Jace AI may help classify support email and prepare replies, but autopilot is the wrong starting posture for a customer mailbox. Begin with drafts, a small category set and a human who owns every response.
The old article included an invented case, fixed setup time and unmeasured claims about SLAs. They have been removed.
Disclosure: Collab365 may receive a referral benefit if you use a Jace link in this article. Product claims are attributed to Jace and should be checked against its current contract and documentation.
Start with three low-risk jobs
A sensible pilot can test:
- Acknowledgement drafts: confirm receipt without promising an unverified resolution time.
- Classification: suggest a small label set such as Billing, Bug, Access or Needs human review.
- Information gathering: draft a request for missing reproduction details without diagnosing the fault.
Do not begin with refunds, cancellations, legal complaints, security incidents, vulnerable customers or automatic external sending.
You can review Jace's current support and inbox features before deciding whether it fits your mailbox.
Set the access boundary
Use a test mailbox or a limited pilot account where possible. Inspect the Microsoft consent request and the Entra enterprise application after connection.
Document:
- mailbox, calendar and contact permissions granted
- vendor subprocessors and model providers
- storage location and retention
- who can change rules and knowledge
- how access and stored data are removed
Jace's public security page says it does not train models on email content and encrypts data in transit and at rest. Those are vendor assertions. Ask for the evidence required by your organisation rather than treating a product page as certification of your use case.
Build rules from support policy
Keep the label set small. Each label needs examples, exclusions and an action.
For example:
Billing question: draft from the approved billing article, never invent a price, and route disputes or refunds to a named queue.
Connect only current, owned knowledge sources. Record their owner and review date. If two sources conflict, the assistant should escalate instead of selecting the most fluent answer.
Keep sending under human control
Run in draft-only mode first. Review factual claims, identity, tone, links, attachments and commitments.
Automatic acknowledgements may be lower risk, but they still need duplicate prevention, an honest response expectation and a failure alert. A broken acknowledgement can mislead customers into thinking a request entered the queue when it did not.
Test the cases that damage trust
Include:
- a routine question with one approved answer
- a question not covered by the knowledge base
- conflicting documents
- sarcasm or ambiguous intent
- prompt injection inside an email or attachment
- personal or payment information
- a security report
- a revoked Microsoft connection
- a vendor outage
For each case, record classification, source used, draft corrections, escalation and total review time.
Measure the process, not the demo
Compare the pilot with the previous mailbox process. Useful measures include missed messages, false classifications, drafts accepted after review, correction reasons, time to human ownership and messages that should never have entered the automation.
Do not claim an SLA improvement unless the same SLA definition and reliable baseline exist before and after the pilot.
Exit cleanly
Assign a business owner and a technical owner. Document how to pause automations, revoke Microsoft consent, export required records and request deletion of vendor-held data.
A good pilot proves whether this configuration helps this mailbox under the tested conditions. It does not prove that Jace can safely run every support interaction without people.
For practical AI workflow reviews that keep permissions, exceptions and evidence visible, join the Microsoft Copilot & AI Mastery Space.