AI is now arriving in mid-market companies in a different way than it did three years ago. The conversation is no longer about pilot projects and proofs of concept. It is about real workflows: predictive forecasting, automated reporting, decision-support layers in commercial and operational systems. The technology is becoming infrastructure, not novelty.
This is, on balance, useful. Faster reporting, deeper modelling, broader pattern detection — these are real capabilities, and they will compound over the next several years. But they do not address the constraint that determines whether a company actually moves: who is accountable for deciding, on what evidence, with what cost of being wrong. AI changes the inputs. It does not change the responsibility.
1. More information does not produce more decisions
The first observation is structural. In most companies under pressure, the binding constraint is not the absence of information. It is the absence of someone willing and able to act on the information that already exists. Adding faster, richer, more granular signals to that situation does not produce action. It produces more material to be reviewed, more variance to be explained, and often more reasons to wait for the next refresh. Reviewing the new information is not free. It absorbs the same finite decision energy that should have been spent acting. In organisations already slow to act, additional information competes with action, rather than enabling it.
The Board that could not decide on a quarterly report cannot decide on a real-time dashboard. The CFO who could not move on monthly variances cannot move on weekly ones. The CEO who deferred on imperfect information will defer on better information, because the binding constraint was never information quality. It was the cost of the decision.
2. AI shifts where accountability sits, but it does not establish it
When a model produces a forecast, an alert, or a recommendation, the accountability question becomes more interesting, not less. Who is accountable for acting on the output? Who is accountable for the model's assumptions? Who is accountable when the recommendation turns out to be wrong? Who is accountable when it is ignored?
Recent academic work has begun to formalise the difficulty. Strategic decisions are characterised by incompleteness and irreducible discretion — they cannot be closed by a model, only informed by one. A second line of research describes an inferential trilemma: when an AI output is non-obvious, the organisation has to distinguish between genuine breakthrough, hallucination, and misalignment. Each of those three readings calls for a different action, and only a human in a defined accountability seat can make the call. The model does not free anyone from that seat. It makes occupying it more demanding.
In well-governed organisations, these questions have answers. The model is a tool; a named person owns the decision the tool informs; the Board has visibility into both. In organisations with weaker governance, AI produces a different pattern. The model becomes a layer that distributes responsibility — the recommendation came from the system, the system was supplied by a vendor, the vendor was approved by IT, and no one is squarely accountable for the action that should have followed. In the post-mortem of a missed decision, each actor can point to another layer. The architecture itself produces a denial of accountability by design. The AI Act's emphasis on human oversight is a regulatory response to exactly this drift.
3. The fastest dashboard does not solve a slow Board
The third pattern is a corollary of the first two. The technical capability of the reporting layer often outpaces the decision-making capability of the governance layer above it. Variances can be detected within the day. Action on those variances may still take weeks, because the structure that owns the action has not been updated to match the speed of the signal.
The result is a particular kind of disorientation. The data layer behaves as if the company is a high-frequency operation. The decision layer behaves as if it is a quarterly one. The gap between them widens with each new piece of capability, and the Board increasingly receives information that it cannot, in practice, act on at the speed the information itself implies. The mismatch is also visible to outsiders. Lenders, sponsors, acquirers see the dashboard and start asking questions at the cadence the dashboard implies. The company has equipped itself with a signalling layer it cannot match in execution. This does not produce paralysis at the technical level. It produces it at the governance level, more visibly than before.
What actually has to change
The companies that benefit from AI in operational contexts share an unfashionable characteristic: they had clear accountability before the technology arrived. The model produces a signal, a named person decides what to do, the Board sees both, and the cycle is short enough that the action can still matter. The technology compresses the loop. It does not invent the loop.
In our experience, the diagnostic question for any AI deployment in a stressed mid-market company is not whether the model is good. It is whether the company would have acted on a clear, manual signal in the same situation. If the answer is no, the AI deployment is solving for the wrong constraint. If the answer is yes, the AI compounds an existing capability. There is rarely a middle case.
Compliance frameworks — the AI Act in Europe, sectoral guidance elsewhere — are now formalising what was always true. The European AI Act's main governance and general-purpose-AI obligations apply from August 2025, with the bulk of the operational rules from August 2026. From that point on, accountability around AI is no longer a question of innovation policy. It is part of ordinary corporate governance. The question is not what the system recommends. The question is who, in this organisation, is responsible for what happens next. Better tools do not change that question. They make the answer more urgent.
---
References
- AI and Strategic Decisions: Facing the Incompleteness, European Business Organization Law Review, January 2026.
- Replace, augment, disrupt: AI & organizational decision-making, Journal of Organization Design, October 2025.
- Timeline for the Implementation of the EU AI Act, European Commission AI Act Service Desk, 2026.