SMEs can afford agentic AI when they deploy it with discipline and focus on ROI at every step. At the AI Frontiers 2025 session organized by IIoT World, panelists from AstraZeneca, Cybus, Infinite Uptime, and Frost & Sullivan confirmed this. AstraZeneca cut AI model build time from three months to one. The same principles scale down to any plant floor.
Below are the most relevant takeaways, drawn directly from the live audience Q&A.
How Do SMEs Prove ROI for Agentic AI?
Q: How is your leadership team addressing challenges in justifying ROI when introducing AI?
Answer — Madalina Ciortan, AstraZeneca: By using AI copilots to accelerate model development, AstraZeneca cut build time from three months to just one. In production, fewer breakdowns and faster troubleshooting reduced downtime—direct savings with measurable impact.
For SMEs: While AstraZeneca is a global company, the principle applies everywhere: ROI is built from measurable, visible wins. For smaller manufacturers, that might mean reducing time spent on paperwork, solving a chronic machine failure, or eliminating a reporting bottleneck. Start small, measure clearly, and expand only after value is proven.
What Are the Hidden Costs of Agentic AI for Manufacturers?
Q: What’s the most underestimated barrier for agentic AI readiness?
Answer — Sebastian Trolli, Frost & Sullivan: Context gaps—missing metadata and semantics—are often overlooked. Without them, agents act on incomplete or ambiguous signals.
For SMEs: If you skip the step of standardizing basic context and metadata up front (for example, asset IDs, machine states, or units of measure), you will face expensive rework later. Every system will need a custom integration. Even a lightweight namespace or simple rules for naming and tagging can save significant money by making systems easier to connect and scale.
Should SMEs Replace Existing Manufacturing Systems for AI?
Q: How can CFOs/COOs move forward while salvaging existing investments?
Answer — Raunak Bhinge, Infinite Uptime: The keys are interoperability and standardization. Use open interfaces so best-of-breed tools can talk to each other, and make sure sensors are correctly applied. AI can’t fix poor sensing.
For SMEs: Protect your balance sheet. Keep your existing MES, historians, and dashboards. Add an integration layer instead of ripping out systems. Standardize sensors and tags so each new project builds on the last one—making every step cheaper.
How Does Edge Computing Cut Manufacturing Cloud Costs?
Q: How do manufacturers integrate edge and cloud properly while avoiding exploding costs?
Answer — Peter Sorowka, Cybus GmbH: Run real-time inference at the edge and use the cloud for training and fleet-level analytics. Architect for reusability—shared pipelines and data services prevent duplication and ballooning cloud spend.
For SMEs: A small industrial PC or gateway can handle local filtering and inference. Only send essential signals to the cloud. This keeps monthly bills predictable and prevents surprise overspend.
How Does a Unified Namespace Reduce Integration Costs?
Q: How can manufacturers resolve OT/IT silos—UNS/MCP, etc.?
Answer — Peter Sorowka, Cybus GmbH: A unified namespace or similar approach ensures real-time data is standardized and accessible. The biggest hurdle is often organizational—procurement must demand that machines come with integration-friendly interfaces.
For SMEs: Every silo costs money because it forces you to pay for duplicate integrations. Standardizing once through a namespace or other shared structure means you don’t keep paying integrators every time you add a new machine or vendor.
How Do Manufacturers Build Trust in Autonomous AI Systems?
Q: How do we build trust so operators and executives allow real-time decisions?
Answer — Sebastian Trolli, Frost & Sullivan; Peter Sorowka, Cybus GmbH; Raunak Bhinge, Infinite Uptime:
- Begin in read-only mode, then move to advisory, then limited autonomy.
- Prove reliability at each stage with audit trails and explainability.
- Combine process data and equipment data to reduce false positives.
For SMEs: A failed rollout can waste scarce resources. By staging autonomy in levels, you only invest further once the system has proven value and earned operator trust. This reduces both financial risk and cultural resistance.
What Steps Should SMEs Follow for Agentic AI?
1. Pick one loss. Target a recurring downtime cause or costly bottleneck.
2. Instrument wisely. Ensure the right sensors are in place and configured.
3. Standardize context. Create a minimal namespace with IDs, states, and units.
4. Go edge-first. Keep control local; send summaries to the cloud.
5. Stage autonomy. Read-only → advisory → constrained write.
6. Track the numbers. Expand only when ROI is demonstrated.
Speakers who contributed to these answers:
• Madalina Ciortan, Director AI, OpsIT — AstraZeneca
• Peter Sorowka, CEO — Cybus GmbH
• Raunak Bhinge, Founder — Infinite Uptime
• Sebastian Trolli, Research Manager & Head of Industrial Automation & Software — Frost & Sullivan
Related articles:
