Exactly one employer in our crawl of 1.1 million British careers pages puts “NLP” in a job title. One. That is not a field in decline — it is a field that stopped announcing itself. The work was absorbed into adverts for machine learning engineers, AI engineers and data scientists, and the name dropped out of recruitment while the problems kept arriving.

You therefore cannot source this hire by title, and no advert market exists to price the skill. You have to know what you are looking for.

The title vanished, the work did not

Consider what lands on a text team’s desk in a normal quarter. Route incoming correspondence to the right queue. Pull named parties, dates and amounts out of contracts. Redact personal data before it reaches a warehouse. Deduplicate a supplier list where one company appears eleven ways.

None of it is glamorous and all of it is constant. The people who do it reliably now call themselves something else, so write the brief around the problems: an advert naming the tasks finds candidates an advert naming the title never will.

Where a few megabytes beats a few billion parameters

The judgement worth paying for is knowing when a trained classifier of a few megabytes is right and a hosted generative model is merely expensive. Four arguments recur, and a good candidate reaches for them unprompted.

  • Cost at volume. A per-call charge is invisible on a thousand documents and ruinous on forty million. A model you host once costs the same whether you succeed or not.
  • Latency. Single-digit milliseconds inside a request path, versus hundreds and a network hop you do not control. If classification sits in front of a user, this decides the architecture.
  • Predictability. The same input gives the same output, and still does after a vendor ships an update nobody told you about. For anything audited, that beats a few accuracy points.
  • Explanation. When a customer appeals a decision, or a regulator asks how a document was classified, you can point at features and weights. “The model felt that way” is not a defence.

This is the instinct that makes a good tabular modeller, and the overlap is real enough that a strong scikit-learn engineer is often the nearest adjacent hire. A candidate who reaches for a large model in every case has one tool.

Classification, extraction and the shape of a document

Three things make up most commercial text work, and each has underestimated depth.

Classification is rarely the tidy two-class problem of a tutorial. Real taxonomies have forty overlapping labels, a long tail that appears twice a year, and cases where a document is genuinely three things at once. Ask how they handled a class with nine training examples, and what they told the business about it.

Entity extraction fails at the edges, not the middle. Names that are also common nouns, addresses split across lines, amounts in three currencies, dates written five ways, and the normalisation deciding whether two extracted strings are one entity. Extraction is the easy half; deciding what the extracted thing is takes the time.

Document structure is the part nobody interviews for and it eats most of the project. Reading order in a two-column PDF, tables spanning pages, repeated headers, scans of varying quality, and the boundary detection that decides where a section ends. A candidate who raises structure before models has shipped something.

Languages that break the naive pipeline

If your corpus is not purely English, this is where you learn how deep somebody’s knowledge goes: the naive pipeline works in English and falls over elsewhere.

Tokenisation is the fault line. Splitting on whitespace is a habit so ingrained that engineers forget it is an assumption, and it is simply wrong for scripts written without spaces between words, where segmentation is itself a modelling problem with its own error rate feeding everything downstream. Morphologically rich languages inflate vocabulary until a word-level model has almost no examples per form. Unicode normalisation decides whether two visually identical strings compare equal. Transliterated names and stray diacritics break lookup tables that tested clean.

Then calibration: one threshold tuned on English is too strict in one language and too permissive in another, and per-language thresholds need per-language labelled data nobody budgeted for. Ask which language gave them the most trouble. The answer is specific or it is invented.

A cohort that has only ever called an API

A meaningful share of people presenting as text specialists have never trained a model. They have written prompts, chained calls and shipped useful things — a real skill, and the subject of hiring an LLM engineer rather than this role — but not the skill that builds a small, cheap, explainable component.

Telling the difference is easy if you ask about the unglamorous middle of a training project. What was your baseline, and did the model beat it. How did you split the data — and if documents arrived in batches from one source, did you split by source. Who wrote the annotation guidelines, and what happened when two annotators disagreed. And what did you do when precision was acceptable and recall was not.

Somebody who has done it answers with irritation and detail, usually about labelling — the arguments a working data scientist will recognise, so one on the panel is a useful cross-check.

The money, and why no benchmark exists here

Treat any figure as indicative, and mean it: with one employer naming the title, there is no advertised market to benchmark against. As a working range, a capable mid-level engineer sits around £55,000 to £75,000, someone senior who has owned a text pipeline in production £80,000 to £110,000, and a day rate of £450 to £700 depending on whether multilingual work is in scope. Push higher where regulated explainability matters: that pool shrinks sharply.

An hour that cannot be bluffed

Skip the take-home. Bring a confusion matrix and ask which errors they would trade away, and what they would tell the business about the trade. Hand them two hundred rows of your real text and a taxonomy, and ask them to write the labelling guideline — the ambiguities they spot in twenty minutes tell you more than a coding exercise. Show them tokeniser output on a non-English sentence and ask what breaks downstream. Then ask them to argue against building a model at all, and see whether they can.