Always-on internal customer-feedback intelligence — without a paid VoC tool
Customer-feedback intelligence, on infrastructure you already own
It reads what the company's own customers are saying — across support tickets, in-product surveys, and public review sites — and turns it into a dashboard, a Monday-morning digest, and plain-English answers on demand. Built inside the warehouse the company already ran, instead of on a ~$10K/year off-the-shelf tool.
Customer feedback at the company lived in half a dozen places — tickets, in-product surveys, public reviews, sales calls — and the only way to synthesize it was a person reading and copy-pasting for days, losing about a month a year. The off-the-shelf fix was a five-figure annual SaaS that would also move every customer quote into another vendor's database.
An always-on internal-feedback intelligence system built inside the warehouse the company already ran — same outcomes as the paid tool (dashboard, weekly digest, ad-hoc plain-English queries with sourced quotes) — for under a fifth of the cost, with customer data that never leaves the building.
| Before | After | |
|---|---|---|
| Where feedback lives | scattered across tickets, surveys, reviews, sales calls | one place, inside the warehouse already running |
| Effort per question | hours of copy-paste and hand-synthesis into a doc | a one-line question, answered with source-linked quotes |
| Cadence | a manual reading sprint once a quarter, if at all | a live dashboard, a Monday digest, and answers on demand |
| Annual cost | ~$10K/yr for the off-the-shelf SaaS | well under $2K/yr, on infrastructure already paid for |
| Where customer data ends up | another vendor's database, behind another login | never leaves the company's own warehouse |
| Time to set up | a months-long vendor rollout | about four weeks to ship v1 |
A paid SaaS line item with the data leaving the building became a quiet always-on capability inside the warehouse the company already ran — for under a fifth of the price, shipped in weeks, with the security review reduced to 'where does the data go?' / 'it doesn't.' The researcher's hand-coded month a year goes back into the work that actually needed a human; everything that was just listening becomes the system's job.
What I built
A customer-feedback intelligence system, built end-to-end inside the company's existing data warehouse instead of on a paid SaaS. It pulls quotes from every channel where customers leave words, runs the AI on them inside the warehouse so the data never leaves the company's tenant, and serves the results three ways the team can actually use.
- A dashboard. Filter by segment, channel, date range, or theme — built on the BI tool the team was already using, sitting on the same auto-tagged quotes the digest and the
/voc-askcommand use. One source of truth, three lenses on it. - A weekly Slack digest. Every Monday morning, an auto-post: the top themes of the week, the biggest movers, the verbatims worth reading. The thing leaders read with their coffee instead of a stale report.
- A
/voc-askcommand. Plain-English questions answered in plain English: "what's driving churn?", "what do users in this region need?", "have we already heard this complaint?" Every answer comes with real customer quotes attached, each linked back to its source ticket, survey response, or review — so the team can verify any claim in one click.
The proof-of-concept ingested thousands of customer quotes across four channels (support tickets, in-product surveys, and two public review sites), auto-tagged each with theme and sentiment, and shipped the dashboard, the digest, and the /voc-ask script in about four weeks.
Why it matters
The team kept asking "have we already heard this complaint?" and the honest answer was "probably, but I don't know where." That isn't a tool problem — it's a "the words are everywhere and listening to all of them is a full-time job nobody has" problem. The off-the-shelf fix solved it with a five-figure annual subscription and another vendor holding copies of every customer quote. The replacement is the same outcomes — dashboard, weekly digest, ad-hoc queries with sourced quotes — built on infrastructure already paid for, with the AI running inside the warehouse so the data never moves. Under a fifth of the cost, weeks to ship instead of months, and a security conversation that gets short.
Normally "always-on customer-feedback intelligence" describes a category of SaaS with a sticker price to match. Here it's a Monday-morning Slack post, a dashboard a designer can filter mid-meeting, and a one-line question that returns the customer's own words — built in-house, on the stack the company was already paying for.
The hard part isn't moving the text — most channels have APIs. It's keeping the AI inside the company's own walls. The off-the-shelf tool would have meant another vendor with copies of every customer quote, another inference pipeline to manage, and another identity model to keep in sync with the customer-ID layer the rest of the company already uses. Here, the LLM runs inside the warehouse itself, on data that never moves: no API keys to rotate, no rate limits to engineer around, no separate tenant to argue about in a security review. That single architectural choice — keep the AI where the data already is — is what makes the cost math work and the rollout fast. The second hard part is the auto-tagging: themes only earn their place by agreeing with a researcher's hand-coded sample, so the dashboard reflects how the team actually thinks about customer feedback, not how an LLM happens to cluster sentences.
The shift is from buying a layer to using the one you already own: the warehouse, the AI inside it, the customer-IDs that tie feedback back to accounts. The same answers — what customers are saying, what's driving churn, what users in a region need — arrive on a dashboard, in a Monday digest, and on demand through a one-line question, for under a fifth of the SaaS the company nearly bought.
Which channels are worth ingesting, how the auto-tagging is calibrated against hand-coded ground truth, and how the in-warehouse AI is sequenced so the cost math holds at full scale are the parts the client paid for.