The AI in Scanner

People sign agreements they have not read. Not out of carelessness: out of arithmetic. The document is long, the decision feels small, and the cost of reading is certain while the cost of not reading is hypothetical.

Scanner is a document-scanning app. Its distinguishing feature is that it uses artificial intelligence to highlight the most critical clauses in a legal agreement, directing the reader to the sections they should read before signing. That sentence is short, and almost all of the product work is inside it.

The problem is attention, not comprehension

It is tempting to frame this as a reading problem, where the answer would be a summary. I do not think that is right. Most people can understand a termination clause perfectly well once they are looking at it. What they cannot do is find it, in 12 pages of even grey text, when nothing on the page indicates that paragraph 9 matters more than paragraph 8.

So the unit of the product is not the document and it is not a summary of the document. It is the clause. The question the software has to answer is narrower and harder than "what does this say": it is "which 3 paragraphs here would change this person's mind, and where are they".

Why clause-level attention is the interesting part

Working at clause level forces a set of decisions that a summary lets you avoid.

• Ranking is the product. Highlighting everything is the same as highlighting nothing. If the tool marks 20 passages, the reader is back where they started, only now with a coloured document. The value is entirely in what it declines to mark.
• Critical is relative to the reader. An auto-renewal term is background noise in one agreement and the whole point of another. What counts as critical depends on what the person stands to lose, which means the ranking cannot be a fixed list of clause types.
• The output has to be checkable. Pointing at a location in the original text lets the reader verify the tool in one glance. A paraphrase does not: to check a paraphrase, you have to read the whole document, which is the task you were trying to avoid.

What it deliberately does not do

It does not give legal advice, and the design should make that obvious rather than disclaim it in small print. The distinction I hold to is between direction and interpretation. Telling someone where to look is direction, and a tool can do it honestly. Telling someone what to do about what they find is interpretation, and that belongs to a lawyer.

Keeping to that line is a product constraint with teeth, because the more useful the tool feels, the more the interface will be read as advice whatever the wording says. The work is in the phrasing, the ordering and the restraint, not in a checkbox.

Why this is where I put the team

Highlighting the clauses that matter is a narrow feature in a category full of broad ones. That is the argument for it. Document scanning is solved; a scanner that answers "what in here should I not miss" is doing something a filing tool does not. The team of 10 works on that question, and the company is structured so it can keep working on it.

Back to the main page. The two published interviews cover this product thinking in conversation form.