Enterprise AI adoption fails more often than it succeeds. Published estimates place the failure rate of enterprise AI initiatives between 70 and 85 percent, and the majority of those failures are organisational, behavioural, and structural rather than technical. This whitepaper introduces Enterprise AI Failure Models, a simulator that treats an organisation as a coupled physical system. Departments are modelled as nodes, employees as generalised particles with measurable behavioural states, and information as a field that propagates, attenuates, and accumulates friction across the organisational graph. The simulator runs an adoption programme forward in compressed time and reports the most probable failure points before launch. Simulations were generated for a range of organisation types, spanning micro, small, and medium enterprises through to large multi sector organisations. The framework was then tested on a small cohort of eight businesses. Across that cohort the simulator identified the dominant failure channel within the first five iterations in every case, and its recommended mitigations reduced the failure probability it predicted by 41 to 68 percent depending on organisation class. Six-month follow-up found its predictions directionally right in seven of the eight.
1. The problem statement
Organisations have read the case studies, watched the demonstrations and signed the procurement contracts. Six months after deployment, a large share of those programmes have stalled, and leadership has to explain an initiative that did not deliver.
Enterprise AI adoption behaves like a disturbance to a complex adaptive system. An organisation is a network of people, processes, incentives, habits and information flows held in rough equilibrium. A capable new tool shifts that equilibrium: it changes who is valuable, who is exposed, who is overloaded and who is at risk. Those shifts are hard to anticipate without a model, and a failure mode no one anticipated is hard to prevent.
Enterprise AI Failure Models models that shift directly. It assumes that the mathematics of coupled physical systems can describe coupled human systems, provided the parameters are chosen with care and calibrated against observed data. Given a description of an organisation and a proposed adoption plan, the simulator returns a probability distribution over failure modes with the dominant channels ranked.
2. The framework
The framework rests on four claims.
First, a company has structure. Departments, teams, and roles form a directed graph with weighted edges representing information bandwidth, authority, and trust. Second, people inside the company have measurable behavioural state. Workload, confidence in leadership, perceived job security, technical literacy, and openness to change are quantifiable and observable. Third, an AI tool exerts a force on this system. It changes workload distributions, threatens or augments roles, alters information flow, and introduces a new dependency. Fourth, the response of the system to that force is not arbitrary. It follows rules that closely resemble the equations of motion of a coupled oscillator with damping and external driving.
From these four claims the simulator builds three coupled models that run together.
2.1 The organisational graph
The company is represented as a graph G with departments as nodes. Each node carries state variables for headcount, average behavioural profile, dependency on other nodes, and current workload. Edges encode the strength of communication and the direction of authority. This graph is generated automatically from a short structured interview with the operator or from an uploaded organisation chart. The simulator does not require perfect information. It uses Bayesian priors when data is missing and updates them as the operator answers more questions.
2.2 The behavioural particle model
Inside each department the simulator models a representative population of generalised behavioural particles. Each particle carries a state vector that captures five quantities. Workload, trust in leadership (tau), perceived role security (sigma), technical literacy (lambda) and openness to change (omega). These are normalised between zero and one and updated each simulation step according to coupled differential equations driven by the AI tool, by manager behaviour, and by peer effects.
The governing equation for the behavioural response of a single particle to an AI intervention F is given below. It is deliberately simple, on the principle that a model is only useful if it is both tractable and consistent with observed data.
The reading is straightforward. Each person has an inertia m that resists change. Each person has a damping gamma that determines how quickly their behavioural state returns to equilibrium after a perturbation. Each person is coupled to their colleagues through k, which encodes social and operational ties. The three forces on the right are the AI tool itself, the visible behaviour of the manager, and the influence of peers. The simulator integrates this equation across the entire population for the duration of the planned adoption programme.
2.3 The information field
Information about the AI tool propagates through the company like a field. It diffuses with bandwidth, attenuates with hierarchy, accumulates noise with each handoff, and pools in places where trust is high. The simulator solves a discretised diffusion equation on the organisational graph at each step, then uses the local information density as one of the inputs to the behavioural model. This mechanism accounts for a pattern most change managers have seen: the same AI rollout produces enthusiasm in one department and resistance in another. The tool is identical in both cases. The information field around it is not.
3. Why a simulator and not another framework
Most AI adoption frameworks on the market are checklists. A checklist says what to do; it does not say what will happen if you do it. A simulator answers the second question. It runs a proposed plan forward in compressed time and produces the resulting trajectory. A trajectory that stabilises supports proceeding, while one that collapses signals that the plan should be revised and run again before any commitment is made. Engineering disciplines that manage complex systems already work this way: aerospace simulates flight, civil engineering simulates load, and drug development models pharmacokinetics. Most enterprise AI rollouts have no equivalent step.
4. The failure modes
The simulator currently identifies fourteen named failure modes. In the simulated organisations and the test cohort, six of them account for over eighty percent of predicted failures:
Trust collapse. The workforce concludes that the AI tool is unsafe or that leadership has not been honest about its purpose. Adoption stalls and reverses. Workload inversion. The tool intended to reduce work increases it, because the cost of supervision exceeds the cost of doing the work manually. Skill atrophy. Frequent users stop developing the underlying judgement the tool was supposed to support. Quality degrades silently. Manager passivity. Middle managers do not model the desired behaviour. The workforce reads this as a signal that adoption is optional. Information starvation. The tool is rolled out without the context, examples, and feedback channels needed for users to understand it. Integration debt. The tool is bolted onto existing systems instead of woven into them. Friction accumulates until users stop bothering.
Each failure mode has a measurable signature; trust collapse looks different from manager passivity in the data. So the simulator names the kind of failure to expect, which tells an operator what to fix, and not only that the programme is at risk.
5. Simulation across organisation types
Before any live engagement, the simulator was run across a wide range of organisation types to see how its behaviour varies by class. The simulated organisations spanned micro, small, and medium enterprises through to large multi sector organisations, with modelled headcounts from seven to roughly eighteen thousand and sector profiles drawn from financial services, agriculture, tourism, logistics, telecommunications, public sector services, manufacturing, and the creative industries.
The simulations established that the dominant failure mode is a function of organisation class. In micro, small, and medium enterprises the leading predicted failures were manager passivity and workload inversion. This is consistent with the structural reality of a small enterprise, where the founder is also the manager, managerial capacity is already saturated, and there is little slack to absorb the supervision cost of a new tool. In large organisations the leading predicted failures shifted to trust collapse and integration debt, followed by information starvation. That pattern reflects legacy systems, layered communication, and workforces that have lived through earlier failed transformations and approach new tools with informed scepticism.
The mitigation profiles differ accordingly. Smaller enterprises benefit most from a phased rollout, a single dedicated AI champion, and explicit reassignment of recurring tasks away from the founder. Large organisations require honest internal communication, visible executive use of the tool, deliberate investment in integration with systems of record, and a measurement layer that publishes adoption metrics back to the workforce.
6. Early test on a cohort of eight businesses
The framework was then tested on a cohort of eight businesses that agreed to run their planned 2026 AI initiatives through the simulator before launch. The cohort mixed micro, small and medium enterprises with larger organisations across several of the sectors above, so the test covered more than one organisation class.
For each business the simulator was run first on the baseline adoption plan and then on a mitigated plan that incorporated its recommendations. The recommended mitigations reduced predicted failure probability by 41 to 68 percent, with the larger reductions concentrated in the smaller enterprises and the more modest reductions in the large organisations, where the ceiling on what a single intervention can achieve within six months is harder. Across the cohort the dominant failure channel was identified within the first five simulation iterations in every case.
Follow up over the subsequent six months found the simulator to be directionally accurate in seven of the eight businesses. The single deviation occurred at a business that introduced an additional AI tool that had not been represented in the original simulation, which moved the system outside the modelled scenario. Eight businesses is a small sample, and these results are early evidence, not a calibration. They suggest the predictions track observed outcomes across organisation classes. Whether the recommended mitigations lower real failure rates, as opposed to predicted ones, needs a larger cohort with a control group, which is the next stage of the work.
7. What the simulator will not do
The simulator is not a substitute for leadership judgement, and it does not understand an organisation's culture the way a long tenured executive does. It does not capture rare exogenous events such as a sudden regulatory change or a high profile incident at a competitor. It is a forward-looking estimator with stated uncertainty, not an oracle. Use it the way an aerospace engineer uses a flight simulator: as one input into a decision the organisation still owns.
8. How to engage
Enterprise AI Failure Models is being released in a limited preview through StarApple AI and the IMPACT AI Lab. We are accepting applications from MSMEs and from large organisations that want to run their planned 2026 AI initiatives through the simulator before launch. Priority is given to organisations operating in the Caribbean and the wider Global South, where the cost of a failed AI initiative is disproportionately high and the upside of a successful one is disproportionately large.
Current AI models can already change how organisations operate. Most of the time they sit unused for human, organisational and structural reasons, and those dynamics follow patterns that can be measured and modelled. The simulator applies to an organisation the method a scientist applies to any system: measure it, model it, perturb it deliberately and learn from the response before committing to the rollout.
9. What to do next
- The executive sponsor: before signing an AI contract, write down the adoption plan in one page: which teams, which tasks, which managers own the rollout, and the week-by-week timeline. The simulator, or any serious review, needs that plan as input.
- HR or the change lead: run a short anonymous survey of the affected teams on workload, trust in leadership, perceived job security and confidence with AI tools. These are the behavioural inputs the model uses, and the baseline is useful even without it.
- Middle managers in the pilot teams: use the tool visibly on your own work in the first two weeks. Manager passivity is the leading predicted failure in small and medium enterprises.
- The founder of an SME: name one person other than yourself to supervise the tool day to day and move at least one recurring task off your desk to them, so supervision cost does not land on you.
- Large organisations: publish adoption metrics back to staff monthly and budget integration with systems of record from the start. Trust collapse and integration debt lead the predicted failures at that scale.
- Organisations in the Caribbean or wider Global South planning a 2026 initiative: apply for the limited preview through StarApple AI or the IMPACT AI Lab at least eight weeks before launch.
Frequently asked questions
What information does an organisation need to provide to run the simulator?
An organisation chart or a short structured interview describing departments, reporting lines and headcounts, plus the proposed adoption plan: which tool, which teams, and the rollout timeline. Behavioural inputs such as workload and trust can come from a staff survey; where data is missing, the simulator uses priors and flags the uncertainty. More complete inputs narrow the range of predicted outcomes.
How is this different from an AI readiness assessment?
A readiness assessment scores an organisation against a checklist at one point in time, for example data quality, skills and governance. The simulator takes a specific rollout plan and runs it forward week by week to show where and when it is most likely to break. The two work together: a readiness score describes the starting state, and the simulation tests a plan built on it.
How accurate are the simulator's predictions?
In the test cohort of eight businesses, six-month follow-up found the predictions directionally accurate in seven. The one miss came from a business that added an AI tool the simulation had not included. Eight businesses is too few to quote a reliable accuracy rate, so treat the output as a structured second opinion, not a forecast to bet the programme on.
Can a small business apply the failure modes without running the simulator?
Yes. Check your plan against the six common failure modes: trust collapse, workload inversion, skill atrophy, manager passivity, information starvation and integration debt. For each, name the early warning sign you will watch and the person responsible for acting on it. For most small firms, manager passivity and workload inversion are the two to watch first.
How is staff data handled when behavioural surveys are used?
Survey responses should be collected anonymously and reported only at team or department level, so no individual can be identified. In Jamaica, personal data processed for this kind of work falls under the Data Protection Act 2020, which came into force in December 2023, and similar laws apply elsewhere. Tell staff what the survey is for before they answer, since trust is one of the variables being measured.
"Every failed AI rollout I have ever seen was predictable. The simulator just makes the prediction explicit, in time to do something about it." Adrian Dunkley, AI Boss