Strong evidence (grade B: first-hand account)Detailed first-hand article by the company's Head of AI with specific metrics and what broke; self-reported.

Dayuse's Two-Person AI Team Handles 1 in 3 Support Tickets and Most Answered US Calls

Dayuse says 1 support ticket in 3 is now handled with no human, and in September 2026 AI alone handled 70% of answered US calls. A two-person AI team built it over eight months on top of Zendesk, HubSpot and Notion.

Who this is about: Day-use hotel booking marketplace (Dayuse, 120 people)

ZendeskHubSpotNotionIn-house MCP serversStartupsAutopilot

Get the prompt

SupportAdded Oct 16, 2026Source: x.com

What it takes

Cost: Cost not statedModels aren't named; the company says model costs are a small part of the overall investment.
Setup time: A few weeksOur estimate from the source, for one business.
Skill needed: Developer skillsYou build or run your own code, even if AI writes it.

Explain it simply

Dayuse, a 120-person company that sells daytime hotel stays, kept its existing support, sales and documentation apps. Its small AI team built its own layer on top that sorts tickets, drafts replies and takes actions in the back office within set permissions. New actions need a person's approval until they are proven safe.

New to AI? See how AI works

What was done

A two-person AI & Innovation team kept Zendesk (support), HubSpot (sales) and Notion (docs) and built its own layer on top: classification, back-office actions, drafts, workflow automation and account prioritisation.

Each business action became a tool with explicit permissions, exposed once over MCP so the website agent, email and phone support, and a ChatGPT app can all use it.

Every morning an agent reviews what the automations sent, team feedback and unanswered requests, then adjusts.

Writes to shared systems need a human's approval until a case is proven eligible for automation, and every call is logged.

How it went

The company reports 1 support ticket in 3 handled with no human, 70% of answered US calls handled by AI alone in September 2026, over a thousand tickets a month handled in under 30 seconds, and unanswered US calls down 4x.

Not everything worked first time: the first search tool shipped in a degraded version, and the voice agent took many iterations before it spoke to customers.
Business teams have also built 26 apps and agents now running in production.
Company self-report by its Head of AI; the numbers weren't independently checked.

Prompt

PromptWritten by our editors from the source

You classify incoming support tickets for [company].

For each ticket: pick one category from [category list], say whether it matches a documented procedure in [docs], and draft a reply using only that procedure.
If the procedure needs a back-office action, name the action and the fields, but don't run it until a person approves.
If no procedure fits, route the ticket to [team] with a one-line summary.
Never promise refunds or exceptions that aren't in the procedure.

Setup notes

Pick one painful, repetitive flow and write down its rules before automating anything.
Turn each business action into a tool with explicit permissions, and keep a human approval on writes until the case is proven.
Measure capacity and cost, not usage.

Business fit

Who this helps, and how:

Booking and marketplace companies: answer routine booking questions and calls without adding support staff as volume grows.
Startups
Hotels and hospitality groups: take routine calls with AI and keep people for the cases that need them.
Local & service

Similar tools

Other tools for this job, and how they differ. Chosen by our editors, never paid for.

Gorgias
Off-the-shelf ecommerce helpdesk whose AI Agent answers chat, email and SMS; phone volume is not automated.
Grok Bot
A managed agent that can work a support queue on a schedule and ask before refunds, without building your own layer.
See it in use: Run customer support with a bot that asks before refunding

“Strong evidence” (grade B) means: The person who did the work describes the setup and results in detail. How we grade. Spotted a mistake? Suggest a correction.