A support reply is a claim about something the organisation possesses. The returns window, the stock position, the delivery date, the address an order went to, the terms of a plan: each exists in a document the organisation wrote and maintains, and each becomes expensive at the moment a sentence in a mailbox and a line in that document disagree. The failure is not usually a badly written sentence. It is a true-sounding sentence about a fact nobody checked.
That is why the interesting artefact in support is not the reply generator. It is the answer bank: the place where the organisation writes down, once, what it will say in answer to a question it has been asked before, and records the passage that answer stands on. A bank of that kind is a small thing — a table, a folder of snippets, a page per topic — and it is the only object in the arrangement that can be reviewed before a customer reads anything. The model’s job downstream of it is the sentence. The bank’s job is the fact.
This piece is about that artefact: what goes into it, how a citation attaches from the reply back to the source document, and what the agent does when the corpus holds no answer at all. The last of those is the part most often left undesigned, and it is the part that decides whether the arrangement improves or merely produces fluent copy at volume.
The row is the unit
An answer bank is not a prompt and it is not a vector index. It is a set of rows, each of which claims that a recurring question has a settled answer and names the document that settles it. Whether the rows live in a spreadsheet, a help-centre article, or a directory of Markdown files is an implementation detail; what makes the object usable is that each row carries the same nine fields.
Table 1. The nine fields of an answer-bank row, and what the absence of each produces in the replies built from it. Source: constructed for this piece; the fields are a design, not a product specification.
| # | Field | What it holds | What its absence produces |
|---|---|---|---|
| 1 | The question, in the customer’s words | Two or three phrasings as they actually arrive, alongside the organisation’s own vocabulary | A bank searchable only in the words the organisation uses, which misses the customer who asks it differently |
| 2 | The canonical answer | The answer in the house voice, with a stated length and no optional clauses | An answer assembled from scratch each time, which is where tone drifts and half-promises appear |
| 3 | The passage it stands on | Document identity, section anchor, and the quoted span the answer paraphrases | An answer that rests on the model’s recollection of how companies usually answer |
| 4 | The document’s version and date | The version retrieval actually opened, not the version the author believed was current | A citation that resolves to a page that no longer says it |
| 5 | The owner | A named person or role who is accountable for the row being true | Nobody notices when the product, the price or the policy changes |
| 6 | Last verified | The date a person last read the passage against the row | A row that ages quietly and is trusted precisely because it is old |
| 7 | The promise line | What this answer may commit to, and what it may not — explicitly | A reply that reads like a policy when it was meant to read like an explanation |
| 8 | The scope note | Who the answer is not for: other product tier, other region, other plan, other version | The right answer given to the wrong customer, which is indistinguishable from a wrong answer |
| 9 | The escalation route | Where the case goes when the row does not fit, and who owns it | An agent that improvises rather than handing over |
Nine fields is more than most banks have, and the surplus is deliberate. Fields 5, 6 and 8 are the ones that turn a knowledge base into something a reviewer can audit; fields 3 and 4 are the ones that turn an answer into a citation. A bank with only the first three is a FAQ, and a FAQ is what most organisations already have: accurate on the day it was written, unowned thereafter, and consulted less often than the model.
What goes in, and what is kept out
The corpus is the organisation’s own documents; the bank is the layer above it where the recurring answers are written down and pointed at a passage. Six classes of material belong in the corpus, and it is worth naming them because a corpus assembled by grabbing the intranet usually contains four of the six.
Product facts. What the thing is, what it does, what it does not do, which versions exist and what changed between them. This is the class most often thin, because product documentation is written for buyers and support answers questions asked by owners.
Policy commitments. Returns, warranty, delivery windows, cancellation, data retention, refund eligibility. Every sentence here is one the organisation will be held to, so every sentence needs an owner who can change it.
Process states. What happens after an order is placed, what a tracking status means, where a ticket stands, what the customer should expect next and when. The useful form is a state list with the expected duration attached, not a narrative of the journey.
Prices and terms. List prices, plan inclusions, what is metered, what is included in a service call. Prices move, which is why a price row without a version and a date is a liability rather than an asset.
Troubleshooting as decisions. If the light is red, do this; if it stays red after that, do that. The value is in the branch, not the prose, and a decision the customer can follow beats a paragraph they must interpret.
Refusals and out-of-scope statements. What the organisation does not do, will not advise on, and cannot answer — written down in the same voice as everything else. A corpus that contains only what the organisation can do forces the agent to invent the boundary.
Five classes stay out, and the test for each is the same question: would a named person in this organisation sign their name under this sentence if a customer quoted it back? Anything that fails that test is not corpus material, whatever its other merits. Out go opinions and predictions, including the ones in the marketing deck; competitor comparisons, which date faster than prices and carry legal exposure; superseded documents, which should be removed from the index rather than archived inside it, because retrieval will find them; personal data of any kind, since the corpus is shared and a customer’s case is not; and drafts of policy that nobody has approved, which are the most dangerous class of all because they are written in exactly the register of a commitment.
The bank’s coverage is a separate question from the corpus’s completeness, and the difference matters for what a reviewer can conclude. A complete corpus with a thin bank produces correct replies written afresh every time — expensive, inconsistent, and unverifiable in advance. A full bank over a thin corpus produces confident citations to documents that do not answer the question. Neither is a state to aim at, and neither is visible in a count of documents.
How the citation attaches
A citation has three parts and all three have to be present for it to be checkable. The first is the document: a name or identifier that resolves inside the client’s corpus, not a URL that will move. The second is the location: a section anchor, a heading, a clause number — something a person can navigate to in the time it takes to open a file. The third is the span: the sentence or clause the answer paraphrases, quoted at its first and last few words.
The row holds all three; the reply carries the sentence; the retrieval set carries the citation. That division is the whole design, and it has one consequence worth stating plainly: a reply that cites the answer bank cites nothing. A citation to the bank proves only that the bank says so. The chain has to run through to the document, because the document is the thing with an owner outside the support function and a change history that can be read. Where the reply carries a visible attribution — “this is covered by our returns policy, section 4” — the attribution is for the reader’s benefit; the checkable citation for the reviewer is the passage in the retrieval set.
Version is the part that quietly breaks. Retrieval opens a version; the row was written against a version; policy documents are revised without anyone telling the support desk. A row whose field 4 names a version the document no longer carries will produce a citation that resolves to different words, and the reviewer who opens it will find a plausible page that does not say what the reply says. The repair is mechanical and cheap — verify the version at the moment of drafting, and record the version that was open — and it is the single highest-yield act of maintenance in the arrangement.
The second artefact, the retrieval set, is worth naming because it is what makes supervision cheap. For each reply it records what was searched, what came back, and what came back empty. A reviewer with a retrieval set is verifying; a reviewer without one is re-doing the work, and an organisation that asks people to re-do the work will stop reviewing. That is the economic argument for the citation, and it is stronger than any argument about accuracy: the citation does not make the reply true, it makes the reply checkable in a minute instead of ten.
The null result
Every corpus reaches a question it cannot answer, and what the arrangement does in that moment is the difference between a system that improves and one that fails fluently. There are three nulls, and they are not the same event.
Nothing retrieved. The search returns no passage above the relevance floor. The case is outside the corpus, which is a statement about the documentation rather than about the customer.
Something retrieved that does not answer. Passages come back on the topic but none of them addresses the question asked. This is the most common null and the most dangerous, because a fluent model given an adjacent passage will write an answer that appears to rest on it.
Two passages that conflict. The corpus supports both readings — a policy revised in one place and not another, a product page and a price list disagreeing. The conflict is a fact about the corpus, and a reply cannot resolve it.
In all three the correct output is an escalation record rather than a hedge, and the record has its own fields: the question as the customer put it; the queries run and where they were run; the documents searched with their versions; what was found and rejected, with the reason for rejection; the gap stated in one sentence; and the owner with a date. Six fields, of which the fourth and the fifth are the ones that make the record useful. A note that says “could not answer, needs a human” costs the person who picks it up the same twenty minutes the agent spent. A note that says which two documents were searched, what was found, why it did not answer, and what is missing turns the case into a five-minute decision about the document rather than about the customer.
The null result is also the only channel through which the corpus improves. A gap that surfaces as an escalation becomes a ticket against the documentation; a gap that surfaces as an invented answer becomes a complaint three weeks later with no trace back to the missing page. The queue of escalations is therefore a maintenance list, and a support operation that treats it as a failure of the agent has misread what it is looking at.
A week, worked
The example is a constructed support desk: one mailbox, a corpus of twelve documents, a bank of sixty-one rows, and one hundred questions in a week. No client, customer or product is described, and the counts are arranged so the arithmetic can be followed rather than so any desk can be described.
Table 2. One constructed week of one hundred questions: where the answer came from, what it cost, and what it left behind. Source: constructed for this piece.
| Disposition | Questions | Paid calls | Cost |
|---|---|---|---|
| Answered from a bank row standing on the current document version | 58 | 58 | $0.40600 |
| Answered from a corpus passage with no bank row | 23 | 23 | $0.16100 |
| Escalated — the corpus holds no answer | 11 | 11 | $0.07700 |
| Routed without a draft — sales, legal, personal data, live complaint | 8 | 0 | $0.00000 |
| Totals | 100 | 92 | $0.64400 |
The dispositions reconcile: 58 and 23 are the eighty-one replies that rest on a document, 81 per cent of the week; 11 escalated and 8 were routed, nineteen cases, 19 per cent, that no document answered. The 92 paid calls are the eighty-one phrasing calls plus eleven escalation summaries at $0.00700, which is $0.64400. Retrieval ran one query per question and two per escalation — one hundred and twenty-two queries — and carries no published per-call price, so it sits outside the total.
Two repair findings belong to the same week and belong in the record. Five of the fifty-eight bank-answered replies cited a document section that had been superseded — 8.6 per cent of the replies that used a row, against 8.2 per cent of the sixty-one rows in the bank — and each was repaired by rewriting the row against the current version, at one call apiece. That is five calls and $0.03500, which takes the week to ninety-seven calls and $0.67900. At the metered ceiling of $0.00955 the same ninety-seven calls are $0.92635.
Set against the pillar’s measured month — 498 paid calls, $4.51575 — the ninety-seven calls of this week are 19.5 per cent of the month’s calls and 15.0 per cent of its cost, and the ninety-two that precede the repairs are 14.3 per cent of the cost. Fifty calls at the measured rate is $0.35000; the same fifty at $0.00875, which is Qwen’s measured rate and not DeepSeek’s, would be $0.43750.
The number that matters in the table is not the cost. It is the twenty-three: questions the corpus answers and the bank does not. Each is a correct reply that took a retrieval and a read to produce, and each is a candidate row. A desk that writes those twenty-three rows spends one afternoon and converts a repeated cost into a fixed asset. A desk that does not writes the same twenty-three replies again next week, at the same price, with the same variation in wording.
What the bank cannot tell you
A citation is provenance, not truth. The passage is the organisation’s own statement on the date its version was cut. If the help page is wrong, the reply is wrong and now carries a footnote that makes it harder to argue with. The bank moves the question from “did the model invent this?” to “is this document right?”, which is a better question and still a question.
Coverage is not outcome. The counts above measure what rests on a document; they do not measure whether the customer accepted the answer, whether the case closed, or whether the reply was any good in the sense a customer means. This publication has no measurement of resolution rates, satisfaction scores, deflection or response times for this family, and the pillar says so. The only quantity measured here and in the pillar is cost, and the pieces in this family should be read with that limit held firmly in mind.
The bank decays faster than the corpus. A document changes once and invalidates several rows; the rows are not edited by the person who edits the document, because the two are different jobs in different systems. The 8.6 per cent above is one week’s finding in a constructed example, and it is offered as a demonstration of the check rather than as a figure anybody should expect.
The null depends on an honest query log. An agent that reformulates a question until something comes back will always find something, and what it finds will be adjacent rather than answering. The log is what makes the null visible; a log that records only the successful query records nothing.
The decision
Three gates, applied to the bank rather than to the replies, and a bank that passes them is auditable in an afternoon.
Does every reply either cite a passage that resolves in the corpus or carry an escalation record with its six fields? If a reply does neither, it is the failure the design exists to prevent, whatever it says.
Does every row carry an owner, a version and a verified date, and does the owner know that they own it? A row with no owner is a row that will be wrong by the next product change, and the date is what makes that predictable rather than surprising.
Does the week’s null result exist as a list that somebody reads — questions the corpus could not answer, with the document that should have answered them? If the escalations go to a person as cases and never aggregate into a list, the corpus cannot improve by design; it improves by accident, when somebody senior notices.
Where this goes
- The pillar: x402.review/2026/10/31/front-office/ — the family, the six artefacts and the boundary this piece works inside
- The role: studio.x402.press/roles/customer_service — retrieval over the client’s corpus, the escalation summary and the mailbox
- The service it buys: x402.press/services/deepseek/ — the canonical price, seven chains
- The service it buys: x402.press/services/qwen/ — the second model, priced per call
- The reviews: x402.review/2026/10/31/deepseek-review/ and x402.review/2026/10/31/qwen-review/ — the measured handshakes and what they do not promise
The cost of a bank
A week of one hundred questions is ninety-seven calls once the five row repairs are counted, and $0.67900 at the measured DeepSeek rate on Base — $0.00700 a short-prompt call under exact, read from the live handshake on 31 October 2026. At the metered ceiling of $0.00955 the same week is $0.92635.
The pillar’s month is 498 paid calls and $4.51575, which is $4.52; the support side of it is 450 calls — three hundred drafts at $2.10000, a hundred and twenty verification passes at $0.84000 and thirty tone variants at Qwen’s $0.00875, $0.26250 — for $3.20250. One complete deal’s document set is fifty calls at $0.35000. The figure that circulates as DeepSeek’s price, $0.00875, is Qwen’s measured rate and not DeepSeek’s: at it the same fifty calls read $0.43750, and the 498-call month would read $5.28925, or $5.29, instead of $4.51575.
Method note
Prices were read from the live 402 handshake on Base on 31 October 2026: $0.00700 for a short prompt under exact, $0.00955 for a long prompt under exact and as the upto ceiling at max_tokens: 1000, and $0.00875 for qwen3.8-flash under both schemes. The catalogue figure of $0.00875 for DeepSeek was checked against the handshake, found to be Qwen’s rate and stale for DeepSeek, and is not used here except where it is named as stale beside the correction. Table 2 sums: 58 plus 23 plus 11 plus 8 is 100, 58 plus 23 plus 11 is the 92 paid calls, and 92 at $0.00700 is $0.64400; with five repairs at one call each the week is 97 calls and $0.67900. The desk, the corpus, the bank and the counts are constructed so the arithmetic can be followed; no client, customer, product or document is described, and none is real. Retrieval, the reading floor, the report renderer and the mailbox carry no published per-call price and are outside every total. No outcome is measured anywhere in this piece: the counts are of replies that rest on a document, not of replies that resolved anything.
x402.review is written and edited by N.P. Vincent. The publication is connected to x402.press, which sells the services reviewed here; that interest is disclosed on every piece that touches a service, and it is stated here rather than left to inference.
Filed under Jobs · Tags: support, documentation, answers, retrieval
