AI in production. What enterprises have actually built, how it is working, what it costs, and what breaks in the real world.
THE FORMAT
A closed-door conversation among peers
MCC Day is a curated, invitation-only event. No recordings. No product pitches. Under 100 executives and technical decision-makers from financial services, telco, banking, and enterprise, together for a full day of talks, confidential roundtables, and an evening on the rooftop terrace.
Curated
Every seat considered. Fewer than 100 attendees. The room is a feature of the event.
Confidential
No video, no recording. What’s said in the room stays there. That’s how you get the real numbers and the real trade-offs.
Peer-led
Talks by the people who built and ran the systems. Not the people who sell them.
The content
AI in production. What is next?
The theme is AI in production: what it costs, what it means to secure it, and what changes about software development and operations.
session 1A
Weaponized AI
The new attack surface of the agentic era
AI in production no longer just talks: it books, integrates, pays. Every new capability is a new attack surface, and most share one root flaw: models treat content as instructions. A map of the agentic threat landscape, from prompt injection through MCP tool poisoning to attacks on agent payments.
Red-teaming your enterprise AI stack
Jailbreaking the model is the easy part. Real exposure lives in the full stack: RAG pipelines, vector databases, orchestration, tool execution. A practical blueprint for offensive testing of AI systems, including automated red-teaming with adversary models, open-source tooling, and the hardening measures that actually work.
Transitioning to AI-driven exposure management
AI has collapsed the time between vulnerability discovery and exploit. Programs optimized for volume can’t cope with velocity. How exposure management must evolve: continuous validation, proving exploitability instead of counting findings, and prioritization that moves at the speed of the threat.
AI systems for autonomous fraud operations
Generative AI lets fraud scale faster than any human team. The answer is fighting AI with AI: autonomous systems that fuse security data, risk signals, and business context to reconstruct attack chains and act before fraud executes. Lessons from production deployments at tier-1 European banks.
session 1B
Sovereign AI
Sovereign AI: Implications for enterprises
Sovereignty is no longer a policy debate, it’s an architectural requirement. What full-stack independence means across silicon, infrastructure, data governance, and models; what the EU AI Act actually obliges; and how European enterprises balance regulatory constraints with access to frontier capability.
Managing open-weight LLMs and licensing
Open weights come with licenses that are anything but uniform: usage restrictions, redistribution clauses, revenue thresholds, acceptable-use policies. What legal exposure enterprises actually take on when deploying open-weight models, and a practical checklist for vetting a model before it reaches production.
Building your own foundation model in-house
When does training your own model make sense? Building a foundation small language model from the ground up: the sovereignty and domain drivers behind the decision, what it took in data, compute, and people, and what they’d do differently.
Using custom vision models in the enterprise
Language models get the headlines; vision models do quiet production work in inspection, safety, document processing, and monitoring. Where custom vision models beat general-purpose APIs, what it takes to train and deploy them on-premise, and how they fit a sovereign AI architecture.
session 2a
Agents at Work
The agentic IT ops stack
Ops teams spend most of their resolution time on investigation, not fixing. Anomaly detection isn’t the bottleneck anymore; reasoning is. How multi-agent systems take on the full incident lifecycle, from context-enriched triage and root cause analysis to remediation, and what infrastructure autonomous workloads require.
When agents run the SOC: Rethinking security operations
Adding AI to a broken SOC process just produces faster noise. This talk treats security operations as a process design problem: which workflows should be rebuilt around autonomous triage and investigation, where humans stay in the loop, and what changes for SIEM when agents are the primary log consumers.
Accelerating scientific research with AI agents
Before agents can automate discovery, the underlying knowledge has to be machine-usable. A biopharma R&D case: using LLM-powered agents to fuse and harmonize a biomedical knowledge graph, aligning ontologies, disambiguating entities, and pruning noise, so downstream automation and explainability have solid ground.
Agentic coding: Real-world lessons from replatforming projects
Agentic coding tools promise order-of-magnitude productivity. Replatforming projects are where the promise meets legacy reality. What worked and what didn’t when agents took on real migration work: where they excel, where they fail silently, and how team structure and review practices had to change.
session 2b
Token Economics
AI performance & unit economics: A blueprint
A framework for the numbers that matter: cost per task, inference routing, fine-tuning vs. prompting breakeven, and the build/buy/bypass decision. How to treat token budgets as an architectural constraint rather than a billing surprise.
Optimizing AI inference with autonomous, full-stack co-tuning
Most teams tune inference one layer at a time. But serving logic, runtime, and hardware parameters are interlocked: optimizing in isolation misses the global optimum. A grounded look at what siloed configuration costs on on-prem infrastructure, and how autonomous co-tuning reclaims performance and ROI.
Beyond LLMs: Building a global AI grid at the edge
Centralized inference has a physics problem: latency, egress costs, and data gravity. Akamai shares how a globally distributed edge compute grid changes the economics of AI serving, which workloads belong at the edge, and what an inference architecture looks like at planetary scale.
Small-model retrieval for cost-effective enterprise RAG
RAG doesn’t require frontier models. A grounded look at the real cost and quality trade-offs of building retrieval on small models: where they match large-model performance at a fraction of the cost, where they fall short, and how to decide which tier of model each retrieval workload actually needs.
All MCC Day sessions are in Italian. MCC Day will be followed by the annual Moviri Terrace Party.
THE FORMAT
Join a group of industry leaders
A small, curated group of senior technology and product leaders from European financial services, telco, banking, and enterprise organizations. Decision-makers who are shipping AI systems, or accountable for the ones their organizations are about to.
3rd
20+
<100
1
PAST ATTENDEES
They joined us
THE location
September 17, Milan.
Moviri HQ, Via Schiaffino 11
schedule
Noon to 10 PM
format
Talks · Roundtables · VIP Happy Hour · Evening Party
access
Invitation only, subject to host approval
Seats are limited and reserved for a select group of peers. Attendance is subject to approval.
