An AI model can produce a convincing answer to a customer in seconds. Giving that same system permission to read an order, check a payment, update a record and send the right response is a more demanding engineering problem. It has to work with existing software, distinguish instructions from untrusted data, respect access controls and leave enough evidence for someone to investigate a mistake.
Those are among the problems behind the International AI Summit 2026, taking place on December 2 at Steigenberger Wiltcher’s in Brussels, with online attendance available. The event, organized by Forum Global, brings together policymakers, researchers, technology companies and technical standards organizations. Its agenda focuses on AI sovereignty, computing infrastructure, deployment, testing and regulation. It is not a confirmed launch event for a new AI model.
Much of the conversation around AI still centres on model capabilities. The Brussels programme asks what happens when those capabilities have to become dependable systems. That distinction could tell us more about the direction of AI in 2027 than another benchmark leaderboard.
From an AI model to a production agent
A chatbot usually generates an answer from a prompt and its available context. An agent can go further: decide which tools to call, retrieve data, follow a multi-step workflow and request an action in another system. In a real application, the model is only one part of that arrangement.
Consider an online retailer using an agent to handle a delivery-address change. The request might pass through a language model that identifies the customer’s intent, an account service that confirms identity, an order-management API that checks whether the parcel has shipped, and a policy layer that decides whether the address may be changed. The system should present the proposed change for confirmation where appropriate, execute it once, and record the outcome. If the carrier refuses the update, it needs a useful failure path rather than a fabricated success message.
That raises familiar software-engineering concerns in an unfamiliar setting: authentication, transactional consistency, retries, idempotency, validation and audit logs. A repeated tool call must not accidentally trigger two refunds or duplicate an order. The agent’s explanation of what happened cannot be the only evidence that it happened.
The gap between experimental AI and these production requirements is substantial. Stanford’s 2026 AI Index reports AI use among 88% of organizations surveyed in 2025, while agent deployment was still in the single digits across nearly all business functions examined. These numbers measure different things, but together they show why adoption of an AI assistant is not the same as operating autonomous workflows at scale.
The summit’s session on moving from experimentation to systemic adoption will examine how AI fits into existing enterprise systems and who is responsible when an automated action causes harm. For engineers, the important question is less whether an agent can complete a task once than whether it can do so consistently, under controlled permissions, across thousands of real cases.
Interoperability is not the same as giving an agent unrestricted access
A business may use separate systems for payments, customer records, inventory, support and logistics. Connecting an AI model to each one through custom code becomes difficult to maintain, especially if different vendors and models are involved.
Two technical developments help explain the direction of the industry. The Model Context Protocol (MCP) provides a common way for AI applications to discover and use external tools or data sources. The Agent2Agent (A2A) protocol addresses communication and task collaboration between independent agent systems. They deal with different integration boundaries; neither makes a connected system secure merely by standardizing the connection.
The engineering challenge moves to the identity and authorization layer. An agent that may read an invoice should not automatically have permission to alter a bank account. Permissions need to be tied to the user, task and target service, with the ability to revoke access and inspect activity.
In an August 2026 technical discussion, the US National Institute of Standards and Technology warned about agents using shared human credentials and long-lived API keys. It points to established identity approaches, including OAuth 2.0 and SPIFFE, as foundations for granting and controlling access. NIST also launched an AI Agent Standards Initiative earlier in 2026 to advance interoperability, security and agent identity.
The practical direction for 2027 is more likely to be controlled integrations—with limited scopes, explicit authorization and reliable logging—than one universal agent connecting to everything with a user’s password.
Security testing changes when AI can take actions
Prompt injection illustrates why adding tools changes the risk. An agent reading a customer email or a web page may encounter text that tells it to ignore its instructions, reveal private data or call a tool it should not use. That text is part of the material the agent was asked to process, not an authorized command.
If the agent has broad access to internal records, the resulting failure is no longer just an incorrect answer. It could become an unauthorized database change or a disclosure through a connected service. The OWASP GenAI Security Project’s 2026 incident review documents how prompt injection, excessive agency and weak permissions can interact in real systems and security research.
Developers therefore need controls outside the model. A permission service should reject an unauthorized API action even if the model requests it. Sensitive operations should have explicit approval paths, tool inputs should be validated, external destinations restricted where appropriate, and credentials stored outside prompts and logs. A model’s refusal to perform a harmful action is useful, but it is not a substitute for enforcing the rule in code.
These are among the reasons the summit’s technical standards session matters. It is scheduled to address testing methodologies, red-teaming, assurance and how technical requirements can be made measurable rather than left as general promises about trustworthy AI.
The benchmark that matters is an entire workflow
Traditional AI benchmarks often measure answers to questions. That is useful for comparing model capabilities, but an agent operating a business process also needs to be evaluated on the actions it performs.
For the delivery-address example, a serious test would include an eligible order, an order that has already shipped, ambiguous customer identity, an unavailable carrier API, a duplicate request and malicious instructions hidden in a customer message. Engineers should measure whether the agent completed the authorized task, whether it attempted anything outside its permissions, how often a human had to intervene and whether recovery worked after a failure.
That means recording tool calls and outcomes, keeping test cases reproducible and comparing a new system version against previously observed failures. A model can improve on a reasoning benchmark while becoming less reliable on a particular operational task; a release process needs to detect that.
This is an active research area, not just a proposed best practice. NIST’s work on evaluation probes for agentic AI examines tools that evaluate claims against trusted documents and produce machine-readable audit trails. Its research focuses on whether source material actually supports an agent’s conclusions, an important part of making long workflows auditable.
For 2027, progress in this area may be as significant as an increase in model capability. Businesses need evidence that a system performs correctly under their own conditions, not only a provider’s demonstration.
AI infrastructure has a memory, power and latency problem
Larger and more capable models are not useful if the infrastructure cannot run them at an acceptable cost or response time. Training receives much of the attention, but inference—the computation required each time a model serves a request—becomes a continuing operational expense when millions of users or long-running agents are involved.
Performance depends on more than raw processor speed. GPU availability, high-bandwidth memory, network capacity, data movement, batching, caching and model optimization all affect how quickly and affordably AI workloads can run. A workflow that makes several model calls and external API requests can be much slower than a single chatbot response, even before accounting for validation or human approval.
The International Energy Agency’s 2026 assessment estimates worldwide data-centre electricity demand at approximately 485 terawatt-hours in 2025 and projects roughly 950 TWh in its central 2030 scenario. These numbers cover all data centres, not AI alone. The same report identifies constraints involving grid connections, electricity equipment and high-bandwidth memory supply, while noting that efficiency per simple AI query has improved significantly.
That mixture matters. A smaller model running efficiently may be preferable for a routine task; more demanding reasoning or video generation may require considerably more computing. Scaling an AI product involves choosing the right model and deployment design as much as acquiring more hardware.
The Brussels agenda includes both data-centre infrastructure and the connectivity required for physical AI, where network reliability and latency become important for systems interacting with machines and devices. Neither topic is solved by publishing better model weights.
Sovereign AI depends on where the system actually runs
The summit also considers AI sovereignty, which has a direct technical meaning beyond national ownership of technology companies. A government or business may need to know where data is processed, which cloud provider hosts a model, who can access logs, whether model weights can be deployed elsewhere and what happens if a vendor changes its terms.
Those questions do not have one universal answer. A self-hosted open-weight model can offer more control over deployment but still depend on imported chips, software and specialist maintenance. A managed cloud model can reduce engineering overhead while introducing dependencies involving pricing, data processing, access and availability. Data residency alone does not answer every sovereignty or security question.
Canada is represented in the summit’s sovereignty discussion: the organizer lists Alexandra Dostal of Innovation, Science and Economic Development Canada on the panel alongside Cohere, the European AI Office and Bulgaria. The conversation is framed around strategic interdependence rather than the assumption that any country can build an entire AI supply chain alone.
Businesses choosing AI services should apply the same care. ZahraAli.ca’s article on AI tools with Canadian roots explains why a Canadian connection does not necessarily establish Canadian ownership or data processing. In technical procurement, the more revealing questions concern deployment regions, data-retention terms, access controls, exportability and how difficult it would be to change providers.
Regulation is becoming an engineering requirement
The EU AI Act is not simply a policy topic for lawyers. Where its requirements apply, teams developing and deploying AI need processes for managing risk, documenting systems, testing performance, maintaining appropriate records and providing human oversight.
According to the European Commission’s July 2026 update, the revised application date for high-risk systems under Annex III is December 2, 2027, while certain high-risk systems embedded in regulated physical products follow on August 2, 2028. Other parts of the law already apply. Not every chatbot or website feature is automatically a high-risk system, and the obligations depend on its use and legal classification.
The difficult part is translating a legal requirement into something a development team can demonstrate. That might involve documented data sources, versioned evaluations, records of changes, clear escalation routes and tests showing that people can intervene when required. The summit’s session on standards and assurance explicitly addresses this move from principles toward measurable technical practices.
The user interface is part of the system too. An AI feature needs to show when a task is in progress, distinguish a proposed action from a completed one and explain errors without misleading the customer. An earlier Figma AI Design Agent test on a bilingual Montréal website found that visually coherent screens could still lack working interactions and language behaviour. Agent-enabled products make that gap more serious because a confusing interface can hide a consequential action.
The same applies to websites that will be used by AI agents. Clear structure and readable content help, but completing an actual task requires consistent forms, predictable application states, authenticated operations and confirmation that a transaction succeeded. ZahraAli.ca’s AI-agent website-readiness test addresses the difference between a site being readable and genuinely usable by an agent.
What the technical discussions in Brussels could tell us
The most useful announcements on December 2 would not necessarily involve a new model. Evidence of reliable agent deployments, interoperable systems that preserve access controls, repeatable evaluation methods or workable plans for computing infrastructure could have a more immediate effect on the software people build next year.
There is a credible case for more agent integration, stronger evaluation tooling and more attention to deployment costs in 2027. There is no equivalent basis for promising that autonomous systems will work safely in every environment or that the industry has solved the problem of trust.
The official International AI Summit programme describes what participants plan to discuss, not what they will conclude. For developers, designers and businesses, its real value will be in whether those discussions produce engineering approaches that can be tested, implemented and maintained after the conference ends.


