Every software vendor now claims to have AI. Most of what's being sold is a search bar connected to an LLM, dressed up as an agent. The actual state of AI in operations is more nuanced, and more interesting, than the marketing suggests.
What AI Agents Are Actually Good At
Anomaly Detection
This is the most mature and genuinely valuable application of AI in operations. An AI agent that has learned your operational baselines, what a normal Monday food cost looks like, what a typical daily revenue range is for each location, can identify deviations the moment they occur. No dashboard checking required.
Pattern Recognition at Scale
Humans are good at noticing patterns in data they're looking at. AI agents are good at finding patterns across datasets too large or complex for a human to review: correlations between weather and demand, between staffing ratios and customer satisfaction scores, between marketing spend and in-store foot traffic.
Natural Language Querying
The ability to ask a question in plain English and get a data-backed answer is genuinely useful, especially for operators who don't have SQL skills. 'What were my top 5 revenue SKUs last week?' should not require an analyst. Modern AI agents can answer this directly.
Where AI Agents Struggle
AI agents are only as good as the data underneath them. If your data is fragmented, stale, or dirty, AI will confidently produce wrong answers. This is the most common failure mode, companies invest in AI before they've solved the data infrastructure problem.
- AI agents can't fix bad data, they amplify it
- Predictions require historical patterns, new businesses or new product lines have cold-start problems
- Causal reasoning is still weak, correlation vs. causation errors are common in automated insights
- Edge cases and novel situations often fall outside the model's training distribution
The Right Way to Deploy AI in Operations
Start with data infrastructure, not AI. Unified, real-time operational data is the prerequisite for everything else. Once your data layer is solid, anomaly detection and alerting are the highest-ROI first AI applications: they're mature, reliable, and immediately valuable.
“The best AI deployment in operations is invisible: it surfaces problems before they become crises, without anyone having to go looking for them.”
Forecasting and recommendation engines come next, once you've built trust in the underlying data and the detection layer. Natural language querying is a nice-to-have that makes the system more accessible, not a foundational capability.
The Bottom Line
AI in operations is real and valuable: but it's an amplifier, not a replacement for good data infrastructure. The companies getting the most value from AI are the ones who solved their data problems first. Build the foundation, then layer AI on top of it.