Natural language processing
Language processing for the text you already have: sorted, extracted, summarized
Natural language processing, or NLP, is software that works with written language. It sorts emails and tickets by type and urgency, pulls out names, dates, amounts and order numbers, and writes a short summary. We build it for the text your business already collects, and the results land in your CRM, your help desk or a weekly report.
- Prices published on the site
- First working version in two weeks
- Support month to month, cancel any time
Most businesses hold text that nobody has time to read
A shared inbox, a year of support tickets, call transcripts, contracts, notes in the CRM. Language processing turns that text into fields: this is a refund request, it is urgent, the order is 48213, the amount is $312.50. Once the text is fields, it can be routed, counted and reported on.
Two tools do the work. A large language model such as Claude or Gemini handles varied, long or ambiguous text. At high volume with fixed categories, a small trained classifier can cost far less per message, so we test both on your messages and pick by volume and accuracy.
From a pile of messages to fields, queues and a weekly report
Comes in
The system
Language processing
- Splits and cleans the text
- Sorts it by type and urgency
- Pulls out names, dates, amounts and order numbers
- Removes personal data where it must not be stored
- Writes a short summary
Lands in
Messages the system is unsure about go to a person with the reason attached. We score the whole chain on a test set of your past messages before it goes live.
Six jobs language processing does well
Ticket sorting
Assigns each new ticket one of your categories, such as billing, delivery or warranty, sets a priority from the wording and fills the order number field.
Field extraction
Pulls names, dates, amounts and reference numbers from emails, notes and contracts into columns, such as the renewal date and notice period of a service contract, for a person to check.
Summaries
A three-line summary of a 40-message email thread or a 20-minute call transcript, placed at the top of the CRM record.
Personal data removal
Masks Social Security numbers, dates of birth and similar personal data before the text is stored or sent to an outside model. Masking is tested on your samples, because automated masking can miss an item.
A weekly themes report
Groups a week of tickets and emails by topic and lists what customers write about most, with example messages.
A small trained classifier
A model trained on your labeled messages that sorts one fixed set of categories at high volume, and costs far less per message than a large model. We test it against the large model on the same messages.
When language processing pays off, and when rules do the job
Worth building when
- Hundreds of messages or documents a day are sorted or read by hand
- The same few fields are retyped into a CRM or spreadsheet from text
- Nobody reads the old tickets, so nobody knows what customers keep asking about
- You can give us 50 to 200 past messages with the outcome your team chose for each
Use what you have when
- You get about 30 emails a day: rules and saved replies in your help desk usually do the job
- The categories are few and the wording hardly varies, so a keyword rule is cheaper and easier to explain
- A form with dropdown fields would collect the same data cleanly at the source
- The text is a scan or a photo, which starts with computer vision
Priced by the kind of text and the systems it reaches
One workflow running end to end, or one internal tool.
Several systems made to work as one: integrations, roles, reports.
One kind of text into one system, for example support tickets sorted and filled into your help desk, usually lands in one process. Several text sources, a trained classifier, redaction rules or a report across systems land in connected systems. Model usage is billed by the provider for the text processed, and we estimate the monthly bill from your real volume in the audit.
Care $300 to $500 a month. Active $750 to $1,500 a month. AI model usage, GPU time and hardware are paid to the provider, on accounts in your name.
Questions about language processing
Is NLP the same as ChatGPT?
No. ChatGPT is one product built on a large language model, and large language models are one tool inside NLP. The field is older and includes classifiers, entity extraction and summarizers, some small enough to run on an ordinary server. We use a large model such as Claude or Gemini where the text varies, and a small trained model where the volume is high and the categories are fixed.
How accurate is it?
That depends on your text, so we measure it. We take 50 to 200 of your past messages with the category your team chose, run the system and show you the score and the cases it got wrong. Messages it is unsure about go to a person. LLM evaluation describes the method.
What happens to personal data in the text?
Where it must not be stored, we mask it before the text is saved or sent to an outside model: Social Security numbers, dates of birth and similar fields. We test the masking on your samples, because no automated masking catches everything. The system can live on your own server. If it calls an outside AI model, we tell you which one and what it sees.
Does it work in other languages?
Claude and Gemini read Spanish and most major languages, and we test on your actual messages in each one you use. The same goes for a small trained classifier, and some languages may need their own labeled examples.
Do we need labeled data of our own?
For a large model, usually a handful: 20 to 50 examples go into the instructions. A small trained classifier needs a few hundred labeled messages, and your team's past tickets with their tags are the labels.
What does it cost to run per message?
With a large model the provider bills by the amount of text processed (tokens), and the total grows with message length and volume. A small trained classifier costs far less per message at high volume, which is why we test both. We estimate the monthly bill from your real message count in the audit.
Related services
- Semantic search Find the right document by meaning and keyword
- Computer vision Read photos and scanned documents
- AI agents Agents that finish a task, with your sign-off
- AI customer support Chat that answers and hands off to a person
- LLM evaluation Test which model does your job best
- AI development Every AI service we build, on one page
Tell us what's slowing you down.
You get a free written review by email within two business days.