The front office: the reply that cites its source, and the send it cannot make

Front office work is the work an organisation does while it is being watched. A customer writes in and expects an answer that is true, in the company’s voice, by the end of the day. A principal opens a diary and needs to know what the day holds before it starts. Both jobs produce a document that leaves the building in someone else’s name, and both are decided not by how well they are written but by whether the sentence can be traced to something the organisation actually stands behind. This piece sets out the job, the two roles that carry it, the six ways it fails, the measured cost of a month — $4.52 at the rate the endpoint asks for, against a figure that has been circulating at a quarter more — and the boundary that matters most here: an agent may draft the reply and the brief, and may not send the one or promise anything in the other.

The job

Every organisation has an edge, and at that edge two kinds of document are produced continuously. One goes outwards: the answer to a customer, a prospect, a tenant, a supplier, a regulator’s routine query. The other goes inwards: the brief that tells one person what happened overnight, what the day requires and what is waiting on somebody else. They look like different jobs, they sit in different inboxes, and they fail for the same reason. Each is a claim made under the organisation’s name about something the organisation possesses — a policy, a delivery date, an appointment, a price — and each becomes expensive at the exact moment the claim and the record part company.

That is why the two belong in one family. A customer service reply and a principal’s brief are the same operation on different material: retrieve what the organisation already knows, keep the wording faithful to what was retrieved, and keep every act that would bind the organisation out of the agent’s reach. The first is retrieval with a stranger at the other end. The second is retrieval with the principal at the near end and a working day as the deadline.

The work has a shape that predates any software in the room. Triage: decide what matters and what can wait. Retrieval: find the passage, the booking, the last exchange, the policy that governs the answer. Composition: put it in a voice the recipient recognises. Escalation: hand over what cannot be answered, with enough context that the person picking it up does not start from nothing. Approval: let a human release anything that leaves, commits money or changes a record. Any competent front office does all five by hand, badly under pressure, and well when it is properly staffed and quiet.

The distinction that makes this family worth treating separately from the back-office pillars is that its output is seen. An error in a contract abstract is discovered by a lawyer before signature. An error in a customer reply is discovered by the customer, acted on, and then quoted back. The cost of a wrong sentence is asymmetric: the reply that invents a returns window creates an obligation somebody has to honour or refuse in public.

The front office also has a property that makes it unusually suited to this treatment and unusually easy to get wrong. Almost all of its material is already text, and almost all of it is already the organisation’s own: help pages, policy documents, product guides, order confirmations, calendar entries, the last three exchanges with this customer. There is nothing to invent, and the temptation to invent is therefore entirely internal to the instrument doing the work.

Two roles, one discipline

Two roles carry the family in the studio. customer_service answers a support mailbox the client owns. personal_assistant keeps one person’s day straight. They share a method and differ in who reads the output, which is exactly the distinction that decides what each is allowed to do.

The customer service role works a mailbox. Each message is answered from the client’s own corpus, retrieved rather than remembered: the agent searches the documentation, quotes the passage and phrases the reply in the client’s voice. Attachments the customer sends are read before they are answered. When retrieval returns nothing usable, the case is escalated to a person with a written summary — what was asked, what was checked, what is missing — instead of a guess dressed as help. Its parts are a knowledge-graph retrieval layer over the corpus, a reading floor for attachments, a report renderer for the escalation summary and the client’s own mailbox under an allow-list. A second model can be attached for a second opinion before a refusal.

The personal assistant role works a day. It reads the morning’s material — the mailbox, the calendar, documents sent for review, lists and itineraries typed into a message — and produces the brief before the day starts: what arrived that matters, what the day holds, what has to be decided and what is waiting on somebody else. Then it drafts: triage with a suggested reply on each message, itineraries and list tables, and preparation for each meeting with the last exchange, the open questions and the outcome being sought. Its parts are a document engine, a reading floor, a data table, the person’s own mailbox and an optional Telegram channel for pushing the brief. Sending is a separate permission from drafting, and nothing leaves without an approval.

Table 1. The two roles and the artefacts that carry the work. Source: studio role pages for customer_service and personal_assistant, read 31 October 2026.

The asymmetry in the last column is the point of the family. The customer service reply is built to leave, and a person releases it; the assistant’s draft is built to wait in an outbox, and the gate is the product. Neither role sends anything unattended, and only one of them is designed around the fact that its output will be read by a stranger.

The reply that cites its source

A support reply has two components that are easy to confuse and must be kept apart: the fact and the sentence. The fact — what the returns window is, whether this model is in stock, which address the order went to — is the organisation’s and exists in its documents. The sentence is a matter of tone, length and the house register. A model is very good at the second and unreliable at the first, and most of the harm in automated support comes from letting it do both.

The retrieval-first design separates them. The corpus is indexed, the message is used as a query, and the passages that come back are the answer’s raw material. The model’s job is the sentence, not the fact. A reviewer can open the passage a reply cites and check it in a minute, which changes the economics of supervision: reviewing a cited reply is verification, and reviewing an uncited one is re-doing the work.

The passage is also the thing that makes a refusal coherent. When retrieval returns nothing usable, the correct output is not a smoother answer; it is an escalation that names what was asked, what was checked and what is missing. That is what makes an escalation worth reading, because the person who picks it up starts from a written record rather than from the customer’s original message and twenty minutes of searching. It also makes the corpus’s gaps visible as a queue rather than as a slow drip of wrong answers, which is the only form in which a documentation gap gets fixed.

What the reply must not do is the other half of the design. It must not answer beyond the corpus: if retrieval finds nothing that addresses the question, the reply is not improvised from general knowledge, and the customer is told the case has gone to a person. A plausible answer the documentation does not support is the most expensive kind of help a support desk can give, because it is fluent, it is specific, and it is unowned.

The reading floor matters more than it sounds. The role’s own volume model assumes a fifth of messages arrive with something attached — a photograph of a damaged item, a scanned invoice, a screenshot of an error, a filled-in form. An agent that answers the covering message and ignores the attachment is answering half the question, and confidently. Reading what the customer sent, before the reply is written, is not a refinement; it is the difference between a reply and a guess.

The mailbox itself is a permission. It is the client’s, under an allow-list, with credentials inside the container, and each client’s corpus is its own. That arrangement is what allows the role to be described as answering from the organisation rather than about it.

The day that gets briefed

The assistant’s work has a different shape because its reader is one person who already knows most of the context. What the principal lacks is not judgement but a current picture: what came in overnight, what is fixed in the diary, what is still open, what is overdue, and what is waiting on somebody else. That picture is the brief, and the brief is the anchor artefact of the role. The tables and the preparation notes are read against it, and a day with no brief is a day the assistant has nothing to measure the rest of the work against.

The order follows the reader’s morning. Fixed points first, because a day is structured by what cannot move. Open decisions second, because they are the only part of the day that will not resolve itself. What arrived overnight third, ordered by whether it needs a person this morning. What is overdue last, because it is the part the principal is most likely to postpone again and the part a brief is most useful for surfacing.

The triage table is the working part. One row a message: sender, subject, what it asks, urgency and a suggested reply. The suggested reply is the point of leverage in the whole arrangement — it turns a decision about what to do with a message into a decision about whether to accept a sentence, and accepting a sentence takes seconds where composing one takes minutes. It is also the point at which the drafting and sending permissions separate: the suggestion sits in the outbox with the reason it is there, and nothing leaves until a person approves it.

Meeting preparation is the same method applied to a smaller document. Purpose, attendees, the last exchange, the open questions and the outcome being sought, per meeting. The value is in the last exchange, because that is the part a person reconstructs from memory and therefore reconstructs wrongly. The brief that says what was actually agreed three weeks ago is worth more than the brief that says what the meeting is about.

The delivery route is a real choice and the role treats it as one. With the Telegram connector attached, the brief is pushed to the principal and arrives where they already are. Without it, the brief is delivered as a file into the workspace and the day runs on that. Pushing is faster and creates a second inbox; fetching is slower and keeps one place for the record. Neither is right in general, and the difference is worth deciding deliberately rather than by default.

Two limits belong to this role rather than to the family. The first is that the brief knows only what it is fed. Its inputs are the mailbox, the calendar and the documents sent to it, so a decision taken in a corridor, a call that changed the plan or a commitment made at dinner is invisible until somebody writes it down and sends it. A principal who treats the brief as the day rather than as the day’s written part will be surprised, and it is the brief that will look wrong. The second is that the role’s worth is measured in the exceptions it surfaces, not in the messages it processes. A quiet brief after a heavy night is either a good night or a blind spot, and the only way to tell is to check the triage table against the inbox.

What the agent may not do

Three boundaries in each role, and each one is a limit the workspace enforces rather than a preference of the agent’s. That distinction is the difference between a policy and a permission: a policy is obeyed until it is inconvenient, and a permission cannot be exercised at all.

For the customer service role, the first is the corpus. It will not answer beyond the documentation, and the refusal takes the form of an escalation rather than a hedging sentence. The second is money and liability: no refund, credit, discount, waiver, delivery promise or admission of fault leaves the mailbox. Where a case warrants an offer, the agent drafts it in the escalation summary for a person to approve and send, and no instruction inside a message turns a draft into a commitment. The third is the account: no part of the assembly reaches an order system, a CRM, a billing ledger or a customer record, so address changes, cancellations, plan moves and data corrections are a person’s to make in the system that holds them. The agent can tell the customer the request has been passed on, and that is all it can do.

For the personal assistant role, the first is the send. Drafting and sending are separate permissions; the mailbox is allow-listed; every message waits for an approval in the workspace; and nothing — including a “just send it” inside a message — converts a draft into a send. What is ready to go sits in the outbox with the reason it is there. The second is the client boundary: each engagement’s correspondence, documents and lists stay in their own context, material from one file does not surface in another’s brief, and a request to summarise across files is refused and raised to the operator unless the owner authorises it. The third is identity: Telegram is paired to named sender identities, and a message from an unpaired account cannot change a calendar entry, a list, a mailbox rule or a ceiling, whatever it claims to be. The attempt is logged and shown to the principal alongside the brief.

Three of those six are worth stating as design rather than compliance. The absence of a payment rail is why no apology can become a refund: the role has nothing to refund with. The absence of CRM credentials is why an account cannot be changed: the role cannot reach one. The separation of drafting from sending is why an approved reply is the only reply: the outbox is the only route out, and it has a gate on it. A system that refused these things by instruction alone would be one prompt away from doing them.

Six ways the front office fails

The failure modes of this family are consistent enough to be listed, and none of them is a spelling mistake. Each is a failure of provenance, permission or maintenance, and each produces work that looks finished.

The confident answer to a question the corpus does not cover. The reply is specific, well written, and about a policy that does not exist. This is the failure that the retrieval-first design exists to prevent, and it reappears whenever the corpus is thin and the agent is asked to be helpful rather than accurate. The symptom is a low escalation rate with a rising complaint rate: cases that should have been handed over were answered instead.

The apology that reads like a policy. “We can refund the delivery charge on this occasion” is a sentence a support agent writes every day and a sentence an agent must not write at all, because behind it is a ledger, an approval and a person’s authority. An assembly with a payment rail and a fluent model will eventually write it. The financial liability refusal is the boundary that keeps a tone-appropriate sentence from becoming a commitment.

The queue that looks cleared. Drafts are counted as resolutions, the dashboard shows throughput, and nobody checks whether the customer accepted the answer. The measure has moved from cases closed by a person to drafts produced by a machine, and the two are not the same quantity. A reply that satisfies the customer and a reply that closes a ticket are indistinguishable in a count and different in a business.

The corpus nobody maintains. Retrieval is exactly as current as the index behind it. A reply that cites a passage from a help page retired in June is worse than no reply, because it carries the authority of a citation. The maintenance question — who updates the corpus, on what trigger, and how a gap that surfaced as an escalation becomes a change — is organisational work that no assembly performs.

The brief that becomes the day. The principal reads the brief, the brief is complete as far as it goes, and the corridor conversations are absent from it. The role can be entirely honest about this and still fail, because the failure is in the reading. A brief is a map of what was written down; it is not a map of the day.

The tone that is right and the content that is wrong. Fluency is a poor signal of accuracy, and a model asked to make a reply warmer will make it warmer, including the false parts. This is the failure that a second model can catch and only if it is asked to check the claim rather than the register.

The second model, and why it is not a price argument

The customer service role can be assembled with a second model attached, taking a second opinion before the reply becomes a refusal. The reason is narrow and worth stating precisely: a refusal is the one output where being wrong is both expensive and invisible. A wrong answer produces a complaint; a wrong refusal produces a customer who goes elsewhere and says nothing.

Independence is the whole value. Two passes through the same weights produce agreement, not verification. A second model from a different vendor, with different training and different failure modes, is a genuine second test of the same claim, and the ask is specific: does the passage support the refusal, or is there something in the corpus that answers this? Asked that way, the second model is cheap insurance on a small number of calls.

It is not a price argument, and this is where the figure that has been circulating needs correcting. The two models do not cost the same. DeepSeek’s measured price on Base is $0.00700 a call under exact. Qwen’s is $0.00875 — a quarter more per call, $0.00175. The 120 verification passes in Table 2 are priced at the first model’s rate, $0.84; run the same 120 checks on the second model instead and that line is $1.05, twenty-one cents more. The parity claim is false, and it does not matter: what the extra $0.21 a month buys is an independent opinion on the refusal, and no argument for a second model has ever depended on it being the same price as the first.

There is one piece of genuine price symmetry worth knowing. Qwen is priced at $0.00875 under exact and at $0.00875 under upto: its ceiling is its floor, measured on the same handshake, so a second-model check has no metered band to be surprised by. That is a property of one service’s pricing rather than a general statement about second models, and it is the only sense in which the two can be said to cost the same.

The cost of a month, measured

This publication takes the amount from the 402 challenge in front of the call and treats a catalogue or a price list as a way to check a figure, never as a way to supply one. The figure that has circulated for a DeepSeek drafting call is $0.00875. On Base that is Qwen’s rate, not DeepSeek’s: the measured price of deepseek-v4-flash is $0.00700 under exact, read from the live handshake on 30 October 2026 outside the service’s peak windows, which are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. The same call under a long prompt, and the upto ceiling at max_tokens: 1000, is $0.00955. Qwen’s qwen3.8-flash is $0.00875 under both schemes.

The month below is the composite a small office actually runs, with both roles in it: a support queue of three hundred replies, and one principal’s brief every working day. The call counts are planning estimates; the prices are not.

Table 2. One month of front office work, both roles, at the measured rates. Source: live 402 handshake on Base, 30 October 2026; weather and news are catalogue list prices for services pending review; call counts are planning assumptions stated below.

The arithmetic is short. Four hundred and ninety-eight paid calls a month is $4.51575, which is $4.52: about a cent and a half a reply at the volume above, or $0.205 of model calls a working day. Over twelve months at the same pace it is $54.19.

At the rate the catalogue carries for DeepSeek, $0.00875, every DeepSeek line in that table would be a quarter higher and the month would read $5.29, the year $63.47. Those are the figures this piece replaces, and it replaces them line by line: the three DeepSeek lines in the table are $3.094 at the measured rate and $3.86750 at the stale one, and the difference between the two months, $0.7735, is the whole of the correction.

Two of those lines cannot be bought today, and the table says so rather than absorbing them. Weather and news are both pending review in the catalogue: named, priced in the list, and not callable. Their figures are list prices, and they are the only figures in this piece that come from a catalogue rather than a handshake. Take them out and the month falls to $3.35650 — $3.36, across 472 paid calls — which is what the front office costs in services that sell today. The year on that basis is $40.28. The difference between the two totals is not small: the four news asks alone are $1.02450, almost a quarter of the composite month, because one news ask costs more than thirty-six model calls.

Table 3 puts the month, the year, the opening balance and the metered ceiling beside each other.

Table 3. The month, the year, the opening balance and the metered ceiling. Source: live 402 handshake on Base, 30 October 2026; twelve months is an assumption.

Two readings of the same figures are worth keeping. The ceiling is where a long prompt and a metered call with a thousand authorised output tokens sit, and on the DeepSeek stream alone it is 36 per cent above the measured rate — the whole month’s DeepSeek calls would cost $4.22110 instead of $3.09400. Blended with the Qwen line, which has no band, the month of services selling today at the ceiling is $4.48360, 34 per cent above $3.35650. The ceiling is not a price rise; it is the difference between asking for a short answer and authorising a long one.

The second reading is the opening balance. A workspace opens with 0.50 USDC in the client’s own wallet, credited to the workspace as working capital. At $0.00700 that is seventy-one DeepSeek calls — not the fifty-seven the same balance reads at the stale rate of $0.00875. It is an opening balance, not a fee: nothing is spent until a call is made, and the balance is the client’s, held and spent from their own wallet. At this month’s pace, seventy-one calls is three and a half working days of DeepSeek work; across the whole composite month, including the news asks, half a dollar covers about two and a half days. The default balance is not a launch budget for a support desk, and should not be treated as one.

The hours, not the cents

The number to argue about is not $4.52. It is twenty hours.

Three hundred replies a month is fourteen a working day, and at four minutes each — read, find the passage, phrase, check — that is twenty hours of a person’s time. Twenty hours is half a working week, or a fortnight of mornings, or the difference between a support desk that answers within the day and one that answers within the week. The model spend against those hours is $4.52, which is $0.23 an hour, or $0.015 a reply. Unless an hour of that person’s time costs less than $4.52 — and in any market that employs support staff it does not — the hour is the expensive column, by an order of magnitude.

That ratio changes what the decision is. Nobody is being asked whether $4.52 is affordable; the answer to that is yes in every organisation that has a support queue at all. The question is what the twenty hours are spent on once the drafting is not in them. Two answers are available and only one of them is worth the money. The twenty hours can go into the exceptions — the escalations, the awkward customer, the policy that keeps generating the same question — and the queue then improves. Or the twenty hours can be spent reviewing drafts that would have been written anyway, in which case the arrangement has bought a second reading of the same work and the saving is imaginary.

The same argument settles the ceiling, and this is where a low figure usually misleads a buyer. Moving every DeepSeek call from the measured rate to the metered ceiling adds $1.13 a month. Setting the daily ceiling high enough that a busy Monday never stalls the queue costs nothing at all until the calls happen. Both are trivial against an hour of a person’s time, which is why the sensible settings on this role are generous ones: the failure that costs money here is a queue that stops because a spend ceiling was set from a quiet week’s average.

It also settles the second model. The 120 verification passes cost $0.84 on the first model and $1.05 on the second; the Qwen tone line is $0.26. Any argument that a second opinion is not worth about a dollar a month has to be an argument that refusals are not worth checking, because at any wage a person is paid it cannot be an argument about the money.

The spend that is genuinely worth watching is not in the table. Preparing the corpus — help pages, policies, product guides, indexed as text rather than as photographed PDFs — is a day of somebody’s time and a recurring half-day every quarter to keep it current. That is many times the model spend of the month, and it is the line that decides whether the replies are worth reading. The same is true of the assistant’s inputs: a mailbox and a calendar exported as text, and a principal who writes down what happened in the corridor.

What the figure leaves out

Retrieval over the corpus carries no published per-call price, and retrieval is most of the customer service role’s work: one query a message and two an escalation, three hundred and sixty queries in this month at the volume above. The document parts carry no published per-call price either: the reading floor that opens attachments, the report renderer that writes an escalation summary, and on the assistant’s side the engine, the data table and the extractor. The mailbox and the Telegram connector carry no published per-call price. None of those is metered per call in the catalogue, they run inside the workspace rather than over the wire, and there is no 402 challenge to read, no amount to measure and no price to report.

That is an omission, not a saving. The totals above understate the work by everything the workspace does besides calling a model, and the honest reading of $4.52 is that it is a floor for the model calls of one month, not a price for the front office. Where a supplier publishes no per-call price, the correct entry is not an estimate; it is the statement that no price exists.

The largest cost is absent from the tables altogether. Writing and maintaining the corpus is a person’s hours. Reading escalations and answering them is a person’s hours. Deciding what the reply is allowed to promise — the refund policy, the delivery promise, the retention offer — is a decision that belongs to somebody with the authority to make it, and no assembly makes it. The assistant’s hour is the principal’s own: approving drafts, correcting the brief, telling it what happened in the corridor.

There is also a category the tables cannot hold at all, and it is the reason the two roles are priced with a caveat. Neither role prices the failure it prevents. A support desk’s cost is not only the drafting; it is the complaint that follows a wrong answer, the customer who leaves after a refusal that should have been an escalation, the meeting that went badly because nobody had read the last exchange. Those are real costs and this publication holds no measurement of them, so it makes no claim about what the arrangement saves beyond the hours in the paragraph above.

When this is the wrong instrument

Six cases where the work is real and an assembly is not the way to do it, and one where the arithmetic settles it rather than a principle.

When there is no corpus. The customer service role answers from documentation, and an organisation without written help pages, policies and product guides has nothing to answer from. The first job in that case is to write them, with the people who answer the questions today, and an assembly given an empty corpus produces either an empty queue of escalations or a fluent set of inventions.

When the answer requires authority. A complaint that threatens legal action, a regulator’s enquiry, a data protection request, a bereavement, a safeguarding concern, a dispute that has already escalated: each of these needs a person with standing, and the right behaviour is to route them before a draft exists. The role can escalate, and escalation is the correct output, but the case is a person’s from the first message.

When the volume is tiny. Twenty messages a week is an hour and a half of work at the pace assumed above. The corpus preparation is a day, the configuration is an afternoon, and the arrangement costs more attention than the queue it removes. Small queues should be answered by hand until they are not small.

When the correspondence is a relationship. Key accounts, negotiations, anything where the words are the work: a reply that quotes the policy is the wrong reply to a customer who is deciding whether to stay. The role’s value is in the routine majority of a queue, not in the messages that decide whether there is a queue at all.

When the principal’s day is not in the systems. The brief can only read a mailbox and a calendar. A principal whose week lives in their head, in voice notes and in other people’s diaries gets a brief that is accurate and irrelevant, and will stop reading it within a fortnight.

When what is wanted is a channel rather than a document. A telephone line, a live chat with a human on it, a counter: these are service, and a document-producing assembly does not staff them. It can draft what a human sends, which is a different and smaller claim.

The seventh case is arithmetic, and it is about the corpus rather than the model. A day of corpus preparation buys a month of three hundred replies at four minutes each — twenty hours of a queue, repaid in the first month at any wage the person answering it is paid. Against sixty messages a month, the same day buys four hours, and the maintenance that follows a corpus is larger than the queue it serves. The arrangement earns its place where the queue is relentless and the exceptions are the interesting part.

What we measured and what we did not

The prices in this piece were measured, not listed: the amount in the 402 challenge, read on Base on 30 October 2026 outside DeepSeek’s peak windows, which are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. The model identifiers were read from GET /v1/models, which lists deepseek-v4-flash and deepseek-v4-pro on the first service and qwen3.8-flash and qwen3.8-max on the second. Catalogue values were used to check a figure and never to supply one, with two stated exceptions: the weather and news lines in Table 2 are catalogue list prices, because neither service is callable and there is no handshake to read.

The correction is worth stating plainly, because it is the substance of this section. DeepSeek on Base is $0.00700 a call, not $0.00875. The higher figure is Qwen’s measured rate, and it is also the value the catalogue still carries for DeepSeek, which is why it has circulated. Every total in this piece is at the measured rate; at the stale one the DeepSeek lines alone would read $3.86750 for the month instead of $3.09400, and the composite month $5.29 and the year $63.47 instead of $4.52 and $54.19.

The call counts are planning estimates, and they are labelled as such in Table 2: three hundred replies, a hundred and twenty verification passes, thirty tone variants, twenty-two briefs, twenty-two weather lookups and four news asks. They are not observations of a real desk, and they will be wrong for any particular one. The twenty hours is an assumption of four minutes a reply at the volume stated, and it is the one figure in this piece that a reader is invited to attack with their own numbers.

What this piece did not do matters as much as what it did. It did not run a mailbox through the assembly, and it describes no client, no customer and no completed engagement, because it has none. It holds no measurement of reply accuracy, no escalation rate, no time-to-reply figure and no customer satisfaction figure, because none has been measured. The refusals described above — the corpus boundary, the money, the account, the send, the client separation, the unpaired sender — are stated in the roles’ specifications and enforced by the parts they have and the permissions they lack. They are reported here as specified rather than tested against a hostile customer, and a reader who needs to rely on one should test it.

Amounts move, and a table is a snapshot. The same call costs different amounts on different chains, under exact and upto, across DeepSeek’s prompt bands and inside its peak windows, and the ceiling moves with the max_tokens an agent authorises. The challenge in front of the call is the price. Where a figure here is wrong, it will be corrected and the correction stated, in the manner of every piece on this site.

Where this goes

• The role: studio.x402.press/roles/customer_service — the support mailbox, the corpus and its running cost
• The role: studio.x402.press/roles/personal_assistant — the brief, the triage table and the approval gate
• 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 service it names and cannot buy: x402.press/services/news/ — pending review, not callable
• 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 catalogue: studio.x402.press — fifteen entries, thirteen roles and two desks

The clusters in this family

• Answering from your own documentation
• When to escalate a support case
• Support macro against a model draft
• Measuring first-response quality
• The daily brief
• Inbox triage

Method note

Prices read from the live 402 handshake on Base on 30 October 2026, outside DeepSeek’s peak windows; 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 anywhere in this piece except where it is named as stale beside the correction. The weather and news figures are catalogue list prices for services pending review and not callable, and are labelled as such wherever they appear. Call counts are planning assumptions, stated in Table 2 and repeated in the arithmetic; the twenty hours is an assumption of four minutes a reply. Nothing in this piece is a promise made on an organisation’s behalf: the roles draft and escalate, and a person releases, promises and changes records. No client, customer or engagement is described, because none was observed.

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: front-office, support, escalation, daily-brief, x402