TL;DR
Named entity recognition (NER) finds names, organizations, places, dates, amounts and custom business values in text, then labels each one so software can use it. Of the five NER approaches, rules win for fixed-format IDs such as policy numbers, and statistical CRF models mostly survive in legacy systems. A transformer, bought as a managed cloud API for standard entities or fine-tuned in-house, pays off for domain entities at high volume. An LLM or a zero-shot model such as GLiNER suits new or shifting entity types, as long as a person checks low-confidence output. Most enterprise pipelines end up as a hybrid, with a different method for each label. Choose by per-label F1 on unseen documents and by cost per accepted record, never by list price alone.
Key Takeaways
Named entity recognition finds and types spans of text, from people, organizations and dates to custom business entities such as policy numbers.
A production system is a pipeline, so OCR, tagging, normalization, linking and human review matter as much as the model.
Pick the approach per entity type: rules for fixed formats, managed APIs for standard entities, fine-tuned models for domain entities at volume and LLMs for shifting schemas.
A 2023 study found a visible gap between GPT-4 and specialized extraction methods, while compact zero-shot models such as GLiNER give a cheaper flexible option.
Measure entity-level precision, recall and F1 per label on a held-out gold set, because an overall score hides weak labels.
LLM token spend is small next to human review time, so compare approaches by cost per accepted record.
Watch on YouTube
Document Intelligence Explained: OCR vs IDP vs LLMs, RAG, Tools, ROI & Security
A Kanerika video on how OCR, document processing and LLMs fit together, the same stack that named entity recognition sits inside.
A Claims Inbox That Never Empties
Picture a claims team at a mid-sized insurer on a Monday morning. Hundreds of emails arrived overnight, each carrying adjuster notes, repair estimates or hospital discharge summaries.
Before anyone can decide on a claim, someone has to find the policy number, the claimant, the provider, the loss date and the amount. That someone is usually a person copying values into the claims system one field at a time, so the backlog grows after every storm.
Named entity recognition is the AI task built to take over that copying. It finds each value in the text, labels its type and passes it on with a confidence score. People then review only the values the model is unsure about. A probability on each answer helps the same way with whole-text decisions such as ticket triage and moderation, which Kanerika’s guide to Jev AI use cases covers.
What Is Named Entity Recognition?
Named entity recognition is a natural language processing task that finds spans of text referring to real-world things and gives each one a type. Take the sentence “Acme Insurance paid $4,200 on March 3.” A model tags “Acme Insurance” as an organization (ORG), “March 3” as a DATE and “$4,200” as MONEY.
The work has two parts. Detection decides where an entity starts and ends, and classification then decides what kind of entity it is. That second step depends on context, because “Amazon shipped the order” names a company while “the Amazon flooded its banks” names a river.
The business case comes from volume. MIT Sloan cites analyst estimates that 80% to 90% of data is unstructured data , so most of it sits in emails, contracts, notes and PDFs. NER then turns that text into fields that a database, a search index or a workflow can use.
Standard Entities vs Custom Business Entities
Off-the-shelf models recognize a fixed set of general types. The English pipelines in spaCy , for example, ship 18 labels trained on OntoNotes 5, from PERSON and ORG to MONEY, DATE and WORK_OF_ART.
Still, most enterprise value sits in entities no general model knows. A policy number or a contract clause type means nothing to a model trained on news text. So those custom entities need patterns, labeled examples or a schema you define yourself.
The table pairs common labels with an illustrative example, as well as the business job each one does. The first six come with standard models, and the last five are custom entities a team defines.
Label Example (illustrative) Enterprise use
PERSON Jane Ortiz Claimant, signatory or customer name ORG Acme Insurance Counterparty such as a vendor or employer GPE / LOC Ohio, Lake Erie Jurisdiction and risk location DATE March 3, 2026 Loss date, contract start or due date MONEY $4,200 Amounts such as claims, invoice totals and fees PERCENT 12% Rates such as interest or discount Custom: POLICY_NUMBER POL-88213 Match a claim to its coverage before payment Custom: CLAIM_ID CLM-2026-00417 Route claims so that none are duplicated Custom: DRUG metformin Pharmacovigilance and clinical coding Custom: SKU SKU-44-910 Order and returns processing Custom: CLAUSE_TYPE Limitation of liability Contract risk review before signing
NER vs Entity Extraction, Entity Linking, Classification, Keyword Extraction, POS Tagging and OCR
NER is often confused with neighboring tasks because they run in the same pipelines. The quickest test is what each one outputs.
Entity extraction is the broader goal, while NER is one method inside it. A full extraction job also pulls relations, such as which party pays which fee, and attributes, such as a currency or a payment term.
The other tasks sit before or after NER. OCR produces text from an image, and NER then reads that text. Entity linking then connects a mention that NER found to a record, such as a customer ID in your CRM. Text classification labels a whole document, while NER labels spans inside it.
Task Output Question it answers How it pairs with NER
Entity extraction Entities plus their relations and attributes Who pays what, to whom and when? NER is one method inside it OCR Text from scans and images What does this page say? Runs first, so its errors flow into NER Entity linking A link from a mention to a known record Which Jane Ortiz is this? Runs after NER so that entities match master data Text classification One label per document or sentence Is this a claim, a complaint or an invoice? Routes each document before NER runs Keyword extraction Important terms without fixed types What is this text about? Gives topics, while NER gives typed entities POS tagging A grammatical tag per word Is this word a noun or a verb? Often used as an input feature in statistical NER NLP (the field) Many tasks such as translation How can software process language? NER is one task inside it
Keep these boundaries in mind when you scope a project. A request for “NER” often means entity extraction plus linking, so it changes what you build and test.
Kanerika Service
AI and Machine Learning Services
Kanerika designs and builds AI solutions for enterprise documents and data, from entity extraction pipelines to model monitoring in production.
Explore AI Services How Named Entity Recognition Works in a Production Pipeline
A production NER system is a pipeline, and the model that tags entities is one step among seven. Many failures happen in the steps around the model, so each step needs an owner and a test.
Consider one scanned claim form, for example, as it moves through the system. It is converted to text and split into tokens, then tagged, cleaned up and matched to existing records before anything reaches the claims platform.
NER is one technique inside broader data extraction work, and the full document workflow is what intelligent document processing packages. The steps below show how the pieces fit, from intake until review.
Ingest and OCR. Pull documents from email, shared drives or APIs, then run OCR on scans and images.Normalize and tokenize. Clean the encoding and fix broken lines so the text splits into tokens the model can read.Tag spans. The model finds where each entity starts and ends.Classify. Each span gets a type, such as ORG, DATE or POLICY_NUMBER, and a confidence score.Post-process. Convert “March 3rd” and “3/3/26” to one date format, then parse amounts and remove duplicates.Link. Match entities to master records, such as the policy table or the vendor list.Validate and route. Apply business rules first, then send low-confidence items to a human review queue.
The graphic below groups these seven steps into five stages, and the last stage catches what the model cannot. A rule such as “the loss date cannot fall after the claim date” flags errors that no confidence score will show. That matters because the model was confident about the wrong value.
Tagging Schemes: BIO, IOB and BILOU
Models learn entity boundaries from tags attached to every token. The tagging scheme decides how those boundaries are written down, so it shapes how your training data gets labeled.
BIO tags the Beginning of an entity, tokens Inside it and tokens Outside any entity. IOB1 is an older variant that uses B only when two entities of the same type sit side by side. BILOU also adds Last and Unit tags for sharper boundary signals. Here is one business sentence tagged all three ways.
Token BIO IOB1 BILOU
Acme B-ORG I-ORG B-ORG
Insurance I-ORG I-ORG L-ORG
paid O O O
$4,200 B-MONEY I-MONEY U-MONEY
to O O O
Jane B-PERSON I-PERSON B-PERSON
Ortiz I-PERSON I-PERSON L-PERSON
on O O O
March B-DATE I-DATE B-DATE
3 I-DATE I-DATE L-DATEWhichever scheme you choose, keep it consistent across annotators and tools. Otherwise, mixed schemes in one training set produce boundary errors that look like model problems.
Worked Example: One Contract Clause to a Validated JSON Record
Here is how a single clause moves through the pipeline. The parties, values and confidence scores below are illustrative, so treat them as a pattern and not a benchmark.
Input clause:
This Services Agreement is entered into on April 1, 2026 between
Northwind Logistics Ltd and Contoso Health Inc. The total fee is
USD 120,000, payable within 45 days of invoice.After tagging, normalization and linking, the pipeline emits this record.
{
"document_type": "services_agreement",
"entities": [
{"text": "April 1, 2026", "label": "EFFECTIVE_DATE",
"normalized": "2026-04-01", "confidence": 0.97},
{"text": "Northwind Logistics Ltd", "label": "ORG",
"role": "provider", "vendor_id": "V-1042", "confidence": 0.95},
{"text": "Contoso Health Inc", "label": "ORG",
"role": "client", "confidence": 0.94},
{"text": "USD 120,000", "label": "MONEY",
"normalized": {"amount": 120000, "currency": "USD"}, "confidence": 0.98},
{"text": "45 days", "label": "PAYMENT_TERM",
"normalized": "P45D", "confidence": 0.71}
],
"validation": {
"vendor_matched": true,
"review_threshold": 0.80,
"route_to_review": ["PAYMENT_TERM"]
}
}Three things happened after tagging. Dates and amounts were normalized into formats a finance system accepts, and then the provider was linked to its vendor master record. The payment term scored 0.71, below this workflow’s illustrative 0.80 threshold, so it went to a reviewer instead of straight through.
NER Approaches Compared: Rules, Statistical Models, Transformers, LLMs and Hybrid
There are five broad ways to build an NER system, and each trades accuracy, cost and upkeep differently. None wins everywhere, so the useful question is which one fits each entity type.
Follow one value through each approach, since the claims inbox makes a good test. A regular expression catches “POL-88213” every time, yet it misses “policy no. 88213” typed by an adjuster.
A conditional random field (CRF), the classic statistical model, copes with more variation because it learns from labeled data and features such as capitalization. Benchmarks like the CoNLL-2003 shared task , which covered English and German, shaped how these systems are still scored.
Fine-tuned transformers go further, since they read meaning from context. Hugging Face calls NER “one of the most common token classification tasks,” and its guide shows how to fine-tune such a model. Teams that want smaller training runs also look at parameter-efficient fine-tuning .
An LLM handles the case the others struggle with, such as a brand-new CLAUSE_TYPE label described in a prompt with no labeled examples. The trade-off, though, is a higher cost per document and less predictable span boundaries. Hybrids keep rules for fixed patterns like the policy number and use a model for everything else.
The Five Approaches Side by Side
The table compares the five on labeled data, accuracy, latency, cost and upkeep. Read it per entity type, because one pipeline often uses three of these rows at once.
Approach Labeled data needed Accuracy profile Latency Main cost driver Maintenance Best fit
Rules and regex None High on fixed formats but poor on variation Very low Engineer time Update when a format changes IDs and codes in known layouts Statistical (CRF) Moderate, plus feature work Fair on standard types Low Feature engineering Retrain and rework features Legacy systems already in place Fine-tuned transformer Labeled set per entity type Strong in-domain Low to moderate Labeling and compute Retrain when documents drift Domain entities at high volume LLM prompting None or a few examples Flexible but uneven on boundaries Higher per call Tokens per document Prompt and schema updates New types at low volume Hybrid Varies by component Strongest when each part is tuned Mixed Integration work Two components to own Mixed document workflows
LLMs for Entity Extraction in Late 2026
Large language models changed what is possible when entity types are new or keep shifting. You describe the entities in a prompt or a JSON schema, and the model returns them with no labeled training set. So LLMs are a fast way to prototype a schema or cover a long tail of rare document types.
The accuracy evidence is more mixed. A 2023 evaluation of GPT-4 , revised in September 2024, found “a visible performance gap” between that model and state-of-the-art information extraction methods. Newer models may narrow that gap, so test them on your own labels.
An LLM can also return a value that never appears in the source text. Those invented values are, in effect, a form of AI hallucinations . A span-based model cannot produce them by design, because it can only tag text that is already there.
Compact models, however, are narrowing the flexibility gap. GLiNER is a bidirectional transformer that extracts any entity type you name at run time. Its 2023 paper reports that it beat ChatGPT and fine-tuned LLMs on zero-shot NER benchmarks, and the GLiNER repository is released under the Apache-2.0 license.
LLM API Prices for Entity Extraction
If you do call a hosted model, the price spread is wide. The table lists published API prices as of November 3, 2026 for models teams test on extraction work, plus two open-weight options you can host yourself.
Model Price per 1M tokens (input / output) Context window Note
GPT-6.1 Sol $2 / $10 1,050,000 tokens Published September 29, 2026; 50% lower when run through Batch or Flex GPT-6.1 Terra $1 / $5 1,050,000 tokens 50% lower when run through Batch or Flex GPT-6 Luna $0.10 / $0.50 1.05M tokens Low-cost OpenAI tier Claude Opus 5.5 $4 / $20 1M tokens Announced September 22, 2026 Claude Sonnet 5.5 $2 / $10 1M tokens Mid-tier Claude Haiku 5.5 $0.10 / $0.50 up to a 100K-token prompt 1M tokens Low-cost tier Gemini 4 Argon $2 / $10 through November 30, 2026; $4 / $20 from December 1, 2026 Up to 1M tokens Price doubles December 1 Gemini 3.8 Flash $0.75 / $3.75 through December 31, 2026, then $1.50 / $7.50 See vendor page Price rises January 1, 2027 Qwen3.8-27B Self-hosted, your compute See model card Open weights, Apache 2.0 DeepSeek V4.1-Flash-1017 Self-hosted, your compute See model card Open weights, MIT license on Hugging Face
Prices change, so check the vendor page before you budget. Extraction prompts are short next to these context windows, which is why the low-cost tiers deserve a test first.
A common pattern combines the two worlds. Teams prototype with an LLM, keep the outputs reviewers confirm as labels, then move high-volume entity types to a cheaper fine-tuned or zero-shot model. For the wider trade-offs, Kanerika’s guides to enterprise LLMs as well as small language models go deeper.
When to Self-Host the Extraction Model
Run extraction on a model you host when the documents cannot leave your environment. Patterns such as virtual large language models and private LLMs keep prompts, outputs and logs inside your own boundary. That control matters most when the text holds PHI or contract terms.
Talk to Kanerika
Planning an NER Pipeline for Your Documents?
Kanerika helps teams define the entity schema, build a labeled gold set and score candidate approaches per label before anything reaches production.
Talk to an AI Expert → NER Tools and Platforms Worth Evaluating
The tool market splits into open-source libraries you host and managed cloud APIs you call. So your choice usually follows where your data already lives and how much control you need over it.
Open-source libraries give full control and no per-call fees, but your team owns hosting, scaling and retraining. Managed APIs remove that work while adding per-unit pricing, regional availability and service limits, so check each limit against the vendor documentation before you commit.
Tool Type What it gives you Check before buying
spaCy Open-source Python library Fast pipelines and English models with 18 OntoNotes labels; v3.8.18 is the newest release on GitHub as of November 3, 2026 Accuracy on your domain text and custom training effort Hugging Face Transformers Open-source library and model hub Token classification fine-tuning for models such as BERT GPU needs and who owns MLOps GLiNER Open-source zero-shot model (Apache-2.0) Extracts entity types you name at run time Accuracy on your labels before rollout Azure Language custom NER Managed, a feature of Azure Language in Foundry Tools Custom entity models built through Microsoft FoundryRegion, pricing tier and data residency Amazon Comprehend Managed AWS API Built-in entities plus custom recognizers with 1 to 25 entity types per model Billing in 100-character units , with a 300-character minimum per requestGoogle Cloud Natural Language Managed API Entity Analysis for person, organization, location, events, products and media Billing per 1,000-character unit , with the first 5,000 units a month freeLLM APIs with a JSON schema Managed model APIs Flexible extraction into a schema you define Token cost, invented values and data handling terms
Free Tiers and Service Changes to Check
The free tiers also differ. Amazon Comprehend gives 50,000 units, or 5 million characters, per API per month for 12 months from the first request. Google’s free allowance, by contrast, resets every month.
One AWS detail is worth checking before you plan around Comprehend. AWS has closed topic modeling, event detection and prompt safety classification to new customers , and it says other Comprehend features are not affected.
Whichever tool you shortlist, read its data handling terms before its accuracy claims. Kanerika’s LLM security guide covers the controls to ask about when documents leave your network.
Datasheet
KlarityIQ: Your Document Intelligence Agent
Specs for KlarityIQ, Kanerika’s document intelligence agent, for teams comparing libraries and cloud APIs with an agent built for enterprise documents.
View the Datasheet → Enterprise NER Use Cases by Industry
Named entity recognition pays off wherever people read documents to copy values into systems. The entities differ by industry, but the pattern holds everywhere: find the values, check them and send them on.
Take one case in depth, KYC screening at a bank. During onboarding, NER pulls names, organizations and addresses from account files and news. Those entities are then matched against sanctions and watch lists, a core step in AI fraud detection .
Finding the names, though, is the easy part. A customer named Jordan can hit a watch-list entry with the same name. An address in the country of Jordan can also trigger an alert. Each weak hit then lands in an analyst’s queue, which is why the banking KPI is false positives and not raw recall.
Entity linking is what shrinks that queue. Once each name is tied to the customer record, the date of birth, address and nationality come with it. So the comparison runs on several fields instead of one string. Many weak hits then clear without a person.
The other industries also follow the same find, check and send pattern. In accounts payable, NER pulls vendor names, invoice numbers and totals as the front end of invoice workflow automation . Many of these workflows start with scanned files, so this is where document intelligence comes in.
NER Use Cases at a Glance
The table pairs each industry with the entities it needs, as well as a KPI that shows whether extraction is working. Pick the KPI before the pilot starts, since it becomes the success test.
Industry Entities Workflow KPI to track
Insurance Policy number, claimant, loss date, amount Claims intake, the first step for AI agents in insurance claims processing Straight-through processing rate Banking Person, organization, address, country KYC and AML screening False positives on watch-list matches Healthcare Drug, condition, patient identifiers Clinical notes and PHI de-identification Recall on identifiers Legal Party, date, governing law, clause type Contract review and AI contract analysis Review hours per contract Accounts payable Vendor, invoice number, total, due date Invoice processing Touchless invoice rate Customer support Product, account ID, order number Ticket routing Misrouted ticket rate Publishing and content Person, organization, product, date Editorial checks alongside AI proofreading , so names and amounts stay consistent Name and figure errors caught before release
NER Inside RAG, Knowledge Graphs and AI Agents
Entities do more than fill database fields. In a RAG pipeline , the entities tagged in each chunk become metadata filters. That way, a question about one policy retrieves passages about that policy and not its neighbors.
A knowledge graph RAG setup goes a step further. Each linked entity becomes a node, and the relations between parties, contracts and amounts become edges a retriever can walk.
AI agents also need the same grounding. Entity linking gives an agent a stable ID to act on, such as a vendor record instead of a vendor name. A source span for each value also lets a reviewer check what the agent read before it acted.
How to Measure NER Accuracy
Accuracy for named entity recognition is measured per entity, not per word or per document. A prediction counts as correct only when the span and the type both match the gold label, which is stricter than it sounds.
Three numbers matter. Precision is the share of predicted entities that are correct, and recall is the share of real entities the model found. F1 is the harmonic mean of the two, so it drops when either one is weak.
Here is a worked example with illustrative numbers. A test set holds 200 gold entities and the model predicts 180, of which 160 match a gold entity exactly.
Precision = 160 / 180 = 88.9%
Recall = 160 / 200 = 80.0%
F1 = 2 × 0.889 × 0.800 / (0.889 + 0.800) = 84.2%
Report these scores per label and overall. A model can score 95% overall and still score 60% on policy numbers, because dates and names dominate the count. So the rare label that matters most is often the one that fails.
The open-source seqeval library computes these scores CoNLL-style, entity by entity. An LLM evaluation framework applies the same discipline when a model, not a tagger, does the extraction.
Decide early, before any comparison, whether partial matches count. Exact-span scoring marks “Acme Insurance Group” wrong when the gold label is “Acme Insurance,” while partial scoring credits the overlap. Pick one rule, write it down and use it in every comparison.
How Much Labeled Data Does NER Need?
There is no universal quality threshold, but vendors publish training floors. Amazon Comprehend requires at least 25 annotations per entity type to train a custom recognizer on plain text. Each model can hold 1 to 25 entity types.
Still, treat that number as the minimum to start training, not a target for accuracy. What matters is the count of examples per entity type, since one contract can hold ten dates and no policy numbers at all. Kanerika’s guide to LLM training covers how labeled data shapes model quality more broadly.
Build a held-out test set before training anything, and never train on it. Have two people annotate a sample of the same documents, then check how often they agree. When annotators disagree on a label, the model will too, so fix the guideline before adding data.
The podcast episode below covers the product side: how KlarityIQ, Kanerika’s document intelligence agent, turns enterprise documents into usable data.
Back to training data: the table sums up what each approach needs before the first training run.
Approach Labeled training data What to prepare
Rules None A test set so you can check precision on real documents Managed API, standard entities None A test set drawn from your own domain Managed API, custom entities Vendor minimums, such as 25 annotations per entity type for Comprehend plain text Annotation guideline and a test set Fine-tuned transformer Labeled examples per entity type Guideline, double annotation and test set LLM or zero-shot model A few examples or none A larger test set, since errors are less predictable
Multilingual and Low-Resource NER
Most NER models and benchmarks favor English. Even CoNLL-2003, the benchmark that shaped the field, covered only English and German. So a high score on it says little about Portuguese claims or Norwegian invoices.
Teams have three practical routes when the text is not English. They can fine-tune a multilingual transformer, or translate labeled data and project the labels back. They can also prompt an LLM in the source language. Whichever route you take, score each language separately on its own gold set, because one blended score hides the weak language.
NER Risks, Governance and Common Mistakes
NER errors are quiet. A wrong date crashes nothing, which is exactly why it slips into a payment, a screening decision or a report. The controls therefore have to be designed in from the start.
The NIST AI Risk Management Framework is a useful frame for those controls. NIST released it on January 26, 2023, then added a Generative AI Profile, NIST-AI-600-1, on July 26, 2024. That profile applies when an LLM does the extraction.
Personal data is the biggest governance issue. NER often runs on text full of names, addresses and account numbers, and it is also the tool used to find and redact them. Under the HIPAA Safe Harbor method , 18 identifier categories must be removed before health data counts as de-identified, starting with names.
So handle these entities under your data governance best practices , with access rules, retention limits and logging. Unstructured data governance and a clear data privacy policy then decide who may see the raw text before redaction runs.
LLM-based extraction adds its own risk, because a model can return a value that is not in the document. Requiring every extracted value to map back to a source span is a simple control, so it belongs in your wider AI governance policy.
Failure mode Control
Ambiguous entities such as “Jordan” Context-aware models and entity linking Nested or overlapping entities A schema that allows nesting, or else separate passes Multilingual text Language detection and per-language evaluation OCR errors cascading into NER OCR confidence thresholds before tagging Drift as documents change Regular per-label scoring on fresh samples Invented spans from LLMs Require a source span for every value PII or PHI exposure Redaction, access control and audit logs No confidence threshold Route low-confidence items to review No owner Name a business owner for each entity schema
Five Common NER Mistakes
Five mistakes come up often in enterprise NER projects, and each one is cheap to avoid early.
Choosing a model before writing the entity schema and annotation guideline.
Reporting one overall F1 score instead of scores per label.
Skipping a held-out test set, so every result looks good.
Sending everything straight through with no review queue.
Treating launch as the finish line, with no drift monitoring or retraining plan.
What Enterprise NER Costs
NER cost is mostly people and process, not model fees. The useful unit is therefore cost per accepted record, meaning a record that passed validation and reached the target system.
The formula is simple. Cost per accepted record equals setup, labeling, processing, review, integration and monitoring costs for a period, divided by the records accepted in that period. A model with cheap inference but a high review rate can therefore cost more than a pricier model that sends few items to people.
Cost component What to estimate
Data preparation OCR, cleaning and document splitting Labeling Hours per document times documents, plus double annotation Model setup Training runs, prompt design or API configuration Processing Units such as API calls, LLM tokens or GPU hours Human review Share of items routed to review times minutes per item Integration Connectors to source and target systems Monitoring and retraining Drift checks, new labels and release cycles
A Worked Cost Example
Here is an illustrative calculation that uses assumed figures, not vendor prices. Take 10,000 documents a month, each ending as an accepted record. Then assume a review cost of $3 per item, about four minutes at $45 an hour. The other costs are the same for both options, so they are left out.
Model A: $0.01 per document to run and 30% sent to review, so $100 + (3,000 × $3) = $9,100, or $0.91 per record.
Model B: $0.10 per document to run and 8% sent to review, so $1,000 + (800 × $3) = $3,400, or $0.34 per record.
The pricier model costs ten times more to run, yet each accepted record costs 63% less, because review time dominates the bill.
Checklist
Enterprise AI Checklist
A downloadable checklist to work through before you commit to an enterprise AI build such as an entity extraction pipeline.
Get the Checklist → What LLM Tokens Add to the Bill
Token prices also tell the same story. Assume the same 10,000 documents, each sending 3,000 input tokens and returning 500 output tokens. That comes to 30 million input and 5 million output tokens a month. At the November 3, 2026 list prices in the model table, that volume costs the following.
Claude Haiku 5.5: $3 + $2.50 = $5.50 a month.
GPT-6.1 Terra: $30 + $25 = $55 a month.
GPT-6.1 Sol, or Gemini 4 Argon through November 30, 2026: $60 + $50 = $110 a month.
Claude Opus 5.5, or Gemini 4 Argon from December 1, 2026: $120 + $100 = $220 a month.
Even the top line is less than a tenth of the $2,400 that Model B spends on 800 reviews. So compare options by cost per accepted record. Never set a price per token against a price per character unit, because they do not measure the same thing.
Decision Framework: Which NER Approach Fits Your Workload?
The right NER approach depends on the entity, the volume and the risk, so many enterprises run more than one. Start from the workload, then let it pick the method.
Fixed-format identifiers need nothing more than rules, and standard entities in a cloud estate fit a managed API. Domain entities at high volume justify a fine-tuned model, because labeling cost spreads across many documents.
After that, volume and data sensitivity are what change the pick. When entity types are new or shifting and volume is low, an LLM or a GLiNER-class zero-shot model with review fits best. Regulated personal data points to a self-hosted model plus redaction.
Scenario Pick Why
Fixed-format IDs such as policy or invoice numbers Rules and regex Precise, cheap and easy to audit A legacy CRF pipeline that already meets its per-label targets Statistical (CRF) Keep it until a label falls short, then replace that label Standard entities when data already sits in one cloud Transformer through a managed API No training needed; you pay per unit processed Domain entities at high volume Fine-tuned transformer Lowest cost per document once trained New or shifting entity types at low volume LLM or zero-shot model with review No labeled data needed before you start Regulated PHI or PII Fine-tuned or zero-shot model you host, plus redaction Data stays inside your boundary Mixed documents with several entity types Hybrid Each label gets the method that scores best on it
The Claims Inbox, Solved Label by Label
Take the claims inbox from the start of this guide. Policy and claim numbers follow a fixed format, so rules handle them. Names, dates and amounts are standard entities, but claims volume changes the math. A fine-tuned model costs less per document than paying per call, so the insurer trains one.
Medical details in attached reports carry PHI, so that model runs inside the insurer’s own environment. New document types go to an LLM with human review until enough labels exist to train on them.
How Kanerika Delivers NER for Enterprise Document Workflows
Kanerika builds NER as part of document workflows that end in a system of record. The work runs in five stages, and each one has a clear exit test before the next begins.
Define. Agree the entity schema with the business owner and build a labeled gold set.Score. Test candidate approaches per label on that gold set.Build. Set up intake, OCR, extraction, validation rules and source-span checks.Review. Route low-confidence values to people and feed confirmed values back as labels.Monitor. Track per-label F1 and drift once the workflow runs in production.
A global travel management company shows how this works in practice. Its invoices arrived as inconsistent PDF and Excel files, so Kanerika built Outlook intake into SharePoint and then OCR plus LLM-based extraction into standardized schemas, with an accuracy feedback loop.
Power Automate validation flagged exceptions for targeted review, while structured JSON flowed to the financial system with FLIP supporting the pipeline. The solution runs across South Africa, Norway and Brazil. Manual invoice work fell by up to 80%, and the team handled up to 4x the volume without a matching rise in cost.
Processing cycles also ran 50% to 60% faster, and extraction accuracy reached 90% to 95% through OCR and LLM models. Every invoice is now traceable across its full lifecycle.
Case Study
80% Less Manual Work in Travel Invoice Processing
A global travel management company moved invoice intake to OCR and LLM-based extraction with validation and targeted review, cutting manual work by 80%, handling 4x the volume at the same cost and reaching 90-95% extraction accuracy.
Read the Case Study → Kanerika’s AI services team delivers builds like this one, and KlarityIQ is its smart document agent that understands business files at scale. For any team starting out, the advice is simple. Start with the two or three entity types that drive a decision, keep a source span for every value and measure cost per accepted record.
Conclusion: What Enterprise Teams Should Do First
Named entity recognition turns unstructured text into typed fields that systems can act on. Each of the five approaches also has a home, from rules for fixed formats to LLMs for new types under review.
The first steps are the same for most teams, whatever the industry.
Pick one workflow and the two or three entity types that drive its decisions.
Write the entity schema and annotation guideline, then label a held-out gold set.
Score two or three candidate approaches per label on that set.
Route low-confidence values to review and keep a source span for every value.
Compare the finalists by cost per accepted record before you scale.
So which of the five NER approaches should you choose? Score them per entity type, and keep the one that wins on per-label F1 and cost per accepted record for each label. For most enterprise workloads, that makes the answer a hybrid, picked label by label.
Frequently Asked Questions
What is named entity recognition?
Named entity recognition is a natural language processing task that finds names of people, organizations, places, dates, amounts and other entities in text. It finds where each entity starts and ends, then assigns it a type. Enterprises use it to turn emails, contracts and reports into structured fields that databases and workflows can use directly.
What is the purpose of NER?
The purpose of NER is to make unstructured text usable by software. It pulls out the specific values a business cares about, such as customer names, policy numbers, dates and totals. Those values can then fill database fields, trigger workflows, power search or support compliance checks without anyone copying them by hand from each document.
What is the use case of NER?
Insurers use NER to pull policy numbers and amounts from claims. Banks screen names in KYC checks, and hospitals find patient identifiers in clinical notes. Legal teams extract parties and dates from contracts, and finance teams read vendor details from invoices. Support teams also tag products and account IDs in tickets so each one reaches the right queue.
What is an example of a named entity?
In the sentence Acme Insurance paid $4,200 to Jane Ortiz on March 3, there are four named entities. Acme Insurance is an organization, $4,200 is a money amount, Jane Ortiz is a person and March 3 is a date. Businesses also define custom entities of their own, such as a policy number, a claim ID or a product SKU.
What is the difference between keyword extraction and named entity recognition?
Keyword extraction finds the important terms that describe what a text is about, and those terms have no fixed type. Named entity recognition finds specific mentions and assigns each one a defined type, such as person, organization or date. Keywords support topic tagging and search, and NER entities fill structured fields that feed downstream systems.
What is the difference between NER and OCR?
OCR converts images of text, such as scanned forms or photos, into machine-readable characters. NER reads that text and identifies which parts are names, dates, amounts or other entities. The two often run together, with OCR first and NER second. Poor OCR quality causes NER errors, so scan quality checks matter in any document pipeline.
What is the difference between classification and named entity recognition?
Text classification assigns one label to a whole document or sentence, such as complaint, invoice or positive sentiment. Named entity recognition labels specific spans inside the text, such as a customer name or an order number. Many pipelines use both, with classification routing each document to the right NER model and NER extracting the fields.
Which model is best for named entity recognition?
No single model is best for every case. Fine-tuned transformers such as BERT-family models are strong for domain entities at high volume, and spaCy suits fast extraction of standard entities. GLiNER-class models handle new entity types without training data, while LLMs such as Claude Sonnet 5.5 suit shifting schemas at low volume. Test candidates per label first.
What is named entity recognition in banking?
In banking, NER extracts names, organizations, addresses, account numbers and dates from onboarding documents, payment messages and news articles. Those entities feed KYC and AML screening, where they are matched against sanctions and watch lists. Banks also use NER to review loan documents and route customer emails, which cuts manual reading and supports audit trails.
What is the difference between NER and POS tagging?
Part-of-speech tagging labels the grammatical role of every word, such as noun, verb or adjective. Named entity recognition labels only the words that form entities and gives them semantic types, such as person or date. POS tags were once a common input feature for statistical NER models, and modern transformer models learn those patterns directly.
What is the full form of NER?
NER stands for named entity recognition. It is sometimes called entity extraction, entity identification or entity chunking. The term covers the task of finding mentions of real-world things in text and assigning each one a type. In enterprise settings it also covers custom entities that a business defines, such as policy numbers, SKUs and contract clause types.
How to evaluate a NER model?
Evaluate a NER model at the entity level on a held-out test set it never saw during training. Compute precision, recall and F1 for each label as well as overall, because rare labels often score lower. Decide in advance whether partial span matches count. Tools such as seqeval calculate CoNLL-style scores for sequence labeling tasks like NER.
What is BIO tagging in named entity recognition?
BIO tagging is a way to label tokens for NER training. B tags the first token of an entity, I tags tokens inside it and O tags tokens outside any entity. In the name Jane Ortiz, Jane gets B-PERSON and Ortiz gets I-PERSON. Variants include IOB1 and BILOU, which adds tags for last tokens and single-token entities.
Can ChatGPT or other LLMs perform named entity recognition?
Yes. Models such as GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon can extract entities from a JSON schema with no labeled training data. They can also return values that never appear in the text. Check each extracted value against its source span, and send low-confidence values to a person for review before they reach a system.
Does named entity recognition work in languages other than English?
Yes, but most models and benchmarks favor English, and CoNLL-2003 covered only English and German. Teams can fine-tune a multilingual transformer, translate labeled data and project the labels back, or prompt an LLM in the source language. Score each language separately on its own gold set, because one blended score hides the weakest language.
Is named entity recognition the same as entity extraction?
Not quite. Named entity recognition is one method inside the broader job of entity extraction. NER finds and types spans such as names, dates and amounts. Entity extraction also captures relations, such as which party pays which fee, and attributes, such as a currency or a payment term. Many projects called NER need both.
How much training data does a custom NER model need?
There is no universal minimum. The amount depends on the approach, the number of entity types and how varied the documents are. Count labeled examples per entity type, not documents. Amazon Comprehend, for example, needs at least 25 annotations per entity type on plain text. Hold out a separate test set and add data where per-label scores stay weak.
How does NER work?
NER works in stages. Text is cleaned and split into tokens, and a model tags each token as the start of an entity, inside one or outside any entity. Each span then gets a type and a confidence score. Production systems add normalization, linking to master records and human review for results with low confidence.
Can named entity recognition detect personal information?
Yes. NER is widely used to find personal information such as names, addresses, phone numbers, account numbers and dates of birth. Teams then mask or redact those values. The HIPAA Safe Harbor method lists 18 identifier categories that must be removed from health data. Test recall carefully, because every missed identifier is a privacy exposure for the business.
What is the difference between NER and entity linking?
NER finds an entity mention in text and assigns it a type, such as person or organization. Entity linking goes one step further and connects that mention to a specific record, such as a customer ID or a knowledge base entry. NER can tell you that Jordan is a person, and entity linking tells you which Jordan the text means.
Is NER considered AI?
Yes. NER is a task within natural language processing, which is a branch of artificial intelligence. Modern NER systems use machine learning models, such as fine-tuned transformers or large language models, that learn patterns from labeled data. Simple rule-based extractors also exist, and many enterprise systems combine learned models with rules for fixed formats.
What is the difference between NER and NLP?
NLP is the whole field of teaching computers to process human language. NER is one task inside that field. Other NLP tasks include sentiment analysis, translation, summarization and text classification. NER focuses only on finding and typing entities in text. It often feeds other NLP or analytics steps that need those entities as structured input.
What are the types of NER?
NER systems are usually grouped by approach. Rule-based NER uses patterns and dictionaries, and statistical NER uses models such as conditional random fields. Deep learning NER uses neural networks, often fine-tuned transformers. LLM-based and zero-shot NER extract entity types named at run time. Many teams combine several of these in one hybrid pipeline.
