The future of AI in three horizons — and what each one asks of us
Predictions about AI age badly. Horizons age better: what is here now, what is arriving, and what is plausible but unproven. Each demands a different posture from institutions.
It is tempting to talk about "the future of AI" as one thing. It is at least three, and conflating them is how organisations end up over-investing in speculation and under-investing in the systems already making decisions about their customers.
Horizon one · Now: capable, ubiquitous, unverified
General-purpose models are already embedded in customer service, underwriting, hiring, clinical triage and public administration. They are capable enough to be trusted and opaque enough that the trust is mostly unearned. The task of this horizon is verification: knowing, for each deployment, what it does well, where it fails, and who is accountable. Most institutions have not finished this work, and the next two horizons will make it harder, not easier.
Horizon two · Arriving: agents that act, not just answer
Systems that plan, call tools, transact and coordinate with other systems are moving from demonstration to deployment. The governance question shifts from "was this answer correct?" to "was this sequence of actions authorised, bounded and reversible?" Evidence chains must extend to actions: what the agent was permitted to do, what it did, and what it would have done had a control not stopped it.
Horizon three · Plausible: systems that improve themselves
Models that materially contribute to designing their successors are a research reality and a deployment possibility. Whether or not that leads anywhere dramatic, it changes the meaning of a version: an assessment of a system that can revise itself is only valid for as long as it does not. The institutional answer is continuous attestation — proof that the system running now is the system that was tested — rather than point-in-time certification.
The systems will keep changing. What must not change is that someone can prove what they are.
What each horizon asks of us
Horizon one asks for discipline: finish the verification of what is already deployed. Horizon two asks for boundaries: define what an agent may do before it does it, and log what it did in a form a third party can read. Horizon three asks for humility: build assurance as a continuous process, because the object being assured will not stand still.
The good news is that the same infrastructure serves all three. A model registry, independent testing, signed approvals, monitoring bound to the approval thresholds, and cryptographic attestation of what is running — build it for horizon one and it extends to the others.
