Agent Flood Filter: a refund experiment shows what is coming to small-business inboxes

The idea is real but early, and so far one serious company is going after it. Lorikeet is the only company I found that sells detection of AI-agent senders in support queues. Its feature is a beta in observation mode, it is tied to Lorikeet's own help-desk platform, and the tools that would act on the traffic are still being built. Gmail users and Shopify stores on Gorgias or plain email have no lightweight add-on that flags agent-written messages, collapses repeat claims from one sender into one thread, and holds follow-ups past a limit. A new product could still win there by being cheap, installing in minutes, and working with any help desk. There are three risks. First, the size of the problem is unproven: the only figure is Lorikeet's vendor claim of up to 3%, and I found no reporting of small merchants being swamped by Pine- or DoNotPay-style agents. Second, Zendesk, Gorgias and Intercom already auto-tag tickets and could ship an agent flag as a feature. Third, detecting AI-written text is unreliable. Regulators such as the Western Australian Ombudsman also caution that using AI does not make a complaint less valid, so throttling replies risks consumer-protection trouble and false positives on real customers. The merge-and-cap logic is safer to build on than the detection score. I found no existing standard or product for the agents.txt idea, a hosted complaints policy page written for consumer agents. The nearest things are general conventions like llms.txt and AGENTS.md, plus scattered experiments. That part looks empty. Its value depends on consumer agents such as Pine choosing to read and follow the page, so it needs partnerships on the agent side and probably works better as a free feature that draws people to the paid inbox filter than as a product on its own.

• Lorikeet (agent traffic detection beta)

  This is the closest direct match. Lorikeet is an AI customer support vendor. It launched a beta that detects and tags support conversations coming from AI agents, as part of a B2A suite it started in September 2026. The beta runs in observation mode first. Lorikeet says it is now building ways to act on that traffic, such as putting humans first in the queue, applying stricter identity checks to agents, and sending agents to an agent-facing concierge. Lorikeet claims up to 3% of inbound support traffic comes from agents. It sells to larger support teams, not to Gmail-based small businesses.

• Lorikeet guide: signs AI agents are contacting your support team

  This Lorikeet content lists detection signals: regular timing, identical phrasing, bursts of messages, and the documented user agents of OpenAI, Anthropic and Perplexity. It shows the incumbent is already teaching this problem to the market. A webinar called 'When the customer is an AI agent' is scheduled for October 21, 2026.

• Winston AI (AI text detection applied to complaints)

  Winston AI is a general AI-text detector. It has been used in research on CFPB complaints to flag complaints written with ChatGPT. It works as a detection building block, not as an inbox product, and it does not merge or throttle messages.

• Zendesk Intelligent Triage / Intercom Fin AI tagging

  These help desks already tag tickets automatically by intent and sentiment, and they could add an 'agent-written' tag fairly easily. I found nothing showing either one detects agent senders or caps follow-ups today.

• Pine AI (demand-side consumer agent)

  Pine is not a competitor. It is a consumer agent that files complaints, disputes charges and negotiates bills by phone and email, and it raised a $25M Series A. It creates the traffic this product would filter, which shows the problem is real.

• DoNotPay (demand-side consumer agent)

  DoNotPay is also not a competitor. Customers use it to challenge transactions, ask for refunds and demand data removal. Merchant teams already run into it, but I found no reporting that these tools are flooding merchants.

Checked 2026-10-18