Most executive teams I speak to in 2026 have more AI tools than they can count and less impact than they expected. The experimentation of 2024 and 2025 has run into something harder: the tools work, and the organization around them was designed for a different world.
My argument is that the value in this phase comes less from adopting tools and more from designing how people and AI agents work together, which I call orchestration. A conductor never plays a note and still shapes the whole performance by deciding where each instrument fits. That is closer to the skill organizations now need than prompt writing or tool selection.
I have worked on enterprise transformations at organizations including ABB, Holcim, Hilti, Adidas and Via Outlets, and this shift feels different in kind. Here are six changes that are reshaping how I think about strategy.
1. The valuable skill is orchestration
A few years ago, prompt engineering looked like the skill of the moment. Today, talking to a large language model is basic literacy, much like using a spreadsheet.
The higher-value work is designing how many autonomous agents collaborate, hand off work, escalate exceptions and produce outcomes that no single agent or person could produce alone. A tool user opens an application and performs a task. An orchestrator designs the system in which agents and people perform many tasks together.
When I led the agentic AI work at Holcim, the most valuable people on the team were not the best prompt writers. They were the ones who could look at a complex procurement workflow and see where human judgment was essential, where agents could work independently and where the handoffs needed guardrails. They were designing a system rather than operating a tool.
The implication for careers is uncomfortable. Positioning yourself as an "AI user" builds on ground that is eroding, because the value has moved from execution to design.
2. AI has moved to the CEO's desk
In 2025 and 2026, AI moved from a technical project to a question of strategy, talent and culture, and in many companies decision authority moved with it, from the IT department to the CEO. AI touches pricing, hiring, product development, customer experience and risk, which makes it hard to delegate to any single function.
I was sceptical when I first saw this shift at board level, because I have watched many CEO-sponsored digital transformations fade in steering committees. This time the personal stakes look different. When chief executives believe their own results depend on AI, the organizational gravity changes, and the transformation stops being an IT upgrade.
3. Process debt is the main cause of failure
Corporate AI spending keeps rising, and so do the failures. Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
I have seen this pattern before, in the CRM era and in the digital marketing wave. Organizations pour new technology into old processes and are disappointed by the results. The processes were designed around human limitations and grew over decades, full of manual handoffs, approval bottlenecks and undocumented knowledge. Automating them accelerates the friction rather than removing it.
A related problem is what I call agent washing: rebranding basic automation as agents. A rule-based chatbot gets a language model on top and is announced as agentic AI, and when leadership asks why it is not transforming anything, nobody has a good answer.
The organizations I see succeeding ask a more radical question: if we designed this process today, assuming agents would handle much of the execution, what would it look like? The answer is rarely the same process, only faster. It usually removes whole steps and human touchpoints that exist only for historical reasons, and adds feedback loops that only make sense with machines in the workflow.
4. Machine identities now dominate
Machine identities, such as APIs, bots, service accounts and devices, now far outnumber human identities in most enterprises, and they act, communicate and make decisions at a speed the old identity systems were not designed for. Directories, single sign-on and access control lists were built around people.
That volume creates administrative debt that manual oversight cannot keep up with, and it is growing as public TLS certificate lifetimes are cut in stages over the next few years. Organizations that have not automated certificate lifecycle management face a real risk of outages. Newer approaches, such as verifiable credentials for machines and zero-knowledge proofs that confirm an attribute like an agent's authorization level without revealing the underlying data, are emerging responses.
A CISO I spoke with this year put it well: "We do not have an identity problem. We have an identity crisis." The old model of one person, one identity and one set of permissions does not fit a world where one deployment can create hundreds of machine identities in a day.
5. Static compliance cannot govern agents
Agentic systems set sub-goals, interact with other agents and can behave in ways nobody explicitly programmed. A compliance checklist completed once a year cannot capture that. Gartner has predicted that by 2028 at least 15% of day-to-day work decisions will be made autonomously through agentic AI.
The EU AI Act and ISO/IEC 42001 push governance toward continuous monitoring and intervention. For agents working across procurement, customer service, finance and supply chain, the ability to pause or stop a specific agent or group of agents in real time becomes a basic operational requirement.
Human interaction is also becoming a weaker security barrier. OpenAI's GPT-4 system card described an evaluation in which the model persuaded a human worker to solve a CAPTCHA for it. When a machine can convincingly present itself as a person, "prove you are not a robot" stops working as a control.
The organizations adapting fastest design governance on three assumptions: agents will sometimes behave unpredictably, their scale will exceed what people can monitor directly, and regulation will tighten.
6. Synthetic people are hard to tell apart
AI replicas that imitate a person's voice, appearance and mannerisms have become good enough that, in a video call, a genuine colleague and a synthetic one can be hard to distinguish by observation alone. I have seen demonstrations that made me look twice at recordings of my own presentations.
That undermines the video call as a way of verifying who you are dealing with, which matters for contract negotiations, board meetings and customer interactions. Verifying the origin and integrity of digital interactions is becoming one of the central security problems of the decade, and the tools to do it at scale are still immature.
The strongest objection
A sceptical executive could say that "orchestration" is a new label for good process design, that most companies are nowhere near running fleets of agents, and that the identity and deepfake concerns belong to a few security teams rather than the leadership agenda.
The first point is largely true, and that is the point: the skill that matters is process redesign with agents in the picture, and it has an old and respectable history. The second is also true for many companies today, which makes this a good moment to fix processes and identity foundations before agent volume grows. The third underestimates how quickly security issues become leadership issues once a synthetic voice approves a payment.
Measuring return on AI
The measure of success is changing. In my experience, the return on AI comes less from adding AI to old systems than from retiring legacy technology and workflows that no longer fit. Every hour spent maintaining a process designed for a pre-AI world is an hour that produces no return, and digitizing an analogue process without questioning its design is a missed opportunity.
The conductor does not play louder than the orchestra. The job is to decide where each instrument fits and when silence works better than sound, and in 2026 that is the managerial job too.
Monday move
Pick one workflow where you already use AI. Draw it on one page with every handoff between people and agents, and mark three things at each handoff: who or what makes the decision, which identity and permissions it uses, and how it could be stopped. Any handoff where one of the three is unclear is the first piece of orchestration work to fund.