madfeed reads the conversations you already have, finds the requests inside them with the exact words behind each one, weighs them by who asked, and helps you tell those people when it ships.
Call with Priya · Northwind Logistics · sample
Priya: Thanks for making time. Overall we like it, a few things come up every week.
Priya: We rebuild this report in a spreadsheet every Monday. If I could just export it, that is two hours back.1
You: Anything else slowing the team down?
Priya: Our security team keeps asking whether you support single sign-on2.
Priya: Alerts in Slack would stop us missing escalations.3
Requests found3 of 3
Export any table to CSV
ReportingNorthwind · $40k/yr
Single sign-on (SAML)
SecurityNorthwind · $24k/yr
Slack alerts for escalations
IntegrationsNorthwind · $12k/yr
3 requests found · 3 quotes verified word for word · nothing is saved until you accept
Try it on your own conversationA sample conversation, played back. Pick another, or replay it.
Bring in feedback from wherever it already lives
Why it exists
The answers are already in your calls, tickets and chats. The work is finding them, proving them and acting on them.
One review queue, every source kept as written
Feedback is scattered across calls, tickets, chats and DMs.
Everything lands in one review queue, each item kept as an untouched source.
The loudest voice wins, and nobody can say what it is worth.
Every request is weighed by the accounts behind it, each counted once.
People who asked never hear back, so they stop asking.
When it ships, you approve a personal note to each person who asked.
What you get
Most tools ask you to trust a summary. This one shows you the sentence.
Evidence
Quotes are checked to appear word for word in the original at the stored position. A made-up quote cannot be saved.
…so that part is fine. Honestly the thing that slows us down is having no way to export the table, we copy it by hand every Monday…
Review
Suggestions wait in a review queue until a person accepts them. Nothing becomes a request, and nothing is sent, without you.
Ask
Ask in plain language: why are trials not converting, who wants SSO. You get quotes back, or a plain “there is no evidence”.
Why are trials not converting?
Five accounts say setup takes too long.
Do customers want a dark mode?
No evidence in your conversations, so no answer.
Prioritise
A request is worth the accounts behind it, summed once per account. Import values from a CSV, or read them from your CRM.
Yearly value of the accounts asking, each counted once
Close the loop
When a request ships, personal drafts are built from each person’s own words. You review and approve; only approved messages go out.
Draft to Priya · uses her own words
You told us the Monday report took two hours. Export to CSV is live today.
How it works
Five steps, and you stay in charge of every one.
Paste one, upload a call, forward email, add a chat window to your site, or let your tools post them in. Card numbers, keys and passwords are hidden before anything is stored.
The AI proposes requests, each with the exact sentence that supports it. You accept, edit, merge or dismiss. Nothing becomes a request until you say so.
Requests are weighed by the accounts that asked and what they are worth. Every quote is one click away, so nobody has to take a number on trust.
Ask questions answered only from real quotes, get cited briefs for the meeting, and share a public roadmap people can vote on.
Mark a request shipped and personal drafts appear, built from each person’s own words. You approve; only approved messages are sent.
One platform
From the first call to the release note, nothing to stitch together.
Example: type a question such as “why are trials not converting?” and get an answer with the quotes behind it, or a plain statement that there is no evidence.
Fig 1.1 · Ask
Plain-language questions, answered only from your stored quotes.
00:42 Priya “We rebuild this every Monday. If I could just export it…”
Fig 1.2 · Calls
Timed transcripts with speakers. Highlight and tag, or take the suggested highlights.
+48NPS, sample
24 free-text answers went to the review queue.
Fig 1.3 · Surveys
NPS, CSAT, choice and free text, with branching. Free text goes to review.
Can I get this out as a file?
Not yet, I will pass it on.
Private note: third time this week.
Send to review queue
Fig 1.4 · Support desk
Website chat and email, private notes, hand-over, saved replies.
Planned
Slack alerts▲ 128
Dark mode▲ 41
In progress
Single sign-on▲ 212
Shipped
Export to CSV▲ 305
Fig 1.5 · Roadmap
Release notes and a roadmap for signed-in visitors.
See the shape of it
Trends, sources and hot spots, drawn from the same quotes you can click through to.
Mentions per week for three requests. Spot the one that is taking off.
| Request | W1 | W2 | W3 | W4 | W5 | W6 | W7 | W8 |
|---|---|---|---|---|---|---|---|---|
| Export to CSV | 2 | 3 | 3 | 5 | 7 | 9 | 12 | 15 |
| Single sign-on | 6 | 6 | 7 | 7 | 8 | 8 | 9 | 9 |
| Slack alerts | 4 | 3 | 5 | 4 | 3 | 4 | 2 | 2 |
Export to CSV went from 2 to 15 mentions a week in eight weeks.
Share of quotes by source, this month against last.
Last month
This month
Chat and surveys are growing. Calls are still the biggest source.
Mentions per tag per week. Darker means more.
Reporting is heating up while onboarding cools off.
Quotes found, accepted by you, grouped into requests, shipped.
| Quotes found | 412 |
|---|---|
| Accepted by you | 164 |
| Grouped into requests | 38 |
| Shipped | 12 |
31 people were told when these shipped.
Who it is for
Pick a role. Each one gets its own story and the questions it would ask.
The difference
Same job: decide what to build from what customers said. Different weekend.
For developers
One request from your help desk, form tool or script, and it lands in the review queue. Nothing is acted on until a person accepts it.
curl -X POST "$API/v1/ext/feedback" \ -H "Authorization: Bearer $KEY" \ -H "Content-Type: application/json" \ -d '{ "text": "We rebuild this every Monday.", "source": "Help desk", "title": "Ticket 4812: export", "externalId": "4812", "email": "priya@example.com" }'Integrations
Anything that can send a web request already works. The rest are switched on once each has been tried against the real service.
Trust and privacy
Rules the product enforces in code, and the numbers we have measured, with how.
40 of 40
trick questions refused
20 unanswerable and 20 with a false premise. A small synthetic set, measured once against one provider. A measurement, not a guarantee.
~2 ms
to load a page of a filtered view over 100,000 quotes
Measured locally on one machine with 50 requests. Real workloads vary.
Nothing unapproved is sent
The send worker itself refuses a message that has not been approved, not just the screen.
Every stored quote must appear word for word in its source at the stored position. Answers cite quotes or say there is no evidence.
Card numbers, keys and passwords are redacted at ingest, before storage and before any AI call. Every message to a customer has an unsubscribe link, and opt-outs are honoured.
Separation is enforced in the database, and every request runs inside it. Teammates only see what their role allows.
Customer text only goes to AI routes marked as allowed to receive it. Every AI call is logged with provider, model, tokens and cost.
Quotes, transcripts, recordings, survey answers and files go together. You can export your data at any time.
A quality page shows how often suggestions were right, and a weekly spot check samples answers for a person to grade.
Read the security page or the list of subprocessors.
Pricing
Prices are announced before launch. This is what each plan includes today.
Try it with your own conversations
For product and research teams
For larger teams and heavier use
Questions
It reads customer conversations from wherever they happen (support chats and email, calls, surveys, forms, team chat), proposes the requests hidden in them with the exact sentence that supports each, lets you accept or reject every suggestion, weighs requests by who asked and what their account is worth, and helps you tell the people who asked when it ships.
It is built so you do not have to take its word. Every quote is checked to appear word for word in the original at the stored position; a made-up quote cannot be saved. Answers cite quotes or say there is no evidence. Suggestions never become requests until a person accepts them, and the platform measures how often its own suggestions were right.
No. When a request ships, it prepares a personal draft for each person who asked, in their own words. You review and approve; only approved messages are sent, and that rule is enforced in the sending code itself, not only on the screen. Every message carries an unsubscribe link and opt-outs are honoured.
Only your workspace, and only the teammates who have a role in it. Each workspace is isolated at the database level. Customer text goes only to AI routes marked as allowed to receive it, is never used to train models, and sensitive numbers (card numbers, keys, passwords) are hidden before storage and before any AI call. The subprocessors page lists every company that handles data for us.
Deleting a person removes their quotes, conversations, transcripts, recordings, survey answers and files. Working data such as past questions and AI cost records also expires on its own schedule, listed in the privacy policy.
No. Paste one conversation and you get cited requests in minutes. Everything else (people and accounts, tags, integrations, surveys) can wait.
Paste a customer conversation and get cited requests in minutes. No setup first.
Paste one conversation. See cited requests.
Start free