Abstract proposals are easy to misread. These pages work through concrete situations where the Databanking model changes the outcome — what information moves, what doesn't, and why it matters.
A retailer needs to know whether a customer is over 18. Step by step through the full query lifecycle — deposit, algorithm submission, sandbox execution, single-bit result, audit trail — with pseudocode and the bit-budget arithmetic.
The UK Online Safety Act requires platforms to check users' ages, and Ofcom is actively enforcing it. Every platform currently runs its own check from scratch. Databanking lets the verification happen once; every subsequent platform gets a yes/no predicate, not the underlying document or face scan.
A taxi operator needs to route a driver to a pickup and drop-off. It does not need a copy of your journey history. A booking token materialises into addresses on the driver's dashboard, is deleted after the ride, and a link token answers audit and discount questions with a yes or a no.
AI has made job applications cheap to write and expensive to read, and recruiters now screen them with AI of their own. Under Databanking a vacancy runs as a registered matching computation over the attested records of Owners who opted in. The Owner receives the offer; the recruiter learns how many matched, and who only when a candidate chooses to respond.
Every hospital, insurer, pharmacy, and research programme currently keeps its own copy of the same medical record. Databanking replaces duplication with controlled access — and lets research consent become a deliberate, per-study choice rather than a blanket waiver.
Training models on personal data raises fears of memorisation, failed anonymisation and models trained from models. Under Databanking the answer starts from what the model is for. General-purpose language models are trained without personal data. Smaller-scope models, such as a credit score or an epidemiological forecast, are trained inside the custodial boundary on the accounts of Owners who opted in and applied to a person as a query, and AI assistants retrieve personal facts rather than memorise them. The trained model is itself a result, and leaves only once it is shown to carry no identifiable information.
To erase your data from a data broker, you currently have to hand that broker the very information you want them to delete. Databanking resolves the paradox: a lossy fingerprint, a sandbox match, and a single yes/no answer — no raw data ever crosses the boundary.
Cloud data is not inherited the way a hard drive is: survivors get an investigation rather than an estate. Inheritance law already knows what should happen — it has nowhere to attach the instruction. Under Databanking the instruction is a clause in the data policy, executed by the custodian once death or incapacity is attested.
Licence-plate cameras turn one narrow investigative need into a population-scale movement database. Under Databanking the warrant authorises a computation instead: it executes inside the custodial boundary and returns only the authorised result — warrant-to-compute, not warrant-to-disclose. Identification becomes a separately authorised step, and some queries need not be possible at all.
The DGA's Chapter III on data intermediation services is the closest real-world precedent. A clause-by-clause comparison of where the two align, where the DGA falls short, and where new legislation would still be needed.