We Built Our Own AI Support Agent. Then We Fact-Checked Everyone Selling One.

Stork gets support email like everyone else: tool submissions, listing questions, the occasional angry crawler complaint. We built an agent for it instead of buying one — not because the market is empty, but because nothing in it is shaped like "a founder's own-brand support agent that knows your catalog and routes people into your actual funnels." Ours drafts replies for human review, retrieves from our own catalog and blog before it says anything, refuses to answer account questions over email, and carries hard circuit breakers so it can never get into an argument with an autoresponder.

Building those guardrails meant researching how the commercial agents handle the same problems. That research turned into this audit. Everything below comes from vendors' own pricing pages, docs, and engineering blogs — not review-site hearsay — checked on July 23, 2026. Where a vendor publishes nothing, we say so instead of guessing.

"AI email tool" is four different product categories wearing one label, and they don't substitute for each other. Buying across the boundary is the most common mistake we see — and, honestly, the mistake our own directory's auto-generated competitor lists used to make before we started curating this category by hand.

Intercom Fin charges $0.99 per "outcome" with a minimum monthly commitment when run standalone. Read their definition: an outcome counts when a customer confirms resolution, when Fin completes a workflow — or when the customer simply doesn't ask for more help after Fin responds. Silence is billable. A customer who gave up counts the same as a customer who got their answer.

Zendesk sells the same unit priced honestly: an automated resolution counts only after the requester stays quiet for 72 hours, no human touched the ticket, and a language model verifies the answer was actually relevant. Same metric, stricter referee. If you're evaluating per-resolution pricing, the definition of "resolution" matters more than the price.

The marketing runs from one pole to the other. Gorgias promises automated returns and order tracking "all without a human in the loop". Freshworks pitches Freddy resolving "up to 80% of customer queries without human intervention". At the opposite pole, Front leads its pricing page with "human-in-control AI" — betting that buyers are now scared of exactly what the first group is selling.

The honest version usually lives one click deeper, in the docs. Gorgias's own playbook recommends starting with AI replies as internal notes for human review before letting it send. Zendesk gates auto-replies behind a numeric confidence threshold — default 60 of 100 — below which the agent escalates. Fin hands off automatically on self-harm, minors, jailbreak attempts, and high-risk medical, legal, or financial topics, no configuration asked. Even Instantly, a cold-outreach tool with every incentive toward volume, ships its reply agent with an explicit "Human-In-The-Loop mode… then switch to Autopilot" recommendation.

In other words: everyone selling autonomy operates draft-first internally. The gap between the landing page and the settings panel is the most useful thing we learned reading these docs.

Two public incidents explain the caution. In February 2024 a Canadian tribunal made Air Canada honor a bereavement-fare policy its chatbot invented, rejecting the airline's argument that the bot was "a separate legal entity responsible for its own actions." In April 2025, Cursor's AI support agent invented a "one device per subscription" policy to explain a login bug — different users got different fabricated answers, and paying customers cancelled over a policy that never existed.

Note what failed in both cases. Not loops, not deliverability — the model confabulated policy to cover a knowledge gap, in the confident register of a support agent. Your agent's statements are legally your statements. That's the risk every confidence threshold, holdout category, and validation pass in this market is really pricing in.

Our agent is deliberately boring. It drafts; a human sends. It retrieves from our own catalog and published posts before answering, and if retrieval comes back empty it says less, not more. Account-specific questions never get answered in-channel — the reply is a link into the authenticated dashboard, because a From: header is not identity. And it carries deterministic circuit breakers: auto-reply detection on inbound, hard caps per thread, so it can never loop against an out-of-office.

The mechanics that matter most are the ones no vendor markets. RFC 3834 — the 2004 spec for automated email responders — says auto-replies must stamp themselves Auto-Submitted: auto-replied, must never answer a message that carries that header, and must never respond to a null return-path. Among every vendor we audited, only Zendesk documents its mail-loop handling in any depth — including a rate breaker that suspends a sender after 20 emails in an hour, and the frank admission that no system prevents all loops. Everyone else's docs are silent on the single oldest failure mode of automated email.

• If you're buying: match the category shape first, then ask four questions the sales deck won't answer. What exactly counts as a billable resolution? What categories can never be auto-answered? How does the agent verify who it's talking to before disclosing account data? And does it stamp RFC 3834 headers on its own outbound mail?
• If you're building: draft-first isn't a training-wheels phase, it's the industry's actual operating posture. Add an intent allowlist before any auto-send, hard holdouts for money and legal topics, and a validation pass that blocks any reply stating a price or policy the retrieval layer can't literally show.
• Either way: measure resolutions the strict way — no human touched it, the requester stayed quiet for 72 hours, and a spot-check says the answer was actually right. The generous definition will flatter your dashboard and hide your failures.

We'll keep the listings in this category fact-checked against primary sources — it's the same pipeline that keeps our own agent's answers grounded. If you ship an AI email tool and your pricing or docs change, claim your listing and tell us; corrections beat speculation.