LLM governance

AI agents

AI agents can speed up operations by processing near real time incoming inquiries. The reliability and preciseness of the responses depends on few factors which we need to assess first. Let’s consider customer support first. The ERP/CRM system is well integrated and contains the full-standing information about the customer, the AI agent should be able to retrieve the exact information and formulate a proper answer without guessing any non-factual information There are multiple available run-books explaining what is the company procedure and how to solve customer issues. Those run-books should be coherent, updated and complete. I suggest to create synthetic data of customer issues to evaluate that LLM answers are reproducible and reliable. There is an extensive documentation around legal risks associated with wrong answers so we can build guardrails which will review the answers before responding to the customer. The automation of labor intensive tasks is a promising cost-cutting in many branches but we need to make sure we properly evaluate the connected risks because failures can have high cost in terms of legal actions and reputation. This is why I suggest a step-by-step approach to enable a reliable transition.

enrollment

PoC

We collect all the input material (API integration, run-books, labelled cases) and we build an off-line model which will evaluate historical data. historical data: check any mis-interpretation in customer inquiries synthetic data: create additional data points to explore less common issues LLM agent: build an tune an LLM agent which performs the best evaluation and has a good cost balance Guardrail: which will further evaluate responses At the end of the PoC many metrics will be computed (response time, reliability, cost reduction/inquiry, false positive rate…) and then decision to proceed to next phase will be compared to business KPIs.

MVP

The PoC code will be cleaned up and put in a dev environment attached to live resources. A portion of the incoming traffic will be elaborated from the AI agents and the results will be compared in a variant testing. It KPIs prove to be as solid in a live situation the solution can be enrolled for a larger portion of traffic. Product The MVP code will be re-written to perform at scale and the portion of human agents will be still kept as control group. Performances will be monitored and we will put alerts for model performances drops or data drifts.

Governance

AI agents don’t take any reliability for their answers. There is no LLM provider able to guarantee the stability of answers which we need to check during PoC. The company should need to decide few crucial points before you put the solution in production to avoid serious unexpected risks. Local/public: do we need to deploy a local model to avoid sensitive information leak (what is the company policy?) Performances: who is responsible for performances and what is the expected SLA Responsibility: who is the ultimate responsible for customer communication and what is the expected behavior for those agents (conservative vs creative) Customer satisfaction: how do we evaluate and consider customer feedback to be able to revert in case the market will badly react

Conclusion

These are few thought regarding the implementation of AI agents for customer support based on my experience. My experience helps to forecast problems which could not be seen at the beginning but thanks to the big exposure I have with telco, e-commerce and media company I believe this is the pathway to follow for such an implementation. First of all we should start the PoC to deeply evaluate whether the expectation of the AI agents meets the business expectation. This step is low-cost and creates many valuable insights about efficiency in operations for the company and adds added value independently on the enrollment of the project.

LLM governance

Agentic workflow is shifting the paradigm from running operations to monitor the operativity of agents. The system should define clear ownership across all the components to allow proper decision making. The key component is traceability: the system should be built so any agent decision is more or less reproducible and we can spot where the issue was initiated. There are as well few main areas where different stakeholders need to monitor and control the status and the success of the system. Those areas should properly define metrics, KPIs, indicators, logs, checks, evaluations… so each owner can control and act on the system. The system should follow some guidelines and all the components should be consider them. The main areas are:

backend

cost and performance control

fitness

how much the results are aligned with the goals

goodness

checks on model performances and edge cases, data drifts and models evaluation

business

is the overall system performing correctly and delivering value

https://orgmode.org

Many stakeholders get lost in defining a proper governance method especially regarding some terms:

traceability

rebuild the actions done by the agents using accurate logs

observability

a much generic term to indicate the collection of metrics around a running system

evaluation

check model performances against typical metrics like: hallucination,

KPI

more business oriented indicators, they should comprise multiple quantities and

(org-html-export-to-html)

backend

fitness

goodness

business

commands

echo "This is my configuration script"
echo "..do something interesting here.."

(defun org-xor (a b) "Exclusive or." (if a (not b) b))