Clarivise Shield — AI email security for Microsoft 365

The problem with blocking

Most email security products sit in front of the mailbox. They hold, quarantine, or reject — which means the day they get it wrong, someone’s contract renewal is sitting in a quarantine folder nobody checks and the business finds out three days later.

For small and mid-sized organizations that is a genuinely bad trade. The cost of a missed legitimate email is immediate and visible; the cost of a phishing email that slips through is probabilistic. IT teams end up tuning filters conservatively, which is the same as not having them.

Shield takes the opposite position: it cannot lose mail, because it never touches delivery.

How it works

Internet
  → Microsoft 365 EOP (standard spam filter)
  → Mailbox (delivered normally)
  → Shield connector polls new mail every 60s via Graph
  → short/wrapped URLs resolved locally (no token cost)
  → analysis pipeline → verdict
  → subject tag + scan-report banner written back into the message
  → result logged; flagged mail added to the admin review queue

Mail arrives the way it always did. Within a minute, it is annotated in place.

The verdict model

Four tiers, each with exactly one defined action — no configuration matrix, no tuning weekend:

Verdict	What happens
Safe	Scan-report banner only
Spam	[SPAM] subject prefix + banner
Suspicious	[SUSPICIOUS] subject prefix + banner
Phishing	[PHISHING] prefix + warning banner + added to review queue

The banner explains the verdict in plain language — which link pointed somewhere other than where it claimed, why the sender domain looked wrong, what the requested action was. That matters more than the score. A user who understands why a message was flagged gets better at spotting the next one; a user who just sees a red bar learns nothing.

Engineering decisions worth calling out

URL resolution happens before the model, not inside it. Shortened and wrapped links (Safelinks, URL-defense products, bit.ly and friends) are unwrapped locally first. The model reasons about real destinations rather than opaque redirectors, and no tokens are spent on string manipulation a hundred lines of Python does better.

The review queue is honestly named. It is a record of flagged messages for admin review, not a quarantine — because in the marking model nothing is being held. “Release” and “Delete” update review status and can optionally remove the tagged copy through Graph. Pre-inbox quarantine would need Exchange transport rules, and that is deliberately out of scope: it would reintroduce exactly the failure mode the product exists to avoid.

The connector is a container, not a function. Polling on a fixed interval with per-tenant state and Graph throttling to respect is a long-running workload, so it runs as a FastAPI service on Azure Container Apps rather than being forced into a serverless timeout budget. Scale-to-zero keeps the cost profile close to serverless anyway.

API keys never leave the server. Analysis runs in an edge function that holds the model credential. Nothing in the mail path — and nothing in the admin dashboard — can read it.

What ships with it

An admin dashboard covering tenant statistics, the review queue, allow and block lists, and settings; a daily summary function; and a set of PowerShell scripts that stand up the Microsoft 365 app registration and verify permissions before anything else is deployed. The setup step is the part most products handwave, and it is where most failed rollouts stall.

Where it stands

Shield is an Ingot Solutions product, deployed and running. The same architecture — poll, enrich, reason, annotate, log — transfers cleanly to any “read a stream of business documents and mark them up” problem, which is most of the AI automation work worth doing.