Prediction is a mug’s game, but some futures announce themselves early if you know where to look. The place to look right now is any industry where AI outputs meet auditors, and the announcement is unambiguous: the next era of artificial intelligence will not be defined by what machines can conclude, but by what they can prove. Call it the receipts era.
The clearest early build of that future is running today in American healthcare, and it is worth a futurist’s close attention, because the pattern it establishes will be everywhere by the decade’s end.
The preview industry
The setup: US insurers covering older adults are paid according to risk scores computed from members’ documented diagnoses, and AI now does much of the reading that produces those diagnoses, parsing years of clinical notes per patient. The money is enormous, the temptation to over-document proved irresistible, and the correction arrived in force. Federal auditors, roughly two thousand of them, now re-check the records on a quarterly cycle, with error rates from samples extrapolated across entire contracts. Spring 2026 reviews found 81 to 91 percent of certain sampled codes unsupported at three plans; one major insurer settled with the Department of Justice for 117.7 million dollars over how its records were assembled.
Watch what that pressure selected for. The AI systems now winning in that market are not the ones with the most impressive raw extraction. They are the ones whose every conclusion ships with evidence: the source sentence in the clinical note, the explicit rule it satisfies, the confidence score, the identity of the human who confirmed it, and a stored trail that lets a hostile examiner reconstruct the decision years later. Practitioners call the standard audit-ready documentation, and it has quietly become the spec that separates deployable AI from demo AI in that industry.
That spec is the future, generalised.
Why receipts win everywhere
Three converging forces make the receipts era close to inevitable.
Regulation is converging on traceability. The EU AI Act demands logging, transparency, and human oversight for high-risk systems. Financial regulators expect model decisions explainable to examiners. Medical device frameworks, employment law, insurance codes: jurisdiction by jurisdiction, the same requirement is crystallising. Nobody is legislating “be accurate.” Everybody is legislating “be able to show your work.”
Liability is finding the gaps. Every consequential AI error now produces the same discovery question: what did the system know, and can you reconstruct why it concluded what it concluded? Organisations that answer with logs settle small. Organisations that answer with shrugs settle large. General counsels have noticed, and general counsels move procurement.
Agents raise the stakes. As AI shifts from answering questions to taking actions, booking, buying, filing, prescribing workflows, the tolerance for unexplainable behaviour collapses. An agent that acts on your behalf without an inspectable decision trail is not an assistant. It is an unindemnified employee. The agentic future everyone is building presupposes the receipts infrastructure almost nobody is talking about.
What the 2030 stack looks like
Extrapolate the healthcare build-out and the standard enterprise AI stack of 2030 comes into focus. Perception layers, large models reading the messy world, wrapped in validation layers that check outputs against explicit, versioned rules. Evidence stores holding immutable source data, with lineage metadata binding every inference to its inputs. Deterministic replay: model versions, rule versions, and data snapshots pinned so any past decision can be reconstructed under the logic of its day. Human checkpoints recorded as first-class events, not workflow decoration. And correction instrumented in both directions, because auditors have learned that systems which only ever err in their owner’s favour are not erring at all.
None of this is speculative hardware. Every component runs in production today in the industry that got audited first. The futurism is merely in the spreading.
The cultural turn
The deeper shift is in what we will mean by trusting a machine. The first AI decade asked for faith in benchmarks: the model scores well, believe it. The receipts era replaces faith with inspectability: believe nothing, verify anything. It is a colder standard and a far healthier one, the same standard we eventually imposed on accounting, aviation, and pharmaceuticals once their outputs mattered enough.
There will be resistance, mostly on cost. Evidence infrastructure is expensive, and for a while, unaccountable AI will remain cheaper and faster. Then the settlements, the clawbacks, and the procurement checklists will do their patient work, as they are doing in healthcare right now, and the market will discover what regulated industries always discover: proof is cheaper than penalty.
So here is the prediction, dated and falsifiable. By 2030, “can it explain itself to an auditor?” will be the first question asked of any consequential AI system, ahead of accuracy, ahead of cost. The machines will still astonish us. They will simply do it with their receipts attached, because we finally asked. The industry that learned it first has the settlement invoices to prove the lesson was worth it.
