Choose a Support Automation Operating Model
Separate Musts from Wants, compare only eligible alternatives and test the leading option against adverse consequences.
24 minutes
Scenario
A subscription software company must choose how to automate first-line support. The options are a vendor-managed agent, an internally configured agent on an approved platform, or an agent-assist-only design. The steering group has mixed mandatory controls with desirable benefits.
- Your role
- Operations lead preparing the decision recommendation
- Method
- Kepner-Tregoe Decision Analysis
Evidence pack
Privacy
The vendor-managed agent retains full prompts for 30 days; policy permits seven days unless a formal exception is approved.
Human review
All three options can route cancellations and refunds above EUR 100 to a human.
Speed
Estimated time to a controlled pilot is 4 weeks for vendor-managed, 9 weeks for internal agent and 3 weeks for agent assist.
Cost
Twelve-month cost estimates are EUR 180k, EUR 260k and EUR 120k respectively, each with a +/-20% confidence range.
Containment
Comparable trials suggest 45%, 55% and 15% autonomous containment respectively; transfer quality has not been tested on company data.
Auditability
The internal and assist designs expose decision traces to the compliance archive; the vendor option exports only outcome logs.
Team capacity
The internal design requires two engineers for nine weeks and there is only one currently available.
Reversibility
Agent assist can be disabled by queue; both autonomous options require a migration of conversation rules to reverse.
Constraints
- Privacy policy and review of consequential actions are mandatory requirements.
- Want weights must be agreed before alternatives are rated.
- The recommendation must include one review trigger and one dissent note.
Case steps
Work through each prompt using the evidence pack. Answers and rubric weights stay protected in the interactive flow.
Write the decision statement, owner, horizon and the three comparable alternatives.
Define evidence-based pass rules for privacy and consequential-action review, then test every alternative.
Define and weight four Wants, rate only eligible alternatives and record confidence beside each important rating.
Recommend one option after documenting two adverse consequences, mitigations, a review trigger and dissent.