Next measurement, not a claim.
Keep axe-core as the engine of record. Invent the translation layer. Never stamp “compliant.”
The tempting product is an AI auditor that replaces consultants. That is a long, credibility-destroying road. Detection was already solved. I scoped a wedge: public-site automated audit, score and matrices a VP and a lawyer can both read, P1/P2 a PM can sequence, patches an engineer can review. “Compliant” vs “this scan’s pass/fail against WCAG 2.1 AA” is the difference between a useful tool and a liability.
Scan
Real-browser crawl. 396 WCAG 2.2 checks.
Score
0–100, weighted. Site map, not one page.
Plan
P1/P2 with hours and projected score.
Fix + re-scan
AI patch, human review, prove the delta.