Agentic AI: Decoding myth vs. fact

The Reality: General agents are masters of none; specificity wins.
It’s tempting to assume that an agent will solve everything right out of the box, but unfortunately, this is not the case. An agent deployed without tweaking or customization tends to be a jack of all trades and a master of none.

For specific, enterprise use cases utilizing your organization’s data, you absolutely must do training, benchmarking, and experimentation. Tightening the scope and fine-tuning the agent for that domain context will lead to better outcomes. Highly specialized agents perform significantly better.

The Reality: Change management is crucial to help people adapt to new ways of working.
Building a great agentic process in a technical environment is only half the battle. Just because it works technically doesn’t guarantee full adoption by the human team. If the agent changes something humans currently do, you will likely still need a human in the loop for judgment calls or regulatory requirements.

You must involve humans early for change management. This is more than just a technical issue; it involves addressing the human fear of losing jobs. Ensure there is alignment of incentives so that the human workers who need to adopt the process are invested in its success. Aligning KPIs to drive human buy-in is also highly recommended.

The Reality: AI is non-deterministic and requires rigorous guardrails.
Production environments expose all sorts of issues, unexpected hallucinations and boundary cases that testing didn’t encounter. Unlike classical software, where you expect the same result every time, AI is non-deterministic.

You need to put guardrails in place to ensure that the variability in the output is within an acceptable comfort level. When you stress test, try to consider all possible scenarios, as real-world use cases always present multiple edge cases. Additionally, tightening the scope of each agent’s responsibilities can help manage variability and maintain control over outcomes.

The Reality: Constant monitoring is essential for stability.
If you have reached production, your work isn’t over. The underlying technology is evolving incredibly fast, with new models being released almost weekly. Furthermore, models can experience drift over time, starting to produce different results.

You must constantly monitor and revisit new models to see if they can improve your agent outcomes. This means your agentic flow must be architected in a way that makes it easy to benchmark and swap models in and out.

When monitoring and evaluating, we typically focus on three dimensions:

1. Quality: Measured by metrics like accuracy, precision, and recall.
2. Cost: Finding models that perform better at a lower cost for certain tasks (often requiring a mixture of different models).
3. Speed: Ensuring the agent performs within the required time constraints.

At Fortude, we have taken these realities to heart and built specialized agentic capabilities into Charlie. Here’s a summary of its capabilities:

1. Signal-based inventory levelling
Retailers often struggle with stock-outs in some stores and excess in others. This agent forecasts demand using sales patterns and demand signals, then recommends proactive stock balancing across locations, reducing lost sales and carrying costs.

In short: Predict demand, rebalance inventory, minimize waste, and avoid stock-outs.

2. Model Context Protocol (MCP) for Infor ERP

Every new automation in Infor M3 required custom API work. Our MCP layer standardizes how AI interacts with Infor, enabling secure, context-aware ERP actions through natural language prompts.

In short: A reusable AI-ready interface for Infor M3 that accelerates automation and ensures secure access.

3. Demand forecasting for fashion
Fashion planners often rely on manual, backward-looking forecasting. This agent combines internal sales and inventory with external signals like weather, promotions, and trend data to produce SKU-level forecasts and reorder suggestions.

In short: AI forecasts future demand using real-time signals and automates PO recommendations for planners.

4. M3 release impact analysis
ERP teams spend hours reviewing release notes and mapping updates to custom configurations. Charlie’s multi-agent workflow reads release notes, checks client setups, and creates Jira tickets automatically.

In short: Automates release note review and impact analysis, cutting effort by up to 90% and preventing missed changes.

……

If you have been considering the possibilities of agentic AI or feeling nervous about its future, we hope this blog helps you navigate those important questions. We also hope it gives you the clarity to plan your next steps based on the realities outlined here.

Fortude’s teams have the expertise to help you build the right data foundation for AI and tap into the agentic capabilities of Charlie, our AI companion. Learn more about preparing for a future powered by agentic AI by booking a call with our team. And if you missed the webinar, you can watch it here.