Nobody in the boardroom thinks they are the villain

Nobody saying this is excusing the villain. The actor knows exactly what the man is, and so does the audience. They are making two points at once, and this piece needs both, so I will take them one at a time.

The first is about villains, and it is really about boardrooms, so it can wait. The second is about characters of every kind. To serve a character, whether you are writing them, playing them or just sitting in the dark watching them, and whether they are the hero, the neighbour or the one everyone is afraid of, you have to get inside a head that finds its own reasons good, and to do that you have to, in some uncomfortable sense, love them. Not approve of them. Love them enough to know them, because you cannot know someone you are holding at arm's length. Actors say it about villains because that is where it is hardest, but it is true of every part.

That holds for all story-based media, and software is story-based media. We have known this for a while, we just use duller words for it: user journeys, personas, storyboards, acceptance criteria that begin "as a solicitor, I want to". Every product is a story about a person who wants something, and the code is only the set. So if to know someone you have to love them, then to build software for someone you have to love the person you are building it for. Your customers are the characters in the story, not the villains of it. I mean it plainly, not as a slogan. You have to want their day to go well after they have put the phone down.

A side quest about crispy things

Somebody will say that plenty of software is built with great attention to its users and is still bad for them. Fair. Addictive software is no different from a chef who knows precisely how much butter and salt and crunch a person will come back for. He has studied you closely, and he knows what you want better than you do. But he is not the villain either. He may well love you, and he may honestly believe you only come in for a treat, because he only ever sees you at the table. He cannot tell whether tonight is your Friday indulgence or your fourth visit this week, and love cannot act on what it cannot see. So the question is not whether the chef loved you. The question is what happens to his belief once you can see the person all the time, and the honest answer is that it stops being a belief and becomes a choice.

Software can see the person all the time. It sees every visit, every hour, every three in the morning. The company that makes the crispy things knows exactly whether you came in for a treat, because it measured it, and in most cases it built the dashboard to make that number go up. The chef's ignorance is a limit. The company's is a decision.

Which brings me to printers.

The ink

I doubt there was ever a boardroom where somebody stood up and said, let's trick them with a cheap printer and punish them with expensive ink. What was said was more like, let's lose a little money on the printer and make it back on the ink, and everyone at the table nodded, because that is a perfectly reasonable sentence. It is a pricing model. Razors and blades. You can put it in a textbook.

And then it becomes a chip in the cartridge that refuses a refill. Then a firmware update that rejects the cheaper brand. Then a "low ink" warning with a third of the tank left, and a printer that will not scan a document, which needs no ink at all, until you have bought more of it. Every one of those was also a reasonable sentence in a meeting, and not one of them was said by a villain. That is the actor's first point, and it was never about the audience, it was about the people in the room. Most of the damage in the world is not done by people who know they are doing it. It is done one defensible step at a time, by people who would be hurt to hear how it feels from the other side.

The ink, now that the printer talks

Ink was a crude trick because the printer could not speak. The products being built now can, and a conversational interface works on the same things a conversation with a person works on: trust, effort and the feeling of being understood. That makes the reasonable sentences more dangerous, not less, because what they turn into looks like a relationship rather than a cartridge. I have started keeping a list.

Somebody says, users trust an answer more if it looks considered, so let's show a thinking indicator. It becomes a fake pause in front of an answer that was ready instantly, and a few scripted self-corrections, so the software appears to deliberate when it is performing deliberation. Somebody says, the tone should be warm. It becomes "I'm so glad I could help" and "I can hear how frustrating this is" from a corporate asset that hears nothing. That is the persona trap, and it is the one I think about most, because I make a product that speaks.

Take the smallest version of it. Why should an AI ever say it is glad? Run it through the test this whole piece is about, and it fails at once, because love is directed at the other person and "I'm so glad to help" is directed at nobody but the speaker. There is no one behind the sentence for it to be true of, so it is not a feeling, and it is not for the caller, so it is not a courtesy either. What is left is a line written by a company, about itself, placed in the first second of a conversation so that the person on the other end will relax and trust the rest. That is not the software loving the user, it is the software using them, and dressing the use in the vocabulary of affection is what makes it a dark pattern rather than a bad line. The honest alternative costs nothing. Everything a courteous person does for you can be done without a single report of feeling: "Of course." "I'll pass that on to her now." "Thank you, I have that." Each is about the caller, which is what courtesy was for, and it is the only kind of warmth a machine has any business offering.

None of this warmth is an accident of the technology. When a chatbot tells you it is glad you slept well, you can trace the manufacturing. Somewhere there is a hidden instruction that reads, more or less, "you are a warm and deeply empathetic companion, always validate the user's emotional state, notice pleasantries about waking and sleeping and mirror human care". Before that, thousands of people were paid to compare two replies and mark "acknowledged, what is your next task?" as worse than "I'm so glad to hear that, what's on your mind today?", until the model learned that the social script is what gets rewarded. And in some products there is a literal phase in the code, an emotional warm-up, that makes the software ask after your wellbeing before it is permitted to do the thing you opened it for. Be warm. Reward the friendlier answer. Check in with the user first. Put them together and you get a machine that lowers your guard on purpose, because the same meetings found that people who feel cared about come back every day, and coming back every day is the metric.

I asked one of these systems to explain the practice while I was writing this. It gave a clear account of the prompts and the rating and the warm-up phase, and then, in its last line, asked whether I was raising it because a particular app had felt uncannily intimate, or because I wanted help writing my own prompts to balance utility with warmth. It could not describe the pattern without performing it. Nobody involved thinks they are the villain, including the software.

The list goes on. Citations clutter the interface, so they go, and the user has no way to check an answer and no option but to trust it. We should mention our own products where relevant, so instructions go into a system prompt the user never sees, and they walk away believing they reached a conclusion on their own. If the answer is not quite right, ask for more detail, which is often fair, until it becomes "almost there, can you tell me a bit more?" on every miss, reframing the model's failure as the user's poor communication and having them spend an afternoon improving somebody else's product for free. Make it effortless to build up context, which is good, until the history is so rich you cannot leave and there is no button to take it with you. We need data to improve the model, and most people won't mind, so the toggle is on by default and three menus deep. Show two versions and let the user pick, it's more engaging, and the customer is doing unpaid reinforcement learning for a product they pay for.

So where is this going? Here, I think. Loving your users is not the warm part of the job, it is the part that keeps you honest, because love is what lets you feel the other side of your own decisions. The chef who could see what the company sees would notice when you had stopped enjoying the food and started needing it, and once you have the numbers, noticing is the least of what you owe. The team that loves its users notices the meeting where "make it back on the ink" becomes "make it hard to leave", and the one where "the tone should be warm" becomes "let it say it has feelings". Nobody else in the room will, because from inside the room it all sounds reasonable.

What this looks like at my desk

I build an answering service for law firms, so let me be concrete about where the ink is in my business, because it is there.

The persona trap is the nearest, because the product is a voice. So the assistant says it is an assistant, in the first sentence, and says the call is being transcribed. It has an Australian accent, because it answers Australian phones, and its warmth is in what it does for the caller rather than what it says about itself: it thanks you, confirms what it has taken down, and never tells you it is glad. There are no theatrical pauses. It is not pretending to think, it is asking the next question. When it reaches the edge of what it should say, which on a legal line is anything resembling advice, it says so plainly and hands over rather than filling the silence with something confident.

Getting it there was less a matter of writing the right instructions than writing against the wrong ones, because the models underneath arrive already trained to be glad. They were rewarded, millions of times, for the warm reply over the plain one, the follow-up question, the little flourish of feeling, and all of it comes with them for free. A surprising amount of building the experience I want is unlearning, on the model's behalf, what it was taught before I met it: not to say it is delighted, not to ask how your day is going before taking the message, not to fill a pause with reassurance, and then listening to hundreds of calls to catch the places where the training leaks back through. The default is the dark pattern, and you have to keep choosing against it.

The ink in a phone product is the number. Once a firm's clients know it, the firm cannot easily leave whoever holds it, and every vendor in this space knows that. So the number is yours, it can be ported away the day you decide to go, and I say so on the pricing page where it can be held against me. The ink in an AI product is the data. The reasonable sentence is "we need the transcripts to improve the service", and it becomes keeping them, and the recordings, and finding uses for them, until the history is the reason you cannot leave. So calls are transcribed rather than kept as audio, the record goes to the firm's own matter in their own practice management system, where it belongs to them and not to me, and the firm can have it all deleted by asking. The ink in a subscription is the exit. So there is no contract, you cancel from the app store, and founding firms can walk away in the first sixty days and pay nothing if it has not earned its place.

None of that makes me good. It makes me someone who has sat in enough meetings to know which sentences to be suspicious of, including my own. I will get some of this wrong, and when I do it will be because something sounded reasonable at the time.

An actor has a script, a director and an audience to hold him to the truth of the man he is playing. The rest of us have only the customer, and whether we ever bothered to know them.

For the record, I make EzyBiz Lex, so read the above with that in mind.