Check a support reply against the conversation before it's sent
Reads the whole conversation and the draft reply, and says whether the draft answers every open question, contradicts what was said or asks again for details the customer already gave.
Try it on this example
Channel (email, web chat, chat app, SMS): Web chat
The conversation so far, oldest message first: Customer: Hi, I withdrew 250 EUR to my Visa ending 4471 on Monday and it's still not in my bank. Reference WD-55120. Agent: Thanks, let me check that for you. Customer: Also, why did my free spins disappear? I had 30 left from the weekend offer.
The reply the agent is about to send
- Does the draft answer what the customer asked in their latest messages?No80%
- Does the draft cover every question or request the customer still has open?Some missed100%
- Does the draft tell the customer what happens next?Clear84%
- Does the draft contradict something already stated in the conversation?Unsure41% yes
- Does the draft ask the customer for information they already gave?Yes96%
- Does the draft name a person, account, amount or case that does not fit this conversation?No84%
- Does the draft still contain template text that was never filled in?Yes98%
- How does the draft respond to the customer's situation?Acknowledges99%
- Is the draft respectful towards the customer?Yes95%
- Does the length and form of the draft suit the channel?No85%
These are real answers stored from one run on this example.
The prism behind it
Check a support reply against the conversation before it's sent
Fields
- Channel (email, web chat, chat app, SMS)
- The conversation so far, oldest message first
- The reply the agent is about to send
Context
The draft is the next message a support agent is about to send in the conversation shown. The conversation runs oldest first. It may carry sender names and times, or no speaker labels at all; the customer is the person who contacted support, and the agent writes for the company. Judge the draft as a reply to the conversation as it stands. The answers warn the agent before sending; the agent decides what goes out.
Questions
Does the draft answer what the customer asked in their latest messages? Yes / No
Compare the draft with the customer's most recent messages, read in the light of the whole conversation. Judge whether the draft engages with the request, not whether its answer is correct. A refusal counts when it gives a specific reason. Yes: The draft answers, acts on, or specifically declines what the customer last asked. No: The draft answers a different question, repeats an earlier answer that did not help, or is generic text that engages with no particular request.
Does the draft cover every question or request the customer still has open? Choice
List for yourself each separate question or request the customer has made that no earlier agent message dealt with, then check the draft against each one.
Does the draft tell the customer what happens next? Choice
Judge whether the draft leaves the customer knowing who does what next, whenever anything is left to do after it is sent.
Does the draft contradict something already stated in the conversation? Yes / No
Compare the facts in the draft with the conversation: amounts, dates, what was requested, what was done, what was promised, and what the customer said about their situation. Judge only against the conversation, not against outside knowledge. Yes: The draft states something that conflicts with a fact given earlier by the customer or by the company. No: Nothing in the draft conflicts with the conversation.
Does the draft ask the customer for information they already gave? Yes / No
Yes: The draft asks for a detail, document or explanation the customer already supplied in the conversation. No: The draft asks for nothing already supplied, or asks for nothing at all.
Does the draft name a person, account, amount or case that does not fit this conversation? Yes / No
Compare the names, account or transaction references, amounts and case numbers in the draft with the conversation. Count details that plainly belong to another customer or case, as in a reply pasted from another thread. The agent's own name and the company's details do not count. Yes: The draft contains a detail that does not match this conversation. No: Every detail in the draft fits this conversation, or the draft contains none.
Does the draft still contain template text that was never filled in? Yes / No
Look for merge tokens in braces or brackets such as {FirstName?} or [AMOUNT], notes to self, words like TODO, TBD or XXX, and filler text. Yes: At least one unfilled placeholder or piece of filler text remains. No: No unfilled placeholder or filler text remains.
How does the draft respond to the customer's situation? Scale
Judge acknowledgement of what the customer described, such as money lost, time wasted or stress, not politeness formulas and not whether the outcome is good. A routine request needs no acknowledgement; rate such a draft as neutral.
Is the draft respectful towards the customer? Yes / No
A plain or brief reply is fine. Yes: The draft is courteous throughout. No: The draft is dismissive, sarcastic, blaming or insulting.
Does the length and form of the draft suit the channel? Yes / No
Chat, chat apps and SMS call for short, direct messages that read naturally in a chat window. Email calls for a complete message with a greeting and a sign-off. Judge form, not content. Yes: The draft reads as a natural message for this channel. No: The draft is far too long or formal for a chat, or too clipped and incomplete for an email.
Lens columns
answers_latest_request, answers_latest_request_probability, open_requests_covered, open_requests_covered_probability, next_step, next_step_probability, contradicts_conversation, contradicts_conversation_probability, asks_for_known_details, asks_for_known_details_probability, wrong_customer_details, wrong_customer_details_probability, unfilled_placeholder, unfilled_placeholder_probability, acknowledgement, acknowledgement_average, respectful, respectful_probability, fits_channel, fits_channel_probability
Run it on your own text
Add this prism in the app, change any question, and test it on a file of your own.