Are Control Theory and World Models a Silver Bullet for M&A Automation?

The AI research community is having a serious conversation about "world models." Yann LeCun, Meta's Chief AI Scientist, has spent years arguing that the next frontier in AI is not larger language models but systems that build genuine internal representations of how things work — systems that predict the consequences of actions before taking them.

Meanwhile, M&A practitioners are stuck in a conversation they have been having for decades: why do so many deals fail to deliver promised value, and can any new technology actually help?

These two conversations rarely intersect. They should.

What LeCun Actually Means by a World Model

In a February 2024 LinkedIn post, LeCun gave a formal definition. Given an observation x(t), a previous state estimate s(t), an action proposal a(t), and a latent variable z(t), a world model computes: h(t) = Enc(x(t)) and s(t+1) = Pred(h(t), s(t), z(t), a(t)). Enc is an encoder mapping raw observations into a compact internal form; Pred forecasts the next state from that representation.

In plain language: the system takes what it can observe, maintains a running model of the underlying situation, and uses z(t) to represent everything it cannot directly see — the irreducible uncertainty, the hidden factors that would fully explain what happens next if only you could measure them.

Introducing JEPA

The architecture LeCun proposes to implement this is called JEPA — Joint Embedding Predictive Architecture. Rather than predicting in raw pixel or token space, JEPA predicts in a compressed abstract representation space, building a system that reasons about consequences rather than interpolating surface patterns. A hierarchical version, H-JEPA, extends this to multiple time scales — short-term, fine-grained predictions feeding into longer-horizon planning.

M&A Is a Control System. It Just Does Not Know It.

A deal process has all the structural features of a control system. The target company and the deal itself are the objects you are trying to steer. The state of the deal comprises everything you would want to know: DD findings across workstreams, financial model updates, integration planning maturity, relationship quality with the seller, regulatory exposure, key employee retention risk. That state is never directly observable. What you actually get are partial observations — data room extracts, expert interviews, management presentations, AI-analyzed contracts — strategically curated by the other side.

The actions your team takes are gate decisions: proceed, renegotiate, request more information, adjust the integration approach, or walk away. The setpoint is what you were aiming for at deal inception — strategic fit criteria, a valuation range, synergy targets. The error signal is the gap between expected and actual findings. Feedback is what you learn post-close: did synergies materialize, did timelines hold, did integration succeed?

Mapped this way, the M&A reference model — with its decision rules, guiding questions, and gate criteria — functions as the controller. It takes the observed error signal and prescribes a response.

The uncomfortable diagnosis this framing surfaces: most M&A processes run open-loop. Teams execute checklists without real-time state estimation. They accumulate data without systematically updating a model of the deal's true state. Actions are taken based on the last management presentation rather than a continuously maintained estimate of reality. This is not a criticism of practitioners — it reflects the genuine difficulty of the problem. But it explains why value destruction rates have remained stubbornly high across decades of process improvement efforts.

Three Things Control Theory Actually Adds

Three concepts from control theory translate directly into actionable diagnostics for M&A process design.

The first is observability. In control theory, a system is observable if you can reconstruct its full internal state from the measurements available. If it is not observable, no amount of analytical effort recovers what you are not measuring. Applied to M&A: given the data objects your due diligence process collects, can you actually determine the deal's true state? If your process model has no mechanism for assessing management depth, the deal state is not observable along that dimension. The fix is not more analysis — it is better sensors: new DD workstreams, different interview structures, third-party reference checks.

The second is Model Predictive Control, or MPC. An MPC controller uses a forward model to simulate outcomes under different action sequences and selects the one that best achieves the objective over a defined horizon. This is exactly what structured scenario modeling does. When an M&A team asks "what happens if we proceed versus renegotiate versus walk away," they are doing informal MPC. The formal version makes explicit which model drives the prediction, what assumptions it rests on, and what criteria define the objective. The M&A reference model with rich data objects are the raw material from which that predictive layer can be built.

The third is hierarchical control. Strategy sets targets that Due Diligence must validate. DD findings drive integration planning. Integration planning cascades into operational execution. Each level operates at a different time scale and precision. This mirrors H-JEPA's multi-level architecture: short-horizon fine-grained predictions feeding objectives to longer-horizon, higher-abstraction planning. Understanding M&A as a hierarchical control problem — rather than a linear phase-gate process — changes how you design escalation logic and when to propagate new information upward.

Where the Analogy Breaks

M&A deals are not stationary systems. Control theory assumes a plant whose dynamics are relatively stable. A target company during an acquisition is neither. Management behavior shifts once a deal is live. Competitors react. Markets move. Regulatory environments change. The plant is actively responding to the fact that you are modeling it — a dynamic with no clean analogue in industrial control systems.

The latent variable z(t) is, in M&A, enormous and often irreducible. LeCun's formulation notes that z captures "the unknown information that would allow us to predict exactly what happens." In robotics, z might be the precise position of a partially occluded object. In M&A, z encompasses hidden liabilities no data room will surface, the seller's true walk-away price, cultural compatibility at the operational level, informal influence networks inside the target, and the reliability of every management assertion.

Another issue: You cannot run controlled experiments. In robotics, a world model trains across thousands of trials with systematic variation. Every M&A deal is unique and non-repeatable. You cannot replay the acquisition with a different opening bid and observe the counterfactual. Empirical calibration of any predictive model is therefore extremely difficult.

The state space is partially qualitative in ways that resist mathematical encoding. Trust between buyer and seller, relationship quality with key employees, cultural alignment — these are real variables that determine deal outcomes and are not easily reduced to numbers without destroying the information that makes them useful.

The feedback loops are slow. A robotics controller gets feedback in milliseconds. M&A feedback arrives over years — long enough that organizational memory of the decisions made and why has often degraded before outcome data arrives.

What Practitioners Should Actually Take Away

The silver bullet does not exist. But the framing is still productive. Here is what I think it concretely implies:

Use observability as a diagnostic. When a task in your M&A process model lacks sufficient data objects to determine a meaningful aspect of deal state, that is an observability gap. It can be fixed — by adding data sources, by redesigning a DD workstream, by building different interview instruments. This is a more precise diagnosis than "we need better due diligence."

Understand automation workflows as the actuator layer. Workflow automation tools, direct API integrations with platforms like DealCloud or SAP, or agent-based systems — execute what the process model prescribes. They are actuators, not controllers. Keeping this separation of concerns explicit improves architecture: the process logic (controller) and the execution layer (actuator) should be designed and maintained separately. Conflating them is a common mistake that makes both harder to improve.

Treat structured process models as proto-world models. A mature M&A reference model — one that maps tasks to goals, classifies problem types, links data objects to decisions, and tracks automation readiness — is encoding, in structured form, how M&A processes behave and what outcomes follow from which actions. The question worth investing in is whether that model can become predictive rather than merely prescriptive. That requires instrumentation: capturing not just what the process says to do, but what was actually done, with what data, and what resulted.

Build deal history systematically. The real long-term opportunity is accumulating enough structured deal data to train actual predictive models for specific M&A subtasks — predicting integration timeline overruns, identifying DD red flag patterns before they surface in full review, or forecasting which synergy categories are most likely to underperform. None of that is possible without deliberate data collection discipline starting now.

Karl Popp leads the M&A Automation Initiative and the Arbeitskreis Digitalisierung within the Bundesverband M&A. He has developed a comprehensive M&A reference model covering the full deal lifecycle, used to assess automation readiness and tool integration across M&A processes.

A Modern Post-Merger Integration Playbook: From M&A Models to AI Solutions
By Dr. Karl Michael Popp

Master integration due diligence to transform your M&A success. Learn more at manda-automation.com