Where does it actually go wrong?

At the boundaries and on your own vocabulary. Ordinary categories such as people and dates are handled well; part numbers, internal codes and contract references are handled less well, and those are frequently the values the automation was built for.

Why are its errors hard to notice?

Because they land in fields. Free text invites careful reading and a populated field does not, so a wrong date passes every downstream step with the confidence that structure implies, and nothing about it looks different from a correct one.

What checking is worth building?

A confidence value, a rule that routes anything unusual to a person, or a cross-check between two extracted values that ought to agree. Each is small, each catches the quiet class of error, and none appears unless somebody specifies it.

Glossary

Named entity recognition

Pulling the people, companies, dates and amounts out of ordinary writing so they can be filed, matched or acted on.

In plain terms

Finding the useful bits inside ordinary writing and turning them into fields. An invoice becomes an amount and a date, an email becomes a company and a person, and the awkward part is deciding exactly where each of those starts and stops.

Why it matters

Because it is the unglamorous step behind a great deal of automation, and because its errors are unusually well disguised. A wrong value in a structured field looks exactly like a right one, and everything downstream treats it with the confidence a field deserves.

How it works

It converts writing into fields, which is what makes the rest possible. Nothing downstream can match, file or total anything until the amounts and names have been lifted out of the sentences they were sitting in.

The finding is largely solved and the boundaries are not. Where a company name ends, whether a title belongs to the person, which of two dates in a paragraph is the one meant: those are the errors, and they are quiet ones.

The output looks authoritative in a way the input never did. Free text is obviously something to be read carefully; a field containing a date is treated as a date, so an error that would have been caught by a reader passes unexamined.

Ambiguity in the source cannot be resolved by the extraction. A document that genuinely does not make clear which party is which cannot be made clear by pulling values out of it, and a system that produces confident fields from it has invented the clarity.

The specific and general cases differ sharply. Ordinary categories like people and dates are well handled; your own part numbers, internal codes and contract references are not, and those are usually the ones the automation actually needed.

It is worth separating from what happens next. Pulling a value out and deciding what it means are different jobs, and a system that does both at once makes it impossible to tell whether a wrong outcome came from a misread field or a bad rule applied to a correct one.

Checking is cheap if it is designed in. A confidence value, a rule that flags anything unusual for a person, or a simple cross-check between two extracted fields all catch the quiet errors, and none of them exists unless somebody asked for it.

The same error, before and after extraction

That promotion is the whole reason to care about this rather than treating it as plumbing. The value of turning writing into fields is precisely that fields can be relied on without being read, which is what makes the automation downstream possible; the cost is that the reliance is extended equally to the values that should not have earned it. Nothing about the mechanism distinguishes the two, and nothing about the interface invites anybody to check. The practical response is not to distrust the whole output, which would forfeit the benefit, but to pick one cheap check and build it in from the start: flag the unusual, route the uncertain to somebody, or cross-check two values that ought to agree with each other. Any one of them turns an invisible failure into a small queue, and a small queue is a thing an organisation knows how to handle.

Seen in the wild

Amounts and dates lifted from documents into fields nobody re-reads.
Adobe Acrobat AI Assistant
Company and contact names pulled from correspondence into a record.
Attio
Internal reference codes extracted less reliably than the ordinary categories around them.
Julius AI

Common misconceptions

People assume

The difficulty is finding the entities.

In fact

Finding is largely settled. The difficulty is the boundary: where a name ends, whether the title belongs to it, and which of several similar values in a document is the one that was meant.

People assume

A structured field is more trustworthy than the text it came from.

In fact

It is more usable and no more correct. A wrong value in a field looks identical to a right one, and everything downstream extends it the confidence a field earns, which the sentence it came from never received.

Questions

Where does it actually go wrong?: At the boundaries and on your own vocabulary. Ordinary categories such as people and dates are handled well; part numbers, internal codes and contract references are handled less well, and those are frequently the values the automation was built for.
Why are its errors hard to notice?: Because they land in fields. Free text invites careful reading and a populated field does not, so a wrong date passes every downstream step with the confidence that structure implies, and nothing about it looks different from a correct one.
What checking is worth building?: A confidence value, a rule that routes anything unusual to a person, or a cross-check between two extracted values that ought to agree. Each is small, each catches the quiet class of error, and none appears unless somebody specifies it.

Key takeaways

Finding is settled; boundaries are where it fails.
Your own codes and references are handled worse than ordinary categories.
A wrong field looks exactly like a right one, and is trusted more than prose.
Design one cheap check in, or the quiet errors go all the way through.

Tools that use this

Adobe Acrobat AI Assistant
Amounts and dates lifted into fields nobody re-reads.
Attio
Names pulled from correspondence into a record.
Julius AI
Internal codes extracted less reliably than ordinary categories.

Last checked August 2026

All glossary terms