First-to-Market Agentforce in DACH

The Brief

Client IT team wanted an externally facing service agent with low latency output for fast response times and high case deflection rate to reduce the workload of the inhouse Service team. On the other hand, as AI is still a rather new territory, with many horror stories. Due to this reason, DPO team wanted to avoid having any data leak and strengthen guard rails.

Constraints

Agentforce already having certain guardrails which are non-editable for the developers yet alone, they are also not sufficient as agent tends to answer all the question users prompted in. Data segregation between externally available vs internal only and security measures to control visibility of the externally available data. Delay in the responses due to high volume data and multi step data processing, in addition to that - external system involved external queries expectation. Salesforce Vector search accuracy.

Options Considered

1. Treat the agent like a chatbot and restrict the use cases/capabilities available, especially the actions within the platform
2. Strict Id&V process supported by extra validation layers for certain use cases that are more data sensitive
3. Introducing a guided deflection flow before the agent to reduce encountering with it

Decision

After careful assessment, final decision was using a hybrid approach of strict Id&V and validation process in addition to a preliminary Agent-borne deflection steps. Deflection steps aimed to solve the issue without any action execution apart from data and article search. Use case planning phase included DPO team to verify the use case outcomes for avoiding any potential data-leak. Multiple specialized Agents created and strict "no-no" use cases defined per Agent to route consumer directly to the Service team instead of trying to solve every issue on the Agent level.

Finally, a detailed audit framework built and implemented using the standard platform capabilities to ensure the safety and accuracy of the Agent.

Outcome

Uneventful months after go-live and 27% reduced case handling time for the Service team.

What I'd do differently?

I would consider adding a multi-layered Agentic handling, using a specialised "front-desk" agent to handle rudimentary topics with only article search functionality and create deeper specialised agent swarm with assign responsibilities to keep have better maintainability and potentially even better accuracy/results at the end.

This was a conscious decision due to timeline, budget and the priorities.