Key Takeaways7
- We advise businesses on AI-agent readiness, so we ran the audit on our own site — it scored 32/100.
- The gap was the last stage, "do business with me": an agent could read us but couldn't complete an action.
- We built an MCP server that books a meeting on our real calendar and logs the lead to our CRM.
- We added the discovery layer: llms.txt, .well-known/mcp.json, agents.json, and ScheduleAction schema.
- The audit's action scores jumped (schema 60→100, programmatic 0→50, agent-access →100; overall 32→45).
- Honest limit: hand-connecting an MCP server is still more friction than a form — this is proof and future-proofing, not today's main booking channel.
- For most businesses, the fundamentals (accurate pages, a real bookable action, structured data) matter more than the agent endpoint.
We spend a lot of time helping businesses answer one question: when a customer sends an AI assistant to find and book a service, will it find yours — and can it actually complete the job? It felt only fair to point that question back at ourselves.
So we ran our own AI-agent readiness audit on bitesizeelephant.com. The score came back: 32 out of 100. Poor. An AI assistant could read us just fine — it simply couldn't do anything.
An agent could read us, but not act
That gap is the last — and most valuable — stage of agent readiness: “do business with me.” Our pages were clean and crawlable, we had an llms.txt file, and our structured data was solid, so the find-me and understand-me stages were strong. But when it came time to take an action, there was nothing an agent could complete. Our only “action” was a contact form — free text, behind a bot-check an assistant can't clear. No bookable slots. No way to finish the task.
What we built
We fixed the thing we tell clients to fix — the last mile. We built a small MCP server (Model Context Protocol, the emerging standard that lets an AI agent take actions on a service) so an assistant can actually book a meeting:
- It reads our real calendar availability — no double-booking.
- It books the slot, creates the calendar invite with a video link, and sends it.
- It logs the new lead to our CRM, right alongside form submissions.
Then we wired up the discovery layer so agents can find all of it: an llms.txt pointer, a /.well-known/mcp.json manifest, an agents.json file, and ScheduleAction structured data that declares “there's a booking you can take here.”
Did it move the needle?
We re-ran the audit. The scores that measure whether an agent can act jumped: action-schema went from 60 to 100, programmatic access from 0 to 50, agent-access to 100. The automated overall score rose from 32 to 45 — and the true number is higher, because the scanner has no way to “see” an MCP booking and under-counts it. By the manual standard the audit actually calls for, an agent can now complete a booking with us end to end. It couldn't before.
The honest part
Here's what dogfooding taught us that a slide deck wouldn't: today, almost nobody is going to hand-connect an MCP server just to book a call. For a human, that's more friction than filling out a form. So this isn't our main booking channel — it's proof that we build what we advise, and it's future-proofing for when assistants start doing this on their own, without the setup step. When that lands, the friction flips: the agent books you in the flow of conversation, and the businesses that are ready get chosen.
The other lesson: the flashy part — the MCP server — wasn't the high-value part. For most businesses, the fundamentals move the needle far more: accurate pages, a real bookable action, consistent listings, clean structured data. The agent endpoint is the last ten percent, not the first.
What this means for you
Agent readiness isn't a strategy you buy; it's a handful of concrete gaps you close. Start by finding out where an assistant gets stuck on your site — then fix the fundamentals before the frontier. If you'd like, we'll run that audit with you and turn it into a prioritized, practical plan.