Check every
call
against your own rules.
Write the questions once. Prismlet answers them on every call, letter, message or time entry, only from the options you wrote, and says how sure it is.
The answers in the drawing are real, stored from one run. Each percentage is how sure it is on this text, not a measured accuracy.
Write your questions once.
Every call.
Same questions.
Same way.
Not the 2% a QA team samples: every call, asked the same way. A person reviews the flags.
When none of your options fit, it says so.
The last call gets Other for how it ended, instead of the nearest option.
Find your job.
- Check every call against your rules, not 2% of them
- Fix time entries before the bill goes out
- Find the complaints nobody logs
- Honour opt-outs however the customer words them
- Catch the messages that change what collections may do next
- Sort the claims inbox before a handler opens it
- Fix vague goods descriptions before the customs filing
- Find the damp, mould and hazards in every repair report
- Catch replies that break your policy before they go
- Check every complaint letter before it goes out
- An inbox that routes itself
- Browse all 128 prisms in the library
See it on an example.
Market (US or UK): US
Channel (call, chat or email): call
Conversation with speaker labels
- What does the customer want to do about their service or contract?Clear request to cancel100%
- What main reason does the customer give for wanting to cancel?Price or a price rise100%
- Does the agent make a retention offer to keep the customer?Yes99%
- Does the agent keep offering or arguing for staying after the customer declines an offer or repeats that they want to leave?Yes96%
- Does the agent put a hurdle in the way of the customer's request to cancel?No71%
- How does the conversation end for the customer's request to cancel?Cancellation done or booked98%
- Does the agent explain what the customer will pay on leaving?Charges explained100%
- Does the customer express dissatisfaction with how their request to cancel is handled?Yes96%
- Does the agent log a complaint about the handling and give the customer the complaint route the rules require?No86%
These are real answers stored from one run on this example.
Matter type: Commercial litigation
Time entry narrative
- Is the narrative a description of work actually done?Yes92%
- Does the entry group two or more distinct tasks under one time charge?Yes95%
- Is any task in the narrative too vague for the client to tell what was done and why?No68%
- Does the entry include clerical or administrative work the client does not pay for?Yes98%
- Does the entry include a meeting, call or email among the firm's own people?Yes97%
- Does the entry include work the guidelines exclude other than clerical work and internal meetings?No88%
- Does the narrative contain anything unfit for the client's lawyers to read?No92%
- Which UTBMS litigation task code fits the main work in the entry?L330 Depositions100%
- Which UTBMS activity code fits the main work in the entry?A103 Draft or revise91%
- What should billing ask for before this entry goes on the bill?Send to billing review for excluded work85%
These are real answers stored from one run on this example.
Channel (call transcript, chat, email, letter, social): chat
The whole contact
- Does the customer express dissatisfaction with something the firm did or failed to do?Yes98%
- Under the complaint definition in the context, is this contact a complaint?Yes97%
- What harm does the customer say they have suffered or may suffer?Money lost or charged100%
- Which of the firm's products is the contact about?Payments and transfers100%
- What is the customer mainly dissatisfied about?Payment or transaction error100%
- Did the agent handle the customer's dissatisfaction as a complaint?Not handled as a complaint100%
- How does the customer's issue stand at the end of the contact?Resolved and accepted100%
- Does the customer ask for money back, compensation or a goodwill payment?Yes99%
- Does the customer say they have gone, or will go, to an ombudsman, regulator, lawyer, the press or social media?No91%
- Is the dissatisfaction about a payment service?Yes98%
- Is there enough in the text to tell whether the customer is dissatisfied, and with what?Yes99%
These are real answers stored from one run on this example.
Channel: SMS
Our message they replied to: Harbor & Pine: 48 hours only! 30% off all outdoor furniture, today and tomorrow. Shop now: harborpine.example/s30 Reply STOP to end texts.
Reply
- Does the reply ask us to stop sending some or all messages?Yes94%
- If the reply asks us to stop, which messages does it ask us to stop?Marketing only93%
- Is it unclear from the reply whether the sender wants messages to stop?No79%
- What is the reply mainly about?Opt-out request94%
- Does the reply say we reached the wrong person?No96%
- Does the reply say the sender never signed up or never gave permission to be contacted?No96%
- Does the reply say the sender has asked us to stop before?Yes97%
- Does the reply mention a lawyer, a lawsuit, a regulator or a spam report?No92%
- Apart from any request to stop, does the reply raise a question or problem a person should answer?Yes98%
- Is the reply abusive or threatening toward the company or its staff?No93%
These are real answers stored from one run on this example.
Who reported it: Resident
Channel: Repairs portal
Repair report
- Does the text report a problem with the condition of the home?Yes99%
- Which hazard does the repair report describe?Damp and mould98%
- How serious is the hazard the repair report describes, by the triage guide?Potential significant hazard87%
- Does the repair report describe a second, separate hazard?Yes98%
- Does the report say someone at higher risk lives in the home?Yes99%
- Does the report say someone's health is already being affected?No88%
- Does the report say the problem was reported before or came back after a repair?Yes98%
- How far has the damp, mould or water damage spread?Widespread95%
- Which trade should attend first for the main problem in the repair report?Damp and mould specialist90%
- Does the report say where the problem is and what is wrong, clearly enough to raise a job?Yes97%
- Does the report express dissatisfaction with the landlord's service, actions or lack of action?Yes98%
These are real answers stored from one run on this example.
Email subject: Re: Claim HH-2291847 kitchen leak - STILL WAITING
Attachment text: AquaFix Plumbing Ltd INVOICE 4471 Date: 28/08/2026 Customer: Mrs S Kemp, 22 Larch Close Trace and access to leak under kitchen sink. Replace failed compression joint on cold feed. Make safe. Labour 3 hrs £165.00 Parts £22.40 Call out £65.00 Subtotal £252.40 VAT £50.48 TOTAL DUE £302.88 ENGINEER REPORT Joint failed due to age. Water has tracked under base units into the subfloor. Units swollen and delaminating, kickboards lifted. Recommend strip out of base units and drying of subfloor before refit.
Email or letter text
- What is this claims item, judged by what the sender needs from us?Progress chaser100%
- Does the sender express dissatisfaction with our service, decision or conduct?Yes99%
- Does a solicitor or other legal representative act for the sender or claimant?No93%
- Does the sender say they are thinking of instructing a solicitor or taking legal action?No91%
- Does the item demand payment or action by a stated date or within a stated period?No67%
- Does the sender mention circumstances that may mean the customer needs extra support?Yes97%
- Does the sender ask us to pay or reimburse something, or say a payment has not arrived?Yes96%
- Does the sender ask the claim handler to decide or authorise something?Unsure43% yes
- How urgent is it to act on this claims item?Urgent100%
- Is the item complete and readable enough to judge what it is?Yes98%
These are real answers stored from one run on this example.
You can't paste your own text here. Real text goes in once we have agreed how it is handled.
Test it on your own data.
- Export a sample your team has already labelled.
- Run it in the app as a CSV, TSV or JSON file of up to 10,000 rows, and get the answered file back.
- Add your team's answers as columns, and the app reports how often Prismlet agreed, question by question and at each cutoff, split by a column such as language, with every miss listed. Your team's answers stay in your browser.
A 39-second silent recording of the app. A CSV of 12 save-desk calls with the team's answers is dropped into Run a whole file. The reference, context and language columns are chosen, and the 12 rows are answered. The view then zooms in on the evaluation report, which shows agreement for each question at every cutoff, the figures split by English and Spanish, and the one miss with how sure the answer was.
What happens to your text.
True today
- Every answer is one of your options, never free text. A Choice answer can also be Other or Not enough information.
- Every result is stored with the fields you sent, every probability and the exact prism version it ran on.
- Changing a cutoff relabels stored results without sending the text to the model again.
- Every query is scoped to your account, and a test checks that one account cannot read another's data.
- API keys are stored hashed. People sign in with a one-time link, not a password.
Not decided yet: model provider, hosting region, retention, training terms, encryption, certifications, sub-processors, a data processing agreement. We answer each in writing before you send real data.
The security pageQuestions, answered.
from a fixed list of options, naturallyCan it make up an answer?
No. Every answer is one of the options the question allows, never free text. A Choice question also allows Other, when none of your options fit, and Not enough information, when the text doesn't say enough. A Scale question always picks one of your levels. It can still pick the wrong option, which is why every answer says how sure it is.
Is a percentage how often it is right?
No. It is how sure Prismlet is on this text. How often it is right on your text is something you measure: run a file with your team's answers as columns, and the app reports how often Prismlet agreed.
What can I send?
Any text: call transcripts, letters, messages, time entries, notes. Send one text per request from your own system, with up to about 100,000 characters across its fields, or run a CSV, TSV or JSON file of up to 10,000 rows from the app. A prism holds up to 100 questions.
Which languages?
The library examples are in English. We have not measured accuracy per language, so measure yours: run a file with your team's answers and split the result by a language column.
Who decides what happens?
A person, or your own rules. Prismlet answers your questions and says how sure it is; it does not act on a person, an account or a record.
Do I need a developer?
Not to start. Write your questions and try them in the app. A developer connects your tools once you're happy with the answers.