Read summarized version with
About project
Working time:
Since 2025
Industry:
Software
The service:
Generative AI
Overview
The customer is a North American B2B software company whose support team serves thousands of business users. Dedicatted built an AI support assistant for its help centre and in-product chat on Amazon Bedrock, then took over its operation.
The Challenge
The assistant was accurate on launch day. Within weeks, product releases and pricing changes began to outpace its knowledge.
1. Knowledge Drift
New features and pricing changes were not reflected in answers.
2. Mis-Routed Hand-Offs
Escalations went to the wrong queue, and customers repeated themselves to agents.
3. Nobody Reading Failures
Failed conversations were logged but never reviewed.
4. The Build Team Moved On
The engineers who built the assistant had moved to other projects.
Technology & Approach
We run the assistant as a product with a monthly operating cycle, and every change is tested before customers see it.
Implementation
Operations Onboarding
Baselined containment, escalation and answer quality, built an evaluation set from 500 real conversations, and mapped each knowledge source to an owner.
Knowledge and Quality
Knowledge is updated with each product release, samples of failed and escalated conversations are reviewed weekly, and prompts and Amazon Bedrock Guardrails are tuned.
Escalation and Personalization
Escalation rules route hand-offs with a conversation summary, and answers use the customer’s plan and account status through governed tools.
Evaluated Releases and Reporting
Every change runs through the evaluation set before release, with a monthly report on containment, escalations, quality findings and cost per conversation.
Customization & Client Adaptation
Operations were tied to the company’s own release cycle and support process.
Knowledge Synced to Releases
Product release notes trigger a knowledge update and a re-run of the affected test questions.
Hand-Offs with Context
Agents receive a summary of the conversation, so customers never repeat themselves.
Answers That Fit the Customer
The assistant answers about the features in the customer’s own plan, not the whole product.
Cost per Conversation Tracked
Model and retrieval costs are reported per conversation each month.
If you find this case interesting, we recommend taking a closer look at
The Results
The assistant now resolves a growing share of conversations without a hand-off, escalations reach the right queue with context, and every change passes the evaluation set before release.
Long-Term Value
- Knowledge in step with every product release.
- An evaluation set that grows from real failures.
- Monthly visibility of containment and cost.
Business Impact
- Support agents focus on complex cases.
- Customers get answers that match their plan.
- No quality drop when the product changes.
