The three fates of AI: Convergent, divergent, deviant The three fates of AI: Convergent, divergent, deviant The same AI can simultaneously stimulate and diminish human interaction and coordination, depending on how its possibilities are used. By Billy Mashele, Lecturer in the Department of Construction Economics and Management at the University of Cape Town. Billy Mashele, lecturer in the Department of Construction Economics and Management at the University of Cape Town. When it comes to the discourse on the fate of artificial intelligence (AI) affordances, I often argue that if you give the same AI tool to a thousand employees, it does not have one fate; it has three. In some hands, its use converges: everyone actualising the same capability in the same sanctioned way, coordination knitting tighter. In others, it diverges: different teams bending the same tool to different legitimate ends, some visible, some quietly off the radar. And in some, it goes deviant: used in ways the organisation’s own policies mark as undesired, by people who are not saboteurs but problem-solvers. Boards ask for adoption numbers as if uptake were a thermostat you could set once. It is not. These three fates are produced by conditions and gates the organisation owns, and the difference between an AI programme that compounds and one that leaks is whether anyone is managing the fates on purpose. One critical line from affordance research is worth regurgitation: some action possibilities are intended, designed into the tool; the more consequential ones are emergent, discovered in use. Convergence is largely the management of the intended; divergence and deviance are what emergence looks like on a payroll. In my previous article in this series, I argued that an affordance is not a feature but a possibility for action, living in the relation between a technology and particular people under particular conditions. That piece asked what opens possibilities. This one asks what happens to them next, because a possibility, once alive in an organisation, has a fate. Adoption of the new tools is uneven, because using them well is not a skill alone; it is skill compounded by attitude and habit of mind. The vocabulary for the fates is newer than the phenomenon. My ongoing research on AI affordances in organisations reframed the distinction between convergent and divergent AI affordances to explain something managers keep observing and dashboards keep missing: the same AI can simultaneously stimulate and diminish human interaction and coordination, depending on how its possibilities are taken up. A landmark MIS Quarterly study gave the third category its name and its crucial nuance: deviant affordances are action possibilities whose actualisation the organisation’s own IT policies specify as undesired, yet which goal-oriented actors actualise anyway, because doing so serves both an individual goal and at least one legitimate organisational goal. The dual-goal clause changes everything. Deviance, in the research sense, is not malice. It is unmet legitimate need finding a route the policy did not provide, which is why research on shadow IT and workarounds keeps finding such use persistent and performance-bearing rather than aberrant. Inside a large financial institution that embraces a human-in-the-loop philosophy for AI integration, a surprising realisation emerges: convergence is engineered, not natural. Where AI programmes converge, someone has refused to let a hundred assistants bloom, concentrating use instead onto productionised, governed capability that can be reused; demonstration circuits and standing forums have converged practice by showing rather than mandating; and risk gates, engaged upstream, have shaped solutions towards sanctioned patterns before they were built. Where those conditions hold, the same AI affordances concentrate coordination, exactly as the convergent pattern predicts. Where they are absent, no slogan produces them. Divergence is equally real and, this is the part transformation decks never show, equally valuable. Citizen-developer platforms show the pattern at scale: the same platform affordance actualised thousands of different ways by businesspeople solving their own problems. Adoption of the new tools is uneven, because using them well is not a skill alone; it is skill compounded by attitude and habit of mind. The value splits too: what an assistant affords one team as personal time-saving, another converts into measurable business outcomes. Divergence is where innovation actually enters an organisation; it is the intake valve for the second-order changes that eventually restructure the place. Those changes cannot be scheduled. Joseph Schumpeter distinguished an adaptive response, which stays within existing practice, from a creative one, which goes beyond it, and warned that the creative kind “can practically never be understood ex ante”. A programme that funds only the divergence it can predict is funding adaptation and calling it innovation. But divergence is also where the map fades. The same affordances, under governance conditions in opposite states, either concentrate coordinated work or disperse it into shadow appropriation. Divergence with conditions is an innovation engine. Divergence without them is how you lose sight of your own estate. Which brings us to deviance, and the two forms it takes. The classic form rides the democratisation wave: as tools spread into capable hands, people appropriate them to their own ends, sanctioned and unsanctioned alike, and the unsanctioned actualisations behave precisely as the deviant-affordance research predicts. Policy says no; a legitimate need says yes; the need wins. The attachment runs deeper than need: experiments published this year find people forming psychological ownership of conversational AI through control, intimate knowledge and investment of self, so retiring a workaround can feel to its owner like confiscation. The second form is stranger and, for governance, more important: deviance by supply. Vendor platforms ship AI features enabled by default, so that unsanctioned use exists before any user has even decided to deviate, and features built to engage ship the attachment with the default. Both forms teach the same lesson. Policy perimeters age like the street maps of a fast-growing city: the capability keeps building suburbs, and people and platforms end up living in the gap. Here is the point I have not seen anywhere in the practitioner conversation, and it is the reason this typology earns a place on the management agenda: deviant actualisations have a fate, and the organisation chooses it. Ungoverned use arrives at a selecting gate, the risk and architecture machinery through which everything must pass, where it meets one of two futures. Either it is re-governed, legitimised and brought onto the platform, seeding genuine structural change: today’s workaround becoming tomorrow’s sanctioned capability. Or it is wound back by the organisation’s antibodies: retired, closed, absorbed into the existing order. Deviance, in other words, is not a leak. It is a tributary, and the gate’s conversion rate, not the deviance rate, is the number that decides whether nonconformance generates performance or merely risk. Which future it meets is settled as much by culture as by policy. Most large organisations inherit two ideas they can neither drop nor merge: a control doctrine their regulator requires and an experimental ethic their AI mandate requires. Where both must stand, the gates that work are reformed rather than removed: checks built into the platform, and risk recast as a partner rather than a policeman. Fears follow the same logic. A worry about one’s job, answered by reskilling, can sit beside adoption; a campaign to stamp the worry out treats a concern the organisation needs as an obstacle to remove. Scholars call generative AI Janus-faced, with promise and peril issuing from one artefact; the fates are where the faces separate, and they separate on management, not on the model. So, manage the fates deliberately, as three different jobs. Engineer convergence where coordination and reuse carry the value: platforms over proliferation, demonstration over mandate, gates that shape early. Protect divergence as your innovation arsenal: enable capable hands, expect and budget for uneven adoption, and resist the urge to standardise the exploration phase into sterility. And treat deviance as information with a routing rule: when unsanctioned use appears, apply the dual-goal test (what legitimate organisational need is this use serving that our policy currently blocks?), then send it to the gate for an explicit verdict: legitimise or retire, rather than to the disciplinary process. And instrument all three, because the research community has a confession worth hearing: audits of the field’s affordance studies find that most never say where a use was born; the birth of an affordance is an opaque black box. With AI, the box is doubled, because the model’s own interior is opaque too. Interpretability researchers are prising that second box open, but models do not reliably report their own reasoning, so measure what the estate does, never what a model says it did. Organisations repeat the omission at scale. Naming and counting the fates opens the box: a reuse ratio for convergence, variance of use for divergence, and for deviance, two numbers: incidence and gate conversion rate. The industry numbers say the clock is running. IBM’s 2026 breach study found shadow AI implicated in 43% of breached organisations’ incidents, more than double the year before; Netskope’s 2026 threat report shows personal-account use of generative AI tools falling from 78% to 47% as managed alternatives arrive; and in OutSystems’ 2026 survey, 94% of organisations worry that AI agents are multiplying faster than governance can follow. Convergence and deviance are rising together. That is not a contradiction; it is the typology working. The deepest point is the one the affordance lens always makes: none of these fates lives in the technology. The same tool can converge, diverge and deviate inside a single organisation in the same year, because fate is decided by conditions, communities and gates, all of which someone owns. Treat the estate as a set of things, and you will only inventory it; treat it as a flow of possibilities being born, travelling and meeting gates, and you can govern it. Your AI estate is running this experiment right now, whether or not anyone is watching. Convergent, divergent, deviant: name the fates, count them and choose them, or they will quietly choose themselves. And the stakes are rising on schedule: the task horizons of frontier systems have been doubling roughly every seven months, and last September’s launches, led by OpenAI’s GPT-6 Astra, an agent sold on operating your software end-to-end, add a deeper shift. When an agent actualises possibilities on your behalf, the actor in the affordance relation is no longer only human, and every fate named here acquires a delegated variant arriving at machine speed. The organisations already counting the human fates will be the only ones that can see the machine ones. * Billy Mashele lectures at the University of Cape Town, where he researches technology adoption and organisational change in financial services. He writes here in his personal capacity. Policy & Procedure Management – Better Governance through efficient group-wise policy management- Governance Platform | msg compliance Sigfox South Africa delivers reliable nationwide 0G IoT connectivity enabling scalable, low-power AI-driven smart solutions. Policy & Procedure Management – Better Governance through efficient group-wise policy management- Governance Platform | msg compliance