Agentic AI in 2027: What Practitioners Actually Need to Solve

Quick Facts

• 40% of companies with revenue over $1B are scaling AI agents (McKinsey, 2026).

• Only 19% of Databricks customers run agents at scale (Databricks, 2026).

• >40% of agentic AI projects are expected to be cancelled by the end of 2027 (Gartner).

• Teams that use evaluation tools bring roughly 6x more AI projects into production (Databricks).

• At DIS 2027, agents are covered on six stages, from engineering to industrial AI.

In 2025 and 2026, AI agents were mostly demonstrations: impressive systems that could plan, call tools and complete tasks in carefully controlled settings. In 2027 they meet production, with real enterprise data, real permissions, real costs and real auditors.

The ambition is clear. McKinsey found that 40% of billion-dollar companies were already scaling agents in 2026. Readiness is another matter. Databricks reported that only around one in five of its customers ran agents at scale. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.

The organisations that succeed will be those that treat agents as an engineering and governance discipline rather than a product feature. That means taking three kinds of problems seriously: engineering, governance, and data and infrastructure.

The first engineering problem is accuracy. A language model produces fluent, confident answers whether or not it has the right information. An agent acting on enterprise data needs an enterprise context layer that grounds its reasoning in the organisation’s actual definitions, documents and rules. Without one, an agent can be confident and wrong.

The second is evaluation. Many teams still lack systematic offline and online evaluation pipelines, so they cannot prove that their agents work or detect when they stop working. Organisations that invest in evaluation bring many more AI projects into production, because they can show stakeholders the evidence.

Memory and state are a third challenge. Agents that run for long periods need to remember the right things and forget the rest, without leaking sensitive information between users or tasks.

Tool access is a fourth. Protocols such as MCP make it easy to connect agents to enterprise systems, which turns exposing tools safely and at scale into a serious architectural decision.

Finally, there is observability and security. When a system makes its own decisions, step-by-step tracing is the only way to debug it. Once agents can act, prompt injection and tool misuse stop being chatbot curiosities and become infrastructure-level security risks.

Governing agents requires a different approach from governing traditional AI models. Gartner has warned that applying the same governance to every agent leads to failure. The level of oversight has to match what an agent can actually do. A system that drafts summaries needs very different controls from one that approves payments or changes production systems.

Scale creates its own challenge. Human-in-the-loop approval works when there are dozens of decisions, but not when there are thousands every day. Organisations need policies, monitoring and escalation mechanisms that keep humans in control without making them a bottleneck.

Accountability must also be designed in advance. When an agent makes the wrong decision, someone has to own the outcome, and organisations should prepare for AI incidents in the same way they prepare for security incidents.

Cost belongs on the governance agenda as well. Agents consume models, tokens, tools and compute on their own. Without a clear consumption strategy, AI costs can grow faster than the value they create.

Agents are not only consumers of enterprise data; they increasingly create and change it. That has consequences for database design, lineage, master data and data quality. As agents run continuously rather than waiting for human prompts, inference becomes the dominant AI workload, and infrastructure strategies built around training need rethinking.

In industrial settings the stakes rise further. Agents begin to interact with machines, robots and control systems, where reliability and safety standards are far stricter than in office software.

The 2027 Data Innovation Summit treats agentic AI as a cross-cutting theme rather than a single track, with relevant sessions on six stages:

• Machine Learning, Agents & Generative AI: context engineering, evaluation, memory, MCP in production and tracing.
• AI Strategy & Governance: autonomy, oversight, accountability and token economics.
• AI Automation & Transformation: agentic workflows and the orchestration of humans and agents.
• Databases, Context & Data Quality: agent memory stores and retrieval architectures.
• AI Infrastructure & Operations: the infrastructure needed for always-on agents.
• Industrial & Physical AI Spotlight: agents that act in the physical world.

The Data Innovation Summit is the largest and most influential annual applied data, analytics and AI event in the Nordics, organised by Hyperight in Stockholm since 2016. The 12th edition, DIS2027X2, takes place on 18–20 May 2027 at Kistamässan in Stockholm under the theme “From Data & AI to Enterprise Capability”.

The summit brings together more than 3,500 practitioners, business leaders, technology providers, start-ups and academics from over 50 countries. The programme runs across fifteen stages: two Foundation Days dedicated to data, analytics and AI capabilities, and an Industry Day covering five sectors.
The summit is a hybrid event, with every session streamed live and available on demand.

Regional editions are held in Dubai, Singapore and Sydney.

Learn more at datainnovationsummit.com.