Use case assessment
Is AI email triage worth building?
Mostly rules. Sender and subject carry more signal than the proposal usually credits.
Sender domain, subject patterns and recipient address carry most of the routing signal. That is a rules problem, and rules explain their decisions.
What it usually means
Sorting a shared inbox — sales, support, invoices, complaints — into queues or owners.
Is it genuinely AI?
Usually not. Look at what you already have on every message: sender domain, whether the sender is a known customer, which alias it arrived at, subject patterns, attachments.
Most shared inboxes route accurately on those alone. An invoice arrives as a PDF from a supplier domain to an accounts alias. That is not a semantic problem.
The case for a model appears when routing depends on what someone is asking rather than who they are — a general inbox where the same sender writes about five different things, and the distinction is in the body.
What the simpler version looks like
Separate aliases. Genuinely: the cheapest email triage system is a different address per purpose, published in the right places.
Then rules on sender and subject, with a default queue. Review the default queue for a fortnight. What sits there is your actual triage problem, and it is usually smaller and stranger than expected.
What it costs to run
Labelled examples per category, maintained as your business changes. Monitoring for drift as senders and terminology shift. Confidence thresholds and a fallback.
Weigh against the rules version, which costs a config change and is auditable when someone asks why an email went where it did.
Where it sits under the EU AI Act
Internal email routing is not an Annex III high-risk use.
If auto-replies are generated, Article 50 disclosure applies — a person receiving an AI-generated response should know, and that has been enforceable since 2 August 2026. Email content is also personal data, so processing location is a GDPR question first.
When it is worth building
When the default queue stays full after rules are tuned, when routing genuinely depends on message content, and when volume makes misrouting expensive.
Not when the problem is that nobody owns the inbox. Routing faster into a queue nobody works is not an improvement.
Common questions
Does email triage need AI?
Usually not. Sender domain, recipient alias, subject patterns and attachments carry most of the routing signal, and rules on those are auditable. A model helps when routing depends on what someone is asking rather than who they are.
What is the cheapest email triage improvement?
Separate aliases per purpose, published in the right places, plus rules on sender and subject with a default queue. Reviewing what lands in the default queue for a fortnight tells you whether you have a real triage problem.
Do AI-generated email replies need disclosure?
Article 50 requires disclosure where a person interacts with an AI system, enforceable since 2 August 2026. A recipient of an AI-generated response should know that is what it is.
Cutting enterprise change requests 15 to 7 minutes — where the scope cut, not the model, produced the result.