One inbox. Thousands of emails. A very small team.
Computers 2 Kids refurbishes donated computers and gets them into the hands of low-income families across the country. Their lifeline to the community is a single shared mailbox, and it was drowning them.
When we audited the inbox, the pattern was stark. The majority of daily volume wasn't customer email at all: system notifications, order confirmations, voicemail alerts, newsletters, and internal mail buried the real inquiries. The real inquiries were highly repetitive, how do I apply, what does it cost, when can I pick up, how do I activate Office, but each one still needed a warm, accurate, personal answer.
Every answer had to be right. Prices, hours, program rules: a wrong answer to a low-income family isn't a typo, it's a broken promise. And some topics, like refunds, damaged devices, and legal requests, should never be answered by a machine. Off-the-shelf auto-responders were a non-starter.
The AI recommends. The safety engine decides.
We designed the agent around one uncompromising principle: the language model never has the last word. Every email flows through a pipeline where a deterministic, fully-tested safety engine, not the AI, makes the final routing decision.
A pre-filter silently sets aside vendor mail, system notices, voicemails, and marketing before any AI is invoked. A language model then reads each genuine inquiry, identifies intent, detects language, and recommends a response, as a suggestion only. The safety engine holds veto power: refunds, damage claims, legal requests, anything with personal risk. Those are hard-routed to a human, always.
For safe topics, the AI writes a natural, personalized reply, but it can only state facts from a closed, human-approved knowledge base. Every URL, price, and phone number is verified against that set, then a second AI acts as an independent judge before anything is queued. Any doubt falls back to a human-approved template. Nothing is ever sent automatically: every reply lands as a ready-to-review draft with the customer's original message embedded below it.
Smart pre-triage
Heuristics and a client-curated sender list silently filter vendor mail, confirmations, and newsletters before any AI runs.
Intent classification
Identifies what the customer actually needs, including multiple intents in a single email, and flags language for bilingual drafting.
Deterministic safety engine
A rules engine with 100% branch test coverage holds veto power. Sensitive topics are hard-routed to humans, no exceptions.
Fact-grounded AI writing
Drafts are written naturally and personally, but every fact is verified against an approved knowledge base. A second AI acts as independent judge before anything is queued.
Full-thread awareness
Tracks whole conversations, reopens threads when customers write back, and marks conversations complete when the customer has been helped.
Shadow mode deployment
Zero sends, ever. Every reply lands as a ready-to-review draft: staff review, adjust, and click send.
