AI Agents in Workforce Communications

In recent conversations about the important role of AI in corporate Australia, there is increasing rhetoric among senior business leaders that AI agents, such as Claude can be "connected" to an organisation's workforce communications network to solve the persistent issue of declining productivity.

At Traffyk.ai we are all excited by the future applications of AI agents, unfortunately it's not that simple.

Why?

1. Foundation models have limited inside knowledgeAI models learn from public information, which does not include organisation specific knowledge. Kevin Weil, the CPO of OpenAI recently commented "The future is going to be incredibly smart broad base models tailored with company specific data…" However, customising AI with a company's internal data and subject matter expertise is complicated and creates challenges if not done securely.

2. Agents' success to-date is limitedCurrently working AI tools can only solve specific, verifiable tasks using well-organised data—coding assistants being the most successful example.

3. High‑profile agent rollouts have already hit costly, very public roadblocksWe have recently seen issues, namely around accuracy and privacy with some of the most widely used examples including Microsoft Co-pilot and Atlassian Intelligence.

4. Human oversight is still neededEven the best coding agents have a pass rate of 50% on coding tasks that take 1 hour (METR 2025).

5. Significant operating costsFor company-wide deployment (like email integration), current AI agents would conservatively cost in the order of $2M USD/yr for a 10,000 employee company.

With current capabilities, agents are not a silver bullet for workforce communication. We are still at a stage where human experts with domain knowledge are required to work with world class data scientists and engineers to properly implement AI in a targeted manner to make up for the shortfalls and keep costs low.

Before betting the office on autonomous agents, organisations need robust data‑governance pipelines, careful cost modelling, and human‑in‑the‑loop safeguards—not hype.Get in contact if you would like to understand more.

‍

Dr. Daniel King

Principal AI and Data Scientist