AI agents are crossing the threshold into enterprise production this year. For an executive, the real question is no longer whether to adopt the technology, but how to arbitrate three gaps that, if left unaddressed, will install lasting dependencies that are hard to undo.
Published on 2026-04-24 by Nathalie Lamborghini Dumas.
AI agents are crossing the threshold into enterprise production this year. For an executive, the real question is no longer whether to adopt the technology, but how to arbitrate three gaps that, if left unaddressed, will install lasting dependencies that are hard to undo.
2026 is the year agentic AI moves out of pilots. CB Insights documents a tenfold increase in M&A activity in the agent space in 2025, with nearly one hundred deals recorded. McKinsey notes that fewer than 2% of companies surveyed in November 2025 report having fully deployed an AI agent. In other words, adoption remains modest, but the structure of the market is crystallizing at high speed.
This asymmetry places executives in front of structural decisions that rarely reach them within a clear arbitration framework: which data to let flow out of the company toward AI platforms, which processes to delegate to agents, which control tools to keep in-house, which vendors to choose before exit costs become prohibitive. These decisions are taken today in day-to-day deployments or in market movements, rather than in the boardroom. Three gaps sum up what is at stake.
In most companies, agents arrive in daily operations before they appear on roadmaps. A prospecting agent here, a customer support agent there, a documentary synthesis agent in legal teams, another in finance. Each use case is adopted individually, often with a different vendor, a different governance layer, and data flowing through technical silos that don't talk to each other.
Taken in isolation, each agent does its job. Taken together, they form an architecture imposed by default, made of decisions no one in the company has formally taken. The leader ends up responsible for a system they did not design and whose interfaces they no longer control.
The cost of this situation appears later. When agents need to work together to shift from fragmented productivity to a strategic lever, it becomes nearly impossible to reassemble the puzzle. Business data is scattered across several proprietary environments that don't communicate. Each attempt at reintegration costs time, technical debt, and operational fragility.
The arbitration to make now is not to slow adoption. It is to decide which architectural decisions must rise to the strategic level, and which can stay operational, before positions lock in.
A chatbot answers. An agent acts. It queries a database, calls an API, modifies a customer record, sends an email on behalf of the company, validates or rejects a file. It is not software in the classical sense, it is a proxy.
This distinction is not semantic. It changes the nature of what you sign. When you deploy an agent, you are not ordering a tool, you are formalizing a delegation. The problem is that most contracts and implementations are framed as tool purchases. Often, no one has traced what is delegated, to whom, with what limits, and with what means of control.
Two technical notions become central here for an executive, even if they sound distant. Observability is the ability to see what the agent does while it does it, not just its final output. Evaluation is the ability to measure continuously whether it does it well, by your rules and not just by those of its provider. The investment fund Bessemer calls these two topics one of the largest unresolved bottlenecks in enterprise AI deployment. Without these two capabilities, facing a customer, a regulator or your own board, it becomes hard to explain what happened, and even harder to demonstrate due diligence.
The arbitration here, too, is simple in statement but strong in consequence. Each deployed agent should come with a formalized mandate: what it is allowed to do, what it is not allowed to do, who watches what it does, and how it is corrected when it drifts. Exactly what you would do for a colleague to whom you entrust a commercial signature. Nothing more, nothing less.
This is the quietest and most structuring gap. While many companies are still in steering committees on their first agent, the most strategic layers of the technical architecture of AI are being acquired, integrated, and closed.
The CB Insights numbers are eloquent. Salesforce conducted nine acquisitions in the agent space in 2025, Workday four. Datadog, the software supervision giant, has already taken positions in four agent observability players. Palo Alto Networks, Check Point, F5, Snyk, ClickHouse, Coralogix, and Anthropic itself acquired in 2025 and early 2026 players positioned on agent security, observability, evaluation, and governance. 54% of companies in this segment are still in early stage, meaning positions are not yet taken, but they are being taken at visible speed.
Translated into executive language, this means that the tools that will tomorrow judge, supervise and audit the agents you use are being absorbed into the platforms that sell them to you. Judge and party are structurally merging. This shift does not produce its effects today. It will produce them at the moment of the first difficult negotiation, the first audit, or the first strategic arbitration on AI.
The arbitration to make is to retain, even modestly, an observation and measurement capability independent of the platforms that supply the agents. Not to internalize everything, which would be unrealistic, but to keep a foothold outside the technical environment of the vendor.
Making architectural choices in 2026 does not mean building a sovereign end-to-end AI infrastructure. No SME or mid-cap can do this, and those who claim to are selling something else. It means, concretely, retaining mastery of a few precise layers.
The data layer first. Business histories, customer references, sectoral data can stay in environments controlled by the company, with interfaces that allow agents to use them without ever losing ownership or traceability. It is not free, but it is technically accessible to any mid-cap.
The separation between the model and its orchestration next. Using a large model as a language provider is one thing. Building all business logic on top of its orchestration platform, the one that makes agents work together, is another. The first dependency is manageable, the second is structural.
Tools that can be replaced by one another, rather than closed suites. An agent that can dialogue with others using open conventions can be changed or complemented without dismantling everything. An agent locked in a proprietary suite no longer can. At purchase, the price difference is minimal. At five years, it is considerable.
An independent observability, even modest. Several tools exist, notably open source, to trace and measure what agents do in production. The annual cost is low for a mid-cap, and this capability becomes the proof, before a regulator or a board, that the delegation entrusted to agents is genuinely under control.
None of these levers is spectacular. All are within reach of a company that decides to put them in place knowingly now, rather than discovering in three years that they were never put in place.
Four questions to bring to the next committee
Rather than another monitoring report, here are four concrete questions to bring directly to the executive committee.
1. Are we able, in committee, to precisely name each AI use in progress or under study in our company, and who is responsible for it? 2. Among the AI uses or projects we are considering, which are simple productivity tools, and which constitute a delegation of decision in the company's name? 3. Which company assets (data, customer relationships, know-how) do we absolutely want to keep under direct control, regardless of which AI provider we choose? 4. On which use cases do we really want to build a differentiated position, and on which do we accept to use what the market offers like everyone else?
The moment when architecture is decided
European SMEs, mid-caps and large groups hold precious assets in this shift. Deep business expertise, sensitive data, trust relationships built over time, mastered regulatory requirements. These assets are only monetized, however, if the architecture they fit into remains theirs. Otherwise they simply feed platforms whose training data they become.
The window is not indefinitely open. It still is. In 2026, it remains possible to make architectural choices that preserve the strategic freedom of the company. In 2028, it will be markedly more difficult, because defaults will be set, integrations locked, dependencies amortized.
Taking back control of these three gaps is not a technology project. It is a governance project. And it is decided now, in executive committees, before being decided by default in everyday deployments.
Sources: CB Insights, "5 AI Agent Predictions for 2026"; McKinsey, "How to build businesses faster and better with AI", March 2026.