Why the bot has to transfer you to a human
A woman rings her lender because the payment date no longer suits her, and the voice which answers is not a person but a program, polite, well informed, and more patient than any clerk could afford to be. Then she says what she wants, which is twelve thousand over nine months rather than fifteen over six, and the program tells her it will pass her to a colleague. She waits. Somewhere a person picks up the transcript and begins again from the beginning.
The same thing happens when a policyholder asks to add a second dog to her cover, and when a supplier disputes an invoice and wants a credit note rather than an explanation. The program can answer and it cannot agree, so everything easy is handled with grace and everything which carries money is handed to somebody else.
It is tempting to say the program lacks judgement, but language models negotiate rather well, and they read a muddled request, half apology and half proposal, better than any form ever did. What they lack is permission of a particular kind: a settled understanding of what the program may agree to, up to what amount and on whose behalf, without which no company would be wise to let it agree to anything at all. The understanding itself is old. Every company already gives one to its people, and calls it a mandate.
The mandate
A mandate is the room within which a person may settle a matter on her own judgement, and the edge beyond which she must ask. The underwriter has an authority above which the risk goes to a senior colleague, the account manager may discount to a certain percentage and not beyond, the collections caller works from a ladder with a ceiling for each size of debt, and the civil servant acts under delegated powers which state the decisions she may take without her minister. In each case the limit is drawn so that the person inside it can be trusted to think for herself, and that, rather than restraint, is its purpose. A mandate is not a script for the conversation but a boundary around what the conversation may conclude.
This is also what a business means by the word policy, before software borrowed it and narrowed it to a question of access: a plan of action, saying what may be agreed, with whom and up to what limit. Give the program the same mandate its human colleague works from and the transfer becomes unnecessary, for every request either falls within the limits or it does not. If it does, the program agrees at once and changes the record. If it does not, it says why, or offers the nearest arrangement which does fit, which is very often what the customer would have accepted in the first place.
What may be governed
The pattern is not peculiar to lending. It belongs to any object which has a life of its own and carries money or an obligation with it, and that is most of the records a company keeps. A loan, a lease, a subscription and an insurance cover are contracts whose terms may be changed. A claim, a refund, a credit note and a settlement are decisions which release money. An order, a return, a permit application and a benefit claim pass through a sequence of states, from requested to approved to paid or refused. For each of them the policy is the same kind of document, saying which state may follow which, who may move the object from one to the next, up to what amount, how often within a period and on what conditions. That is all a policy is.
What arrives at the door is never so orderly. It comes as language, in an email, in the transcript of a call, in a form with a box for free text, in a photograph of a veterinary bill taken on a kitchen table, and reading that and turning it into a proposed change to a record is work a language model does well. Holding the policy, checking the proposal against it and writing down what was decided is work of a different kind, better done by something which cannot be persuaded. The model takes the mess, and a control layer keeps the policy and a book of everything done under it, kept the way a bank keeps its books: every transaction is posted to it, and everything is reconciled against it to the cent.
Where the limits live
For any of this to be safe, the limits must live somewhere the model cannot reach them. A limit written into the model's instructions is a request rather than a rule, and a determined customer can talk a model out of a request in a few messages. So the limits are held in a separate layer, one the model can propose to but cannot edit, and that layer takes its facts from the records rather than from the conversation: the balance from the loan, the cover from the policy, the flags from the customer's file, the day's running total from a counter which every program shares. The proposal is set against those facts before the record is touched, and it either passes or it does not. Either way there is a receipt, naming the rule.
The model does not confirm that it has kept the rules. The rules are applied to what the model proposes, by something it cannot reach, and that is the difference between a guideline, which a model may or may not follow, and a control, which it cannot evade.
Where people come in
Some requests will fall outside the mandate, and what happens then is where a well-made system shows itself. A customer may ask for more than the limit allows, and the program neither refuses nor pretends but carries the proposal to the person named as approver, with the record, the limit and its own reasons attached, for her to decide from wherever she happens to be working. A customer may ask for something the mandate never contemplated, a change of currency or a co-signer, and the matter goes to a person who can decide, with the transcript. And a customer's file may carry a flag, of hardship, of dispute or of vulnerability, in which case the program is stopped before any offer is made, because these are the circumstances in which a person's situation matters more than any band on any ladder.
In every case a human being sees only the exceptions, and sees them with their context already assembled. The alternative, a queue holding every request, is not a control at all. Nobody reads it.
When the limits move
Nor is the mandate fixed. A discount ceiling is raised, an underwriting limit is reviewed, a settlement band is widened after some argument between a lender and the firm which collects for it, and with a program in place only the manner of the change is new. One person proposes the revised policy and another approves it, the next request runs under the new version, and the receipts written earlier that day keep the old one, so that an auditor can see without asking anyone which decisions were made under which limits. Nothing is redeployed. The counters carry on as though nothing had happened, which for a daily total is exactly right.
The same question, asked of the law
The question is no longer only a private one. For consumer lenders in the European Union, the revised Consumer Credit Directive requires a creditor to exercise reasonable forbearance before it begins enforcement, with the individual customer's circumstances taken into account (Directive (EU) 2023/2225, Article 35, applying from 20 November 2026). It names the forms forbearance may take, from extending the term to forgiving part of the debt, and then draws a boundary of its own: the terms may be changed without a fresh assessment of creditworthiness, provided that the total the customer will pay does not significantly increase. A list of permitted changes and a limit on their sum is a mandate, and here it has been written by a legislature.
The state has begun to ask the same question of itself. It has been proposed in Estonia that the AI agents acting for its people and companies be given a legal identity of their own, so that it is always known on whose behalf an agent acts and with what legal powers (Delfi juhtkiri: Mineviku poole õhkavat Eestit ootab vaesus ja paigaltammumine). Legal powers are a mandate under another name, and an agent which files a permit for a citizen or claims a benefit on her behalf needs what the collections program needs: a defined set of things it may do, a limit on each and a record of every occasion on which it did one.
So the woman on the telephone gets her nine months, or eight, and a record of why. The lender gets a program which can say yes as well as explain, and a book in which every yes and every no is written down with the rule which decided it and the version of the policy in force at the time. The mess of the request, the half sentences and the photograph of the bill, stays with the thing which is good at mess, and what was decided, and under what authority, goes in the book: stategram.io
