Journal · Product · 6 October 2026 · 17:48 SAST · by Jeremy Rich

Context-sensitive help, on steroids

At the end of the Utrecht project in 2001 I spent my last few weeks writing the user manuals, and I have a quiet suspicion that the people who needed them most never opened them. That’s no criticism of the bank’s staff. A manual answers the questions its author expected, in the order he expected them, and when you have an account open and a client on the phone you don’t have time to find the chapter. Context-sensitive help was supposed to fix that, and mostly it opened the same manual at a slightly better page.

The assistant is what I wanted back then. There’s an Ask the assistant button in the top bar of every screen. You type a question in plain words, and the screen you’re on becomes part of the question.

Asking how

The first kind of question is the one the manual was for. “The amount on a transaction is wrong, what can I do?” “How do I write off a bad debt?” “Why can’t I release this payment file?”

You get a few sentences back, with links to the guide articles the answer came from, so the detail is one click away if you want it. When the answer names something you can do, it links to the screen where you do it. The assistant reads the same guide we write for people, and the public half of it is on this site, so the help and the documentation can’t quietly drift apart.

Asking about this

The second kind of question is where it earns its keep. Open an account’s transactions and ask “what is the largest card purchase?”, and it knows which account you mean without being told. Ask it to “chart card purchases per month” and you get bars, and clicking a bar takes you to the transactions behind it. Then “show me the accounting entries for that one”, and you get a table with every row linked to its record and a link to the full list.

The screen also tells the assistant what it’s showing. Ask “can I add an additional investment?” on a deposit that still has a settlement pending, and the assistant sees the pending settlement and answers with it in mind, instead of reciting the general rule and leaving you to find the exception the hard way. Five screens send their facts today, and the list will grow.

An answer comes back as text, a single number, a table or a chart, and you can watch each lookup it makes while it works, with a Stop button if it’s heading somewhere unhelpful.

The bit the accountants will like

The model never does the arithmetic. It decides which list to read and how to group it, and then our own code reads the rows and adds them up and hands the total back for the model to put into a sentence. That’s why every total can show you exactly which rows it came from. A language model that sounds very sure about a sum is the last thing a bank needs, so we didn’t give it the chance.

It reads with your permissions and nothing more. If you can’t see the books, neither can it, and it tells you so in one sentence and answers with what it could read. It never posts, approves or changes anything. An answer stops after ten lookups or sixty seconds and says why, which is usually a sign the question wanted to be two questions.

Who asked what

Every question is kept: each lookup the assistant made, what it asked for, how many rows came back, and exactly what it saw, including the facts the screen sent with the question. A checker with the review permission can open anyone’s conversation and replay it in order, and that keeps working after the assistant is switched off.

The assistant is off until your organisation switches it on, and asking and reviewing are separate permissions. The question and whatever is on screen go to the model with it, customer details included, so your organisation writes the note your staff see about that. Today the model is Anthropic’s Claude, on your own key if you’d rather name one.

I’d still write the manual. The difference now is that something reads it for you.

Written by Jeremy Rich.