AI agents are moving from clever demonstrations into business systems that make decisions, change records, and act on behalf of users. That shift creates a new challenge: accurate data alone does not guarantee a correct outcome, because agents also need stable context, clear controls, and a complete view of what happened.
The next phase of enterprise AI will be shaped by the systems behind the agent. Data stores, retrieval platforms, security controls, GPUs, and developer tools now determine whether an agent can explain its decision, protect sensitive information, and operate without creating new risks.
Context Has Become the Core Battleground
CTGT is attacking one of the hardest agent problems with Concordance, an open source tool designed to trace why AI agents make inconsistent decisions. In one test, CTGT replaced identifiers with stable labels, sorted the list, and isolated each test inside a fresh organizational environment. Five subsequent runs each produced the same 34 obligations.
That result points toward a larger goal for the next-generation AI stack: unifying data and knowledge without borders. Nearly twice as many companies now name live-system queries as the main context source for their AI agents, showing how much agents depend on current business information rather than isolated model knowledge.
OpenAI, Keewano, and Snowflake supply that missing context in different ways. KeewanoDB gives every user, device, and agent its own table, and queries never join across them. Keewano says agents cannot explain why something happened because the sequence was taken apart before they ever queried it.
That design places sequence and context at the center of agent reliability. Felix Van de Maele, founder and CEO of Collibra, says enterprises running more than a handful of agents in production face risk, cost, and oversight burdens that grow exponentially with every added agent. In a conversation with VentureBeat founder and CEO Matt Marshall at VentureBeat’s AI Week event in New York, Van de Maele traced most agent failures to the business context those agents receive.
His prescription is direct: companies need deterministic controls on agent behavior before their agent count grows. The same principle appears in the broader guidance that scaling AI agents requires three things: inventory, contracts, and maintained context.
Agents Are Moving Into Real Business Workflows
Siena plans to expand its “Agent of Record” by combining purchase history with support and social information. Its website describes agents that can update delivery addresses, create return labels, process refunds, and arrange replacements. These are not passive chat responses; they are business actions that require permission, history, and reliable records.
Enterprise leaders are also watching how AI changes engineering itself. The share of developers using AI to write half or more of their code jumped from 12% to 42% year over year in the latest BairesDev survey. The implication for enterprise engineering organizations is clear: AI adoption does not eliminate the cost of expertise; it changes where that expertise is required.
That expertise now includes checking data quality, designing agent boundaries, and deciding when automated changes can reach production. Fifty-six percent of enterprises with AI agents allow automated production changes, down from 75% in July. At the same time, 64% of enterprises traced a confidently wrong AI agent answer to problems in their own company data during the past six months.
Semantic layers help detect those errors and affect decision-making processes. Yet only 2% of enterprises say the model was the main reason they chose an AI platform. Flexibility and control matter more: 78% of enterprises value those qualities over model alignment when selecting platforms and shaping orchestration decisions.
Platform competition reflects that priority. Forty-eight percent of organizations running AI in production name a Google platform as their primary AI platform, more than twice OpenAI’s share and ahead of Azure. OpenAI’s guardrails secure 40% of enterprise AI agents, but 43% of organizations say attackers are ahead, while 59% have faced security incidents or near-misses.
The Infrastructure Race Is Getting More Intense
Infrastructure costs can rise before an agent completes a single task. Up to 60% of GPU compute goes toward recomputing context that large language models already processed. Alan McSeveney, field CTO at WEKA, said the waste comes from how models hold conversation history in GPU memory, and argued that fixing it lets companies serve several times more users on hardware they already own.
AWS is pitching OpenSearch ahead of OpenSearchCon North America as one platform for retrieval, serverless economics, and agent memory. That combination shows how search, cost management, and context storage are converging inside the AI stack.
Privacy is another infrastructure test. Cohere’s Model Vault encrypts AI inference so even Cohere cannot see enterprise customers’ data. Cohere plans to open source the full serving stack used in Model Vault, allowing independent auditors to validate that it does not log, export, or leak data.
Google is also expanding a pilot program that pays developers and small businesses for proprietary offline code and other data. Google has not specified which products will use code obtained through the pilot, how much it pays, or what rights it receives under a typical agreement. Those unanswered details put contracts and data rights alongside model performance in the center of enterprise AI planning.
Even the style of generated text remains measurable. Claude Opus 5.5 uses em-dashes 99% less often and sounds more human, yet it still exhibits 2,548 AI writing tells. All models receive one fixed prompt, and the team does not specifically instruct them to imitate human authors or suppress known tells.
The financial pressure reaches the cloud market itself. Gartner predicts that more than 80% of neocloud providers will cease operating independently by 2030. John Lovelock, Gartner’s chief forecaster, discussed the firm’s latest report on “Squawk on the Street,” adding a stark warning to an industry racing to build AI infrastructure.
The message across these developments is hard to miss: AI agents will not scale on intelligence alone. They need traceable decisions, maintained context, deterministic controls, protected data, efficient compute, and contracts that define who can use the information. The companies that build those foundations now will decide whether the next generation of agents becomes a trusted operating layer—or an expensive source of decisions nobody can explain.
Based on
