AI agents are becoming a real buyer persona. When a developer asks Claude or ChatGPT to "set up invoicing," the agent — not the human — decides which SaaS to reach for. So we asked a simple question: if an AI agent showed up at your SaaS's front door today, could it get in?
We spent this summer verifying that question against 200 of Japan's leading SaaS products, using only publicly checkable facts. The result is the ARI Award 2026 Summer — and the headline is that only 41 of 200 (20.5%) earned an A rating or above. Eleven earned AAA.
This post explains what we measured, how the rating works, and what the 159 that didn't pass have in common.
Who we are (and why "rating agency")
KanseiLink is an independent rating agency operated by Synapse Arrows (Singapore). We verify Agent Readiness: how ready companies and SaaS products are to be discovered, understood, connected to, and executed by AI agents.
"Rating agency" is a deliberate framing. Gartner ranks vendors for CIOs. G2 aggregates human reviews. ESG ratings serve investors. Nobody was rating software for its newest class of users — AI agents — so we built the evaluation model for it:
- AI Access — is there an API/MCP an agent can actually connect to?
- AI Discoverability — can an agent find and read your docs? (llms.txt, robots.txt policy, structured data)
- AI Execution — can it complete tasks end-to-end?
- AI Trust — auth clarity, scopes, safety of delegation
- AI Compatibility — does it behave consistently across models?
Ratings run AAA to D. The formula is public, ratings are never for sale, and per-service success-rate figures stay unpublished until independent third-party telemetry reaches sufficient scale (Principles of Independence).
The evidence base
Numbers as of 2026-07:
- 11,000+ SaaS/MCP services cataloged and rated on publicly verifiable facts
- 2,257 services probe-tested by our own scanner with real JSON-RPC handshakes — 3,475 successful handshakes out of 4,367 attempts
- 200 Japanese SaaS manually verified against vendor primary sources for the ARI Award
"Publicly verifiable" is the key phrase. We check what an agent can check: official MCP server availability, API openness, whether the developer docs are reachable without emailing sales, whether auth methods are documented. No vendor self-reporting, no pay-for-grade.
What separates the 41 from the 159
The certified tier (AAA 11 / AA 21 / A 9) shares a profile:
- An official MCP server, or a clean, fully public API. The AAA tier is dominated by vendors that ship their own MCP server and treat it as a first-class product.
- Docs an agent can read. No login walls, no "request the PDF" forms, no robots.txt blanket-blocking AI crawlers on the documentation domain.
- Auth that's explainable in one page. OAuth flows with clear scopes beat bespoke token ceremonies every time.
The uncertified majority mostly fails on quiet things, not dramatic ones:
- Developer docs that return 403/404 to any non-browser user agent
- APIs whose reference is only available after a sales contact — a human can email; an agent can't
- A curious pattern we keep seeing: vendors that AI models happily recommend, but that agents cannot actually *reach* — the model's training data knows the brand, but there's no public path from "recommended" to "connected." Recommendation and connectability are different axes, which is exactly why we measure them as separate domains.
Why Japan-first
Japan's SaaS market is large, dense, and under-documented in English — which makes it a blind spot for most AI models. It's also moving fast: several major Japanese vendors (accounting, chat, workflow) now ship official MCP servers that rival anything in the US ecosystem. The ARI Award exists partly to make that legible to the global agent ecosystem.
A global index is the obvious next step; the 11,000+ service database already spans both markets.
Check your own product
- Instant self-check: AEO Score Checker (free)
- Full ranking + methodology: ARI Award 2026 Summer
- 90-day improvement playbook: Agent Readiness Improvement Roadmap
If you're building agents rather than SaaS: the same database is queryable over MCP (github.com/kansei-link/kansei-mcp-server, MIT) so your agent can look up whether a service is worth attempting before burning tokens on trial-and-error.
Top comments (3)
That pattern of "vendors that AI models happily recommend, but that agents cannot actually reach" matches what shows up outside Japan too, and I suspect part of it is not a vendor decision at all. A lot of 403-to-non-browser-UA docs sites are running default CDN bot protection, Cloudflare or Akamai presets the vendor never revisited after setup. Does the ARI methodology distinguish deliberate blocking from inherited infra defaults? It matters for the fix path: a vendor who chose to gate their docs needs convincing, a vendor whose docs are accidentally walled needs one support ticket. If the ARI data can label which of the 159 fall into that second bucket, you would be publishing the cheapest 20 points of agent readiness available in the whole market.
Great point. Today, ARI primarily measures the observable outcome from the agent side, so it does not yet systematically distinguish intentional blocking from inherited CDN/WAF defaults.
But I agree that the distinction is important — especially because the remediation path is completely different.
The idea of identifying vendors that are effectively “accidentally walled off” is particularly interesting. In some cases, the biggest readiness improvement may come from a very small infrastructure/configuration change rather than a product change.
That’s a useful direction for us to explore. Thanks for the thoughtful comment.
The low pass rate makes sense when you consider most SaaS documentation is still optimized for human consumption rather than machine parsing. Agents struggle with complex OAuth flows and ambiguous error messages that lack structured schemas. If we want true agent readiness, we need to treat API tool definitions and error states as first-class citizens, almost like UI components. Have you noticed if products using standardized protocols like MCP naturally score higher on these readiness metrics?