Price is the smallest lever on your desk and the only one with a number on it. A submission lands in the last week of the renewal season. Three years of losses, a schedule of values with the same address typed three different ways, and a broker’s note saying it’s much like the account you wrote in the spring. You have an hour for it, and the pile behind it hasn’t stopped growing since the weekend.

If your book is motor or household, none of this is about you. That was settled by tables and screens long before any of the current noise. For everyone else, that hour goes on two different kinds of work and only one of them has a name. Pricing what has already happened is the named one. Deciding about what hasn’t is the other, and on most days it never announces itself, because the difference only shows up on the day it matters.

The first one is going, and I’m not going to be coy about it. Loss history is data, and a model is a machine for eating loss history. It will triage the submissions, price the ordinary risk better than you do, benchmark the terms against everything else in the market, and handle a renewal where nothing has changed. It will do it at the end of the renewal season without getting tired, which is when your book gets its worst decisions made. Anybody in your trade who insists a person prices small commercial better is going to lose that argument in front of the people who hold the pen on it.

The mistake about your job starts in the part of the lever that has no number on it. Where there is a technical rate it comes from the actuaries, and in the lines this post is about there often isn’t one, so you build the number yourself out of exposure and whatever the market is doing that month. Either way the number is the smaller part. Selection, structure and refusal are the rest of it. The deductible, the limit, the sub-limit, what sits inside the aggregate, the exclusion you insisted on, the wording you would not accept. Premium is one number among all of that. The trouble lives in a sentence somebody drafted years ago that a court will read in circumstances nobody had thought of.

The thing that actually kills insurers isn’t the enormous single loss. Everybody plans for the enormous single loss. That is what the reinsurance is for, and the whole market has rehearsed it. What ends companies is a book full of separate risks that turn out to be one risk. Every account sound on its own. Every one of them leaning, without anybody writing it down, on the same supplier, the same assumption inside the same standard clause, the same piece of infrastructure nobody thought to ask about. You wrote a portfolio. You owned a single position.

The machine’s confidence will be highest exactly there, because every one of those accounts resembles the history closely, and resemblance is what confidence is made of. Each price is defensible. The portfolio is a bomb. And this is the part I would put in front of any underwriter and ask them to shoot down: a loss with no precedent does not arrive looking strange. It arrives looking ordinary, many times over, one broker after another.

Somebody has to see it, and the only people who ever have are the ones who watched a book turn. You bind something today and find out later, sometimes at the end of a career, and it comes back as a reserve moving on a decision you can barely remember making. Not the big losses. The small steady wrongness, arriving late, about business you thought was fine. That is the slowest lesson in finance, nobody has found another way to teach it, and it is why the dull business matters. Ordinary risks are the only thing you get enough of to have a book of your own. If the model writes all of them, the person behind you sees referrals and nothing else, which sounds like a promotion and is nothing of the kind. An exception means nothing without a mass of ordinary underneath it to be exceptional against.

Some of this has nothing to do with data at all. You learn to decline by living through a soft market with your own name on the declinature. Everybody around you is writing the business, the broker is disappointed in you, your line is being filled by somebody down the street, and the numbers say you are wrong for a long stretch. Then it turns, and you find out. Reading about a cycle is not the same as having sat in one and held. If the ordinary book is machine written, that pressure never lands on a person early enough in their career to shape them, and the willingness to say no stops being a memory and becomes an opinion.

There is a serious objection to all of this, and your exposure management people will make it before you do. Aggregation is not unmanaged. There are catastrophe models, stress tests, disaster scenarios, accumulation limits by zone and class, and a whole function that exists to ask exactly the question I have just described. That is all true, and any post pretending otherwise has earned whatever it gets.

So let me keep my claim small. Everything in those models is a peril somebody already named. The scenario exists because a person once looked at a pile of paper and got an itch, and said out loud that these accounts might be the same account. That naming is the act. It has never come out of the data, because before the first loss there is no data, and after it there is no shortage of people willing to model it. What worries me is that the naming gets done by people who spent years being wrong slowly, and we are quietly closing the only place that ever happened.

And your trade is built so that none of this has to be noticed for years. Insurance is the one business that takes the money before it knows what the goods cost. A book written on borrowed confidence looks like a good year, then another good year, and the accounts agree with you the whole way through. By the time the reserves move, the way of pricing that did it has been the house standard for years. It is in the model everybody inherited, and the underwriters who might have argued with it are running somebody else’s division.

Take two accounts you have written this year, different brokers, different classes, and read the two wordings side by side until you find the assumption they are both resting on. If nothing turns up, read them again, because a real one is never in the same paragraph twice. Then work out whether anybody starting this year will ever hold enough business to run that test at all. I know what I would bet on that, and this isn’t the post where I write it down.

The work you did in your first years wasn’t design and everybody in the studio knew it. You resized other people’s layouts, you set price lists, you built the deck for somebody else’s meeting, and you redrew a logo from a photograph of a logo because a photograph of a logo was all the client had. Then you went home and did your own work, which was better, and which nobody paid for.

That production tier is going first and going fastest, and I doubt anybody in your trade will argue with me about it. My claim is narrower and less comfortable. Those jobs were the whole training scheme. There has never been another one, and nothing that is arriving now does what they did.

Start with type, because it is the clearest case. You set a headline, somebody senior looked at it for about a second, said it was loose, and walked off. You fixed it, or you thought you fixed it, and they said it again. This happened on jobs where the headline didn’t matter, week after week, until the day you saw it yourself before anybody told you. None of that arrived as a rule, because there isn’t one to arrive. What you were building was a threshold for wrongness, and a threshold only calibrates against attempts you made yourself and cared about.

Then production, which designers outside the trade forget and designers inside it never do. A file comes back from a printer with the type filled in, or a color that was gorgeous on screen turns to mud on uncoated stock, or a laminate eats the finish somebody sold to the client. You find out about bleed by not leaving any. You learn what a printer will phone you about by being phoned. All of that knowledge came from small jobs, because a small job is where you are allowed to be the reason something went wrong.

And then the clients. Somebody says make it pop, and every designer alive has heard that sentence and winced. What it usually means is that they can’t tell where to look first, which is a real problem and a fair complaint, and their proposed fix is wrong because fixing it is not their job. You learn that translation by getting it wrong. You do the obvious thing, you make everything louder, the client is still unhappy and can’t say why, and eventually you work out that the answer was to take something away. Nobody teaches that in an afternoon. You learn it on a leaflet nobody kept.

Set that against what a junior gets offered instead. Options, at speed, in quantity, on demand. And this is where I think the substitution fails, because being handed a screen of things you didn’t make isn’t repetition. It is browsing. The old loop had you commit to a choice, defend it out loud, and find out you were wrong about something you actually believed. That is what a crit is for, and it needs a piece of work you own. Pick one of forty and you have expressed a preference, and you can be made to defend that too. I think that is a thinner kind of being wrong than the one you get from having made the thing yourself. I’m guessing at the size of the gap, though, and an art director who trains people for a living might tell me there isn’t one.

The awkward part sits right there. Directing one of these tools well takes exactly the eye that the old jobs built. You have to look at four passable images and know that the third is closest, and that the type is off, and that the whole set is solving the wrong problem. That is evaluation, and evaluation was never free. It came from years of making the bad version yourself. So the tool works best in the hands of people who are already good, which is a lovely arrangement for them and a closed door for everybody behind them.

There is an economic version of this that studios will discover the hard way. Juniors were never charity. They were profitable, because production hours billed and cost little, and the training happened inside work that was already paying its way. Nobody sat down and decided to run a school. Take away the billable production and the junior stops being a small profit with a lesson attached and becomes a straight cost with a promise attached. Studios are not sentimental about costs and never have been.

Art direction is a taught craft, and none of the teaching has ever needed equipment. People are trained in it on purpose, by looking hard at work they did not make, saying what is wrong with it and why, and being told they are wrong about that. If an eye can be built that way, then a junior with forty options and a good director at their shoulder is not starved of anything. They are in a tighter loop than you ever had, with more attempts in a week than you used to get in a season.

My answer, and you can weigh it against your own years, is that the loop which taught you had a client in it. Somebody outside the studio was unhappy, or the printer rang, or the thing went out into the world and was wrong in a way nobody in the room had predicted. A director’s verdict arrives before the work has met anybody, and it is one person’s. I don’t think the tighter loop replaces that, and I can’t prove it either. Nobody has run that experiment long enough to know, because everybody teaching it now learned the old way.

Replacing the loop deliberately would mean studios funding work that earns nothing so that somebody learns to see, and that is exactly the sort of spending that dies the first time a year goes badly. A few good places will do it, mostly because one stubborn person insists. Most won’t, and will complain later about the standard of the people applying to them, and will not put the two things together.

Look at the smallest job in the studio this week, the one nobody wants and nobody will ever see. Before you feed it to anything, hand it to whoever joined most recently, and let them carry it all the way out to the printer with their own name on it. It will cost you an afternoon and a slice of the margin. Nobody is going to make you do that, and nothing in the way a studio gets run will ever suggest it.

Elongate member. It’s a rod, the attorney who wrote it knows it’s a rod, and they know that you know. It went in that way to keep the claim broad and your search off the shelf where the rods live. Most of what is hard about your work is already in those two words. The office would describe the job as one thing. Search the prior art, meaning everything the world had published before this application was filed, apply the law to the claims, write an action, read the reply, allow it or don’t. Every word of that is true, and it is two different jobs, counted and paid as one.

The first job is retrieval, and I think it goes completely. Prior art is the best behaved body of documents anywhere. It is published, it is dated, it is classified by people who spent whole careers on the classification, and the families are linked across every office that has ever handled the same invention. Art in other languages stopped being a wall some time ago. Searching a corpus like that by meaning rather than by words is precisely what these systems are good at, and you already work with tools that are a long way past keyword lists. Some of you will be glad.

There is one honest complication in that, and it matters later. Searching well depends on reading the claim properly, which is where the elongate member comes back in. A good part of what looks like search skill is really claim construction, and I would put that in the second job rather than the first. That is somebody who has never examined an application telling you what your own craft is made of, and I know how that reads. Examiners would say that building a strategy, and knowing when you have searched enough to stop, is the skill, and that it doesn’t come apart into halves the way I’ve just cut it. Which of us is right isn’t something I can settle. I’m guessing, and I’d rather say so. What I’ll stand behind is narrower: the part that really is finding documents gets cheap fast.

Novelty gets cheap with it, and I don’t think that one is controversial. Does one document, on its own, already contain every element of this claim. That is a comparison. It is checkable, it has a right answer, and a machine that is good at retrieval is good at the comparison too.

Then there is the second job, which is a peculiar thing to explain to anybody outside your trade. You have to decide whether the claimed thing would have been obvious. Not obvious to you, sitting there with the application open in front of you and the whole idea explained on page four. Obvious to somebody the law invents for the purpose. That person knows everything published in the field up to a particular date and nothing after it. They are thoroughly competent and have no imagination at all. They have never seen this application, because at the moment the question is asked it has not been written. The entire value of that construction sits in what it is denied.

Which is why hindsight is the permanent hazard of your trade. Every examiner is warned about it early and still catches themselves at it years later, because once you have read the disclosure you know how the pieces go together. The application has just taught you. Two documents that sat in different classes and never met now look like an obvious pair, and they look that way because somebody showed you the join.

So look carefully at what is being offered to you as help. A model has read everything, including a great deal published after the date that matters. Its whole competence is putting together things that had not been put together. It has read the application, and an application is, among other things, a set of instructions for how to combine the references. Ask it whether a claim is obvious and it will answer, fluently, every time, with a reason that reads beautifully. What you have been handed is the most productive source of hindsight anybody has built, and hindsight is the one error this part of the law exists to prevent. A date filter keeps late documents out of a search. Nothing filters what a judgment is made of.

You have to say why a person in that invented position would have picked up those two documents, put them together, and had a reasonable expectation that the result would work. Your reason then gets read by an attorney whose client is paying to take it apart, and it may be read again by people above you both. A plausible reason produced after the fact is not a weaker version of that. It is the specific thing the requirement was written to keep out.

I don’t want to leave you with the comfortable reading, because I don’t think it holds. Your trade is not squeezed by machines first. It is squeezed by counts. Output is measured, everybody knows their own numbers, and every tool that has ever arrived in that building was absorbed as a raised expectation rather than as breathing room. If a system drafts the action and you review it, the real question is whether reviewing a fluent draft is the same activity as forming a judgment. I don’t believe it is. A convincing document pulls the reader toward agreeing with it, and it pulls hardest when the reader is behind.

The pressure from the other direction is worse. Applicants have the same tools you do. Drafting is cheap now, filings rise, and any rejection you write can be answered at length by the same afternoon, at a length nobody would have paid for before. The load grows at both ends while the part of your work that used to absorb it is the part being taken away. How that resolves is beyond me, and I am not going to pretend otherwise. It is the part of the problem that never turns up in anybody’s plan to speed things up.

There is a version of your office where the second job becomes the whole job, and the counts fall to match, because a decision that has to be defensible in writing is worth more than three that are not. There is another version where the drafts arrive faster and the counts rise, and the judgment quietly becomes a checkbox on somebody else’s document. Nothing technical decides which of those happens. It gets decided by whoever sets the numbers, and the numbers have never once been set to favor the slower half of the work.

Watch what your own hand does while you read the reason to combine in the last action you signed, out loud, to yourself. If it goes to the application, the thing you were judging has just taught you the answer. Then be honest about how much of your own pile gets decided that way. I have no idea what the answer is, and I notice that everybody outside your office is certain they do.

The spec came in on Monday and you read it twice. The second read is where you start subtracting. It wants years in a tool that has not existed that long, plus the industry, plus the certificate, plus the office days nobody wants, at a band the last two people you sent turned down flat. You have had this spec before under a different job title, and you will have it again on Thursday.

So let me put my claim early, because it is not a kind one. Everything visible in your week goes. Sourcing goes, and half of it went years ago, when the databases got good and every recruiter alive started running the same search against the same people. The screening call goes, or at least the part of it that checks a person against a document goes. Ask a recruiter what the rest of that call is and you get the same three answers: whether the person actually wants to move, what the notice period really is, and whether they will take the counteroffer when it comes. None of that is on a document and none of it goes. Scheduling goes, and so does turning somebody’s history into your agency’s template, which is a mercy nobody will mourn. The weekly pipeline report goes. The first message to a stranger goes too, and I want to come back to that one, because it isn’t going the way people assume.

You are the only person in the whole arrangement who hears both sides say what they actually mean, because neither of them can say it to the other.

Start with the spec, because it is not a description of a person. It is what a manager writes when they have a problem they haven’t diagnosed yet. The honest version of that document is a handful of sentences nobody puts in writing. The last one left and I think some of that was me. I want somebody senior enough not to need managing, because I have no time to manage anybody. My band sits under the market, finance won’t move it, and I am quietly hoping you’ll find somebody who doesn’t know that. We tried to promote from inside and it didn’t work, and I can’t put the reason in an email.

You know those sentences because managers say them to you. They say them to you and not to the candidate, not to their own people team, and often not out loud to themselves. A model reading the written spec will do exactly what the document says, beautifully, at any hour, forever. Every name it returns will match. The role will still be open in the spring, and the failure will be recorded as a tight market rather than a wrong document, which is the most comfortable explanation available to everybody in the chain. Now the next spec gets drafted from the last one, which was also wrong, and the wrongness sets like concrete.

The other half is harder and it is the one I would defend in a room full of your colleagues. The person you most want is not looking. Your good placements are people who were fine. Not miserable, not underpaid, not passed over, just fine, and a fine person has no reason to take your call and every reason to ignore it. What moves somebody like that is almost never the job description. It is a person they have some reason to trust saying, without decoration, that the thing they are mildly tired of does not have to be permanent.

Which is why the story about automated outreach is backwards. When it costs nothing to write a warm, personal-sounding message to a stranger, everybody writes one, and the good ones drown alongside the bad. The inbox of every decent engineer is already full of notes that use their first name, mention their last project and mean nothing at all. Each of those spends a little of the credit the whole channel runs on. A better message won’t save it. Being somebody who is already known, or introduced, or vouched for by a person whose opinion carries will, and there is no cheap way to become that.

So you stand between two parties, each holding a fact that would help the other, and neither able to say it. The manager cannot tell a candidate that the last person left because the job is hard to do under him. The candidate cannot tell the manager that she would take a sideways move and less money to stop running a team, because in that room it sounds like no ambition. Both of them will tell you. Not because you are trusted exactly, but because you are the only one in the conversation who is not the other side. That is the position, and it is not charm or people skills. It is a fact about who is able to hear what.

The trouble is that the act worth paying for is the one your fee punishes. Telling a client their spec describes nobody slows the placement, irritates the person who approves the invoice, and gets you paid later or not at all. Sending four more names against the wrong document gets you paid this month. Anybody who has worked on contingency knows which of those the arrangement rewards.

I don’t think the trade survives at anything close to its present size, and I would rather say so than dress it up. Most recruitment is the volume model, and the volume model is precisely what just got cheap. Agencies built on sending many people quickly against a spec nobody questioned are being paid for the activity that is disappearing. The reputation is part of the same story. Your trade got disliked by being the loudest thing in everyone’s inbox, and every tool it was sold made it louder.

And here is where I am guessing, so treat it as a guess. I don’t know whether the second half survives the flood either. It is entirely possible that people stop answering anybody they have not already met, and that the call which gets picked up comes from a former colleague rather than from a recruiter at all. If that happens, the position I have described still exists and simply stops being a job. Nobody selling you software knows the answer to that one, whatever the pitch says.

Picture the manager who sent you the last spec you worked, alone in his own office, saying out loud the sentences he would never put in an email. Get them into his voice rather than yours. Then read them back to him on the phone as a question, and notice which line he doesn’t argue with. Nobody in your trade writes that version down, and most of you have it composed by the second paragraph anyway.

Not the columnist, and not the desk. This is for whoever on your paper still has to ring a person up and say the thing that person does not want to hear, out loud, with their own name attached to it.

Ask a room of reporters what the job actually is and you get two answers, and both are right. Knowing what is a story. Checking whether it’s true. The rest of the work is not now handled by something else, whatever anybody claims, and you’d stop reading if I said it was. What I want to do instead is take one task, the one you’d hand to somebody else if you were allowed, and look at why it stays.

Start with where copy comes from, because the trade is honest about this in private and coy about it in public. Most of what fills a publication arrives from somebody who wants it in there. Releases, embargoed reports, listings, council agendas, court lists, company filings, the wires, the trade body with a survey, the exclusive that is really a placement. None of that is new and none of it is shameful. It’s the supply, and no newsroom has ever run without it.

What has changed is that the supply is now bottomless and much better dressed. Every organization you deal with can produce endless clean, quotable, on-message text at no cost. The press office answers you inside the hour, at any hour, in complete sentences, and says nothing at all. Your inbox is fuller than it has ever been and emptier at the same time.

Nothing in that flow will ever hand you the story somebody is keeping out.

That’s not a complaint about public relations. It’s arithmetic. A story worth the name exists because a person has decided it should not exist, and everything that arrives through a channel arrives because somebody wanted it to. The material you need is not in the record. It’s inside a person who has chosen not to say it, and inside another person who will lose something when it’s said. No feed carries that. Nothing can be summarized out of a pile of documents that were never written.

So, the task. You have a document, or two sources, or the strong sense that you know what happened. Now you have to put it to them.

Everything in that call is a decision, and none of it is retrieval. How much do you show. Name the document and you have told them who leaked it, and that person may lose their job by teatime. Hold it back and they say it’s nonsense, and you have burned your one call. When you ring matters, and there is a slightly grubby art to catching somebody before their advisers are awake, which the trade knows about and rarely writes down.

You say who you are in the first sentence, which is the moment you give away every advantage you had.

Then you say the sentence, and the reply is evidence. Not just the words. The pause before the words. The denial that is too specific, and answers a question you had not asked yet. The one that is technically exact and covers a slightly different thing. The offer of a friendly briefing on a different subject entirely. None of that is confirmation and nothing runs on a pause. It is a direction. It tells you which document to ask for next, and which of the four people who could have signed it is worth another call, and you get it live, in somebody’s voice, once.

There is a version of this call that costs you more. The knock on a door where somebody died last night. Most people outside the trade think that’s ghoulish and some of it is, and reporters who do it badly are a decent part of why the public dislikes you. What gets left out is how often the family wants to talk, and how much it matters that the person who turns up is careful, and unhurried, and prepared to leave. That cannot be sent. Somebody has to stand there and be told to go away, or not.

A great deal of verification is tooling and always has been. Checking a register, pulling a filing, working out where a photograph was taken, matching a face against an older image. That work has got faster and better and you should use every bit of it, and anybody who tells you otherwise is romanticizing a trade that never had time for romance.

The bigger problem isn’t the tooling. Making difficult calls is not, on its own, a full week. Most days are the routine flow, and a job made entirely of exceptions has too few hours in it to be a job. Worse, a publisher can decide it doesn’t want the exception at all. Stories that people fight to keep out are slow, legally expensive, and no better for traffic than the cheap stuff. That decision has been taken in real newsrooms, more than once, and nothing about the current machinery makes it less tempting.

There’s one more thing the trade knows and the public doesn’t, which is that the call itself can lose you the story. Give somebody proper notice and you have also given them time. Time to reach a judge, where that works, time to put their own statement out first, time to start deleting things. You still make the call. The rule that you put it to them is not a courtesy, it’s the difference between journalism and something else, and it costs you real stories every year.

None of that makes the exception safe, and on its own it is not a living. It is only the one thing no supply can produce, because supply runs on somebody wanting the thing published, and this is the single exchange where everyone involved would rather it wasn’t.

There’s a release sitting in your inbox now that reads beautifully and tells you nothing. Find the line in it that somebody argued about before it went out, the one that got softened on the way. Then work out who lost that argument, ring them, and ask what the sentence said first.

Once a year you hand the whole thing to an accountant who charges more than you do and gets thanked by name for it. That is the trade’s oldest complaint and it isn’t what this post is about. This one is for whoever keeps the books for a business with a yard, a van and an owner who still answers his own phone. Not the finance department of somewhere large. You, and the handful of clients whose whole year you carry around in your head.

Ask anybody what a bookkeeper does and they’ll say it’s entry. They picture typing numbers off paper into a screen, and they have just worked out that a machine can do that, and they are pleased with themselves. You lost the typing years ago. The bank feed arrived, then the rules that code the regulars, then the phone that reads a receipt off a photograph, and the trade quietly reshaped itself around all of it long before anybody found it interesting.

So let’s set out what a month actually is. Reconciling. Chasing the client for the paperwork they have now promised more than once. Coding the odd ones that no rule catches. The tax return, the payroll, the supplier run, the polite letters to people who haven’t paid. Month end journals so the numbers mean something. Then, once a year, that handover to the accountant.

Most of that is being eaten and you can see it happening from where you sit. The regular coding happens without you now. Reconciling is going the same way. The software already flags a duplicate payment and an amount that doesn’t look like the others, and it does it faster than you would.

Here’s the part nobody has ever put a price on.

Once a month, you are the only outsider who looks at that business honestly, up close, one transaction at a time. Think about who else sees it. An accountant sees it once a year, tidied up, when everything in it is already history. A bank sees a balance, the customers see a van, and the owner sees what he needs to see, because he is inside it and running on the story he told himself in January. You see the thing itself, in detail, before anyone has arranged it. That isn’t a skill, it’s a position, and it is made of three plain things. Close enough to see the detail. Often enough to know what has changed since last time. Holding the keys to all of it at once, the bank, the invoices, the payroll, the ledger. Take away any one of the three and the position is gone.

It is also how you know, in the first week of the month, what’s going to go wrong in the summer.

The account looks healthy. It looks healthy because a large slice of what’s sitting in there is tax the business collected on the government’s behalf and hasn’t sent on yet. That money was never his. He has just quoted a job on the strength of the balance, or put a deposit on a van, and nobody in his life has said the sentence out loud. The software will not ring him. It will produce a correct report, on time, and file it in a folder he doesn’t open.

You ring him. That’s the job. It has never appeared on an invoice.

The rest of the list is the same shape. A direct debit that’s been running since he cancelled the service years ago. A supplier whose prices have crept, one small rise at a time, in a way no single invoice shows. Drawings running ahead of what the business actually earned, which nobody notices until the year end and a tax bill. A payment to a company nobody in the office has heard of. An employee’s own bank details sitting quietly on a supplier record.

A lot of that noticing is being automated too, and I think more of it will be. The flag on the odd payment is real. The cash forecast that says the tax bill is coming is real, and it’s better than the one most people do in their heads.

What isn’t in any of it is the rest of what you know. That he isn’t answering emails this month. That his wife does the invoicing and something has gone wrong between them. That the big customer has gone quiet, and last time that happened he borrowed from his brother and didn’t tell anybody. If you go in on a Tuesday, you know the yard is quieter than it was, the good foreman has left, there’s a new van outside that nobody mentioned, and the son is suddenly around a lot. None of that is in the ledger. All of it turns up in the numbers later, and by then it’s the accountant’s problem and it’s too late to be anybody’s opportunity.

Most bookkeepers now work from a screen at home and never set foot in the place, and the yard is not the standard case. It doesn’t have to be. The position was never standing in a yard, it was being inside the same business every month with everything in front of you, and a bookkeeper who is in the bank feed on the fourth of every month has that. What she has less of is the accidental signal, the things nobody would think to send her. So the good ones ring for no reason at all, which is not a soft skill, it’s a sampling method. Anybody who has worked both ways will tell you the remote version needs that call more, not less.

That leaves an uncomfortable piece of arithmetic, and it’s the reason I don’t think this trade is safe just because the noticing is. Your fee is priced on the part that’s disappearing. Transactions, hours, entries, the volume of things touched. The position isn’t any of those. It’s a standing arrangement to be inside somebody’s business every month, whether or not there turns out to be anything in there. A thing like that has to be bought before it pays, or it isn’t in place on the month it matters. Nobody buys it that way. They buy transactions and get the position thrown in, which worked fine while the transactions were expensive. A buyer who has just watched the visible half of the work fall away is not in a generous mood about the invisible half.

Open last month’s file for the client you know best and read down it until you reach the one sentence you’d say to the owner if you had him on the phone with nothing else on the agenda. It’s in there, and it will take you a minute. Send the reports without it, the way you always do, and count how long it takes anybody to ask you for the sentence. I have a view about what that says about your fee, and I’m going to keep it to myself.

An apology first, on behalf of everybody who repeated it. It became a fashionable thing to say at conferences, years ago, that we should stop training you, and the people who said it most confidently had never set foot in a reporting room. Meanwhile the lists got longer, the studies got bigger, and there were never more of you. You have been more polite about this than we deserved.

So let me concede the part your trade is tired of arguing about, and then say the thing I actually came to say. Some of the detection is good. On narrow, well-defined tasks a model finds what it was trained to find, reliably, and it doesn’t get worse at the end of a long list. Anyone who has watched a flagged bleed jump the queue and get a patient into a scanner sooner is not going to be talked out of it, and shouldn’t be.

That is a claim about detection. It is not a claim about reading, and the difference is the whole post.

Your day is protocolling, priors, the phone and the dictation. Deciding what the scan should even be, from a request that says abdominal pain, query cause, and nothing else. Pulling the old study and working out whether the thing is bigger. Being interrupted by a surgeon who wants a look now. Telling whoever is at the scanner whether to add the extra sequence. Then the meeting where the oncologists and surgeons sit down with the images and somebody has to say, out loud, what they think that is. Then a list that is never empty, sent overnight to a service in another time zone and coming back for you to live with in the morning.

A great deal of that is being automated and some of it should be. Measurement, comparison with the prior, the segmenting, the reconstruction that lets a scan be quicker and cleaner. The draft on the normal chest film. The chasing of follow-ups that everybody hates and nobody has time for. Good. None of that is the part I came to defend.

The report doesn’t go, because a report is not a description of an image. It’s a signed instruction about what happens next to a person, addressed to a particular reader who will act on it. Change the reader and the same pixels produce different sentences. Change what the referrer can do about it and they change again. That isn’t sloppiness, it’s the job, and it’s why two competent radiologists can write two different true reports on the same study.

Which brings me to the small thing you saw that had nothing to do with why the scan was ordered.

You know the one. It sits at the edge of the field, it is almost certainly nothing, and you have to decide whether to put it in writing. Mention it and you start a chain of appointments and repeat scans and a frightened person who now has a thing, and sometimes a needle goes into something that was never going to harm them. Leave it out, and if it turns out to matter, the name on the report is yours. Nothing in the image tells you which way to go. You settle it with the patient’s age, the referrer, what the follow-up will do to this particular person, and what you can defend later.

Then there’s the phrase. Clinical correlation is advised. Cannot be excluded. Everyone in medicine makes jokes about that language, and most of the jokes miss what it’s doing. Sometimes it is genuine uncertainty written down honestly. Sometimes it is a person moving a decision to somebody with information they don’t have. And sometimes, let’s be adults, it is a person covering themselves. All three are the same act, which is placing the weight of what comes next somewhere specific.

A machine will flag the incidental. It will flag it every time, because flagging costs it nothing. It has no clinic to fill, no patient who can’t sleep, no colleague who has to do the biopsy. The cost of a flag lands entirely on people who are not the flagger, and somebody in the building has to be willing to take a flag and do nothing about it, in writing, with their name on the line.

That is what a tool of this kind actually does, and it is worth saying plainly. A tool that points at things does not settle anything. It moves the question one step along, from whether there is something there to what happens now that somebody has been told there might be. The second question has never been an imaging question. It gets answered in clinics, in waiting lists, and in a conversation with a frightened person who will remember it for years.

The argument that should worry you isn’t about machines at all. If the read comes with the scan, why does the answerable person have to be a radiologist? Let the clinician who ordered it take the report and own the consequence. They know the patient, which you usually don’t. Nothing about the position requires your training, only somebody’s signature.

I think that’s the actual threat to your trade and I don’t think it’s silly. My answer is only that the position doesn’t disappear, it moves, and it lands on somebody with less imaging training and less time, who will then need somewhere to send the hard ones. Where those go is not a technical question. It’s a question about who a health system is willing to have blamed.

And I’m not going to end this kindly, because the honest reading isn’t kind. Every tool your trade has been given has raised what a day is expected to contain. The flags arrive faster than the reporting time you save, and the being-answerable is not a moment at the end of the study. It’s the posture you hold for the whole list, and it’s what makes the ordinary parts tiring. That doesn’t get lighter when the ordinary parts go. It gets more concentrated.

Pick the last incidental you decided not to chase, one of the small ones you settled without slowing down. Work out what you actually used to settle it, and how much of that came from the images. Then answer something I can’t. Who else in that building could have made the same call, and still be willing to explain it a year from now?

Document management. Diary management. Drafting support. Assisting with disclosure and liaising with counsel. Every line of that was accurate when somebody wrote it, and it is still accurate, and none of it is the job.

Start with the honest part, because the trade already knows it and pretending otherwise wastes your time. First-pass review has been going for years, since long before any of the current noise. Software that ranks documents by how likely they are to matter turned out to work well enough, and it cost a fraction of a room full of people. What’s arrived since then takes the next layer. A serviceable summary of a long transcript. A chronology built off the dates in the file. A first draft of the letter that follows the template, the index, the schedule, the routine instrument. Checking the citations. Working out the deadline from the rule.

Some of that is a relief and some of it isn’t, because a fair amount of it was the billable part, and a job that loses its billable hours has a problem whatever else is true about it.

So what’s left isn’t a task at all. It’s a location.

Take the day it goes wrong. Counsel is on their feet, a witness has just said something that doesn’t match a document, and the room has about the length of one sentence to decide what happens next. A hand comes back without a head turning. Whatever your jurisdiction calls the paginated set everybody in the room works from, you built it. You know the letter is in there. You know it’s the copy with the earlier date, because the argument about which date was right went on and on. The page arrives in the hand. Nobody says anything about it afterward.

That is not retrieval. Retrieval is the part a machine now does better than any of us, and if the question had come in writing with time to answer it, nobody would need you. The speed comes from somewhere else. You have been inside that file since the start, so the document, its wording, the fight about it and the other side’s copy with the wrong date are already in your head. Nothing is being searched. It’s being held. And the knowledge on its own is only a memo that somebody reads later. What makes it worth paying for is that it’s in the room while the room is still moving.

And that is worth saying carefully, because it cuts both ways. The months in the file are exactly what’s being taken away. If the review is done by something else, the person standing in the room on the day has a thinner picture of what’s in there, and nobody in the trade has worked out yet what that does. I can’t either. I suspect the answer is not comfortable and I’d rather say that than pretend the residue is safe.

The trial is the sharp version, but the ordinary version is everywhere in your week. A witness turns up early and has to be met at reception, kept away from the others, given a glass of water. Somebody has to notice that this one is about to come apart, and say so to counsel before they reach the door. A client arrives without an appointment because the letter frightened them. Completion runs late and needs a person holding the signature pages and the room. On the day of the deadline the filing system goes down, so a human being walks it over, and the woman on the counter knows your name.

None of that is written down anywhere as a duty. It happens because you were there.

A great deal of this trade now works from a kitchen table, and the sky did not fall. Bundles get built remotely. Hearings run on a video link, witnesses get met by nobody at all, and mostly it’s fine. If presence were doing the heavy lifting every day, none of that would have worked as well as it has. So the claim isn’t that you have to commute. It’s narrower. It’s that the work concentrates onto particular days and particular rooms, and on those days somebody has to already be there and already know, because a room moves at a speed that has no queue in it. You can’t summon that person. Whoever is present is the only candidate.

I don’t think this post has earned the right to be encouraging, so here is the part I dislike about your position. The title says assistant. You have been at the firm longer than the associate you are assisting, and you both know it, and neither of you says so. Pay follows titles, not work. And the firm can only see you through the parts of your week that leave a record, which are precisely the parts being taken out. The tasks that leave are the ones that were easy to see and easy to count. What’s left is real, and it is nearly invisible from head office, and those two facts have never sat comfortably together in any trade I’ve looked at.

The firms that work this out will pay for presence on purpose, and describe it properly, and stop calling it assistance. What that costs them is somebody who is not billing anything while nothing is happening, which is a harder line to defend on a budget than any of the ones being removed.

For a week, keep a note of every time somebody came and found you in person instead of writing it down. Note what they wanted and whether the answer existed anywhere in the file. That list is your actual job description. Now notice that you have never once put a single item from it on a timesheet.

The first piece of your work to go was gisting. Somebody used to send you a pile of documents in a language they couldn’t read and pay you to say roughly what was in them, so they could work out which ones mattered. Nobody buys that now, and honestly nobody should.

You saw this coming before the rest of us did, which is why I want to start the series with your trade rather than a safer one. That roles hollow out while the job title survives is not a new claim on this blog, and the only thing these thirty posts add is the specific trade instead of the general shape. So let’s be specific. The volume work is going. The first drafts are going. Somebody still has to decide what a sentence is for, and that part was never really about words.

The version going around is that you’re a proofreader now. There’s something in it. Post-editing pays less per word than translating and takes a strange kind of attention, and you can only do so much of it in a day before your eye stops catching things. But the squeeze didn’t start here. The memory discounts were already shaving your invoice years before any of this, the grid of percentages that decides a repeat is worth a fraction and a near match is worth a bit more than that. The machine walked into a trade that had already agreed, in writing, that some of your words were worth less than others.

Clients mostly don’t know this part. A bad translation used to announce itself. It arrived stiff and strange, and the client could see something was wrong even if they couldn’t say what. That was useful. The output you get now is smooth, confident, idiomatic, and every so often it has quietly dropped a negative or swapped a term for its near neighbor. Nothing on the surface tells you which sentence to distrust, so you have to read all of them properly, which is the tiring half of the job. You are being paid less to do the harder part.

One decision holds the whole of this trade, and I want to put it in front of you before anything else.

You’re going into a language with two words for you. One you’d use with a stranger, one you’d use with a friend. On an account you have had for years this is not a live question. It was settled once, it sits in the style guide or in the termbase, and you would have to go digging to find out who decided it. The live version is the new client, the file that landed on Friday, no guide, nobody to ask. The source has one word for you and nothing in it says whether this company talks to its own staff as adults or as employees. You won’t find the tone of the head office in there either. You put the question to them and it comes back as “either is fine.” It is never either. The same client will write in six months to ask why the handbook sounds like a letter from a bank.

A machine will pick one. It will pick consistently, at speed, with no idea it made a decision, and once that choice has run through a company’s handbooks and safety notices and internal mail, it is the company’s voice. That is register, the distance you keep from the person reading, and it isn’t a small thing to set by accident. Somewhere in that company a laminated card is screwed to the wall beside a machine, telling the man on the night shift to keep his hands clear, and it speaks to him the way a bank speaks to a stranger. Nobody chose that. It got chosen.

That is the shape of the whole job. The source text is not a brief. It’s one attempt at the thing, made by somebody with a different reader in mind, and your work is to notice what each sentence is doing and then find out whether it can be done that way over here. A safety instruction that assumes a building code the other country doesn’t have. A contract clause naming a remedy that doesn’t exist in the target jurisdiction. A line telling the reader to ring their family doctor, in a country where that isn’t how anybody gets seen. A joke, which has to be replaced by a different joke or dropped, and either way somebody has to decide. None of that information is in the source and none of it is in the target language. It sits with the person who has been in both rooms.

The machine is good, and the trade doesn’t always say so out loud. On clean, well-written, in-domain prose it produces something a tired human at the end of a long list would be pleased to have written. Saying otherwise in front of a client who can read both texts loses you the argument and some of your credibility with it. The trade’s real complaint isn’t that the output is bad. It’s that judging whether it’s bad takes a professional, and professionals are exactly what the buyer thinks they’re no longer paying for.

And I don’t want to leave you with the comfortable version of this, because the comfortable version is wrong. Saying that the interesting work survives is no use to somebody whose interesting work was subsidized by the boring work. The volume paid the rent. If the pile at the bottom goes, the top doesn’t stay where it was, it just gets more crowded. Some people reading this will not have a trade at the end of it, and it won’t be because they weren’t good.

It holds where being wrong costs money or gives offense, and it holds for a reason worth being clear about. It isn’t that the machine can’t produce the sentence. Often it can. It’s that somebody has to know what the sentence is for, who is going to read it, and what they’ll do next, and then be the person who settled it. Interpreting is a different trade with a different answer, and I’m not going to pretend I can cover both in one go. But the written version of this is a job about two worlds that can’t hear each other, and the words were always the smaller half of it.

The next time a client tells you either is fine, don’t let it go past. Write down which one you picked and the one line of reasoning behind it, and leave it in the file where the next translator will find it. Nobody is going to ask you for that line. It is still the only record anywhere of how that company decided to speak to its own people.