Case studies
Customer Intelligence in action.
Two realistic scenarios show what happens when fragmented feedback meets governed agentic analysis — and how to reproduce the results on your own data. They are illustrative — built on real platform capabilities, not invented testimonials — and they map directly to analyses you can run yourself today.
A note on sourcing
These case studies are illustrative scenarios. They describe realistic deployments of InsightNarrator based on the platform's actual capabilities — the connectors, analysis types, and agent pipeline all exist and work as described. The organizations, however, are composites. We'd rather show you exactly what the platform can do with plausible data than attach a logo to a story we can't fully verify. If you want a live proof-of-concept on your own data, that's what the free trial is for.
Case study · Enterprise · Banking
How a European bank consolidated 5 VOC sources into one governed intelligence layer.
The Challenge
The bank ran customer experience measurement the way most large banks do: in silos. The central CX team owned the Medallia NPS program. Each country team managed its own internal surveys. Reputation sat with marketing, who watched Trustpilot and Google Maps branch reviews in separate dashboards. The branch network had its own contact-form feedback that rarely reached headquarters in a usable form.
None of this was broken individually. All of it was broken together.
The CX team could tell the board that NPS moved from 32 to 35 in a quarter. They could not tell the board why, and they could not tell the board whether the "why" in Germany was the same as the "why" in France. Sentiment analysis existed, but it ran per channel, on different taxonomies, with different definitions — so a "negative" in Trustpilot and a "negative" in the NPS verbatim meant different things.
Three weeks of analyst time per reporting cycle was spent reconciling formats, hand-coding a sample of verbatims, and producing a slide deck that was already stale by the time it reached the risk committee.
The Solution
The bank connected all five feedback sources to a single InsightNarrator workspace. Data was stored and processed in the EU, with LLM inference in Switzerland under EU adequacy. The connectors were direct: Medallia and the internal survey via API, Trustpilot and Google Maps via the review-source layer, branch contact forms via scheduled CSV upload.
With the sources unified into canonical datasets, the team did three things:
- 1
Built a governed taxonomy. Before any analysis ran, the CX lead defined the bank's categories and subcategories — Mortgage, Cards, Digital Banking, Branch Service, Complaints Handling — and the metrics and glossary terms that went with them. This is the Knowledge Pack: human-approved definitions the agents are required to use. No more invented categories.
- 2
Ran ABSA across all channels. Aspect-Based Sentiment Analysis scored every verbatim against the governed taxonomy. Because the taxonomy was shared across all five sources, sentiment was finally comparable. A “negative” meant the same thing in a Trustpilot review and an NPS verbatim.
- 3
Ran C-Risk across all channels. Complaint Risk scoring crossed the ABSA results with severity and risk factors. The output was a risk matrix: category × subcategory × severity. The team could see, for the first time, the specific cells generating complaint-risk signal — not a sentiment index, but a ranked list of problems with verbatim counts attached.
The multi-agent pipeline handled the orchestration: Data agent normalized, Domain agent applied the taxonomy, Sentiment and Themes agents ran the ABSA, Risk agent computed C-Risk, QA agent challenged the findings before delivery.
The Results
| Metric | Before | After |
|---|---|---|
| Analysis cycle time | 3 weeks | 2 days |
| Feedback sources reconciled | 5 separate | 1 unified |
| Analysis types per cycle | 2 (manual coding + sentiment) | 15+ (ABSA, C-Risk, themes, trends, segments) |
| Time to surface top complaint-risk cell | ~21 days | Same-day on weekly refresh |
| Report staleness on arrival at risk committee | 2–3 weeks | Under 48 hours |
The single highest-impact finding: the cell Mortgage → Processing Time generated the most severe complaint-risk signal across all five channels and all five markets. It had been invisible before because no single channel showed it strongly enough to trigger attention — but aggregated and scored, it was the clear #1.
Weekly C-Risk dashboards now ship to branch managers by region, with the top three risk cells and their verbatim counts.
At a glance
- Industry
- Retail banking
- Size
- 50,000+ employees
- Geography
- 5 European markets
- Feedback sources connected
- 5 (Medallia, Trustpilot, Google Maps, internal NPS survey, branch contact form)
- Primary analyses run
- ABSA · C-Risk · Theme evolution
- Top finding
- “Mortgage → Processing Time” was the #1 complaint-risk cell across all channels
- Cycle time
- 3 weeks → 2 days
Connect 5 sources and run ABSA + C-Risk
Start free trial3 weeks → 2 days
Analysis cycle time
5 → 1
Sources reconciled into one view
15+
Analysis types per cycle
< 48h
Report freshness at risk committee
We spent three years trying to reconcile five feedback channels into one view. The reconciliation was never the bottleneck — the definitions were. Once the agents were disciplined by a shared taxonomy, the analysis was almost the easy part.
Case study · Market research agency
How a market research agency cut verbatim analysis time by 70%.
At a glance
- Industry
- Market research agency
- Size
- 45 employees, 15 active clients
- Geography
- DACH region
- Core pain
- Manual coding of open-ended responses consuming 60% of analyst hours
- Primary analyses deployed
- 43-type library, per-client taxonomy
- Top result
- Analysis time down 70%; per-project margin up from 22% to 38%
Upload a dataset and let Scoty recommend an analysis plan
Start free trialThe Challenge
The agency's value proposition was depth: qualitative insight that automated tools couldn't match. In practice, that meant senior analysts spent the majority of their week hand-coding open-ended responses — line by line, code frame open on one monitor, spreadsheet on the other. It was careful work. It was also expensive, slow, and impossible to scale without hiring.
Three problems compounded:
- 1
Coding was inconsistent across analysts. Two people coding the same 5,000 verbatims produced noticeably different code distributions. Inter-rater reliability existed on paper; in practice it drifted.
- 2
Reports were thin relative to the effort. Because coding consumed so much time, each client deliverable typically contained three analysis types: a sentiment breakdown, a top-themes list, and a few pulled quotes. The underlying data supported far richer analysis — there was just no time to run it.
- 3
Margin was eroding. Clients wanted more analysis for the same fee. The agency couldn't raise prices without justifying the increase, and couldn't justify it without showing more value.
The Solution
The agency deployed InsightNarrator as its core analysis engine, with a separate workspace per client and a per-client governed taxonomy. The code frames the analysts had been using by hand became the Knowledge Pack definitions — the same categories, the same subcategories, now enforced consistently across every project.
Each new client project followed the same flow:
- 1
Upload or connect the response dataset (CSV, Excel, or a survey-platform connector).
- 2
Scoty, the analysis-recommender agent, reviewed the dataset and suggested an analysis plan based on the data shape and the client's question. Analysts could accept, edit, or override.
- 3
The multi-agent pipeline ran the analyses — typically 12 to 18 of the 43 available types per project, depending on the client question. Sentiment drivers, theme extraction, root cause on detractors, segment comparisons, trend over waves.
- 4
The QA agent challenged the findings before they reached the analyst. Anything the QA agent flagged went to a human reviewer — but the vast majority passed without issue.
- 5
Analysts focused on interpretation and narrative, not coding. The hours previously spent on manual coding went into writing the actual report.
Crucially, the agency did not fire its analysts. It redeployed them. The work that required judgment — framing the client's question, interpreting the findings, writing the narrative — was the work the agency was actually being paid for. Coding was not.
The Results
| Metric | Before | After |
|---|---|---|
| Verbatim analysis time per project | Manual coding | −70% |
| Analysis types per client report | 3 | 15+ |
| Inter-rater consistency | Drifted by analyst | Enforced by governed taxonomy |
| Per-project gross margin | 22% | 38% |
| Senior analyst hours on coding | ~60% | ~15% |
| Senior analyst hours on interpretation & narrative | ~25% | ~65% |
The margin improvement had two causes, both real. First, the direct cost of coding dropped sharply. Second — and more importantly for client retention — the agency could deliver substantially richer reports (15+ analysis types instead of 3) for the same fee, which made renewal conversations easier and let the agency justify a price increase on new engagements.
−70%
Verbatim analysis time per project
3 → 15+
Analysis types per client report
22% → 38%
Per-project gross margin
~65%
Analyst hours on interpretation
We didn't become an AI agency. We became an agency where our best people do the work only our best people can do, and the machine handles the coding we were never paid enough to do by hand anyway.
Which profile are you?
Two different problems. One platform.
Primary pain
Enterprise bank
Fragmented feedback sources, no unified view
MR agency
Manual coding consuming analyst capacity
Feedback sources
Enterprise bank
5 external channels, enterprise scale
MR agency
Client-supplied datasets, per-project
Key capability used
Enterprise bank
Connectors + ABSA + C-Risk + multi-market rollup
MR agency
43-type analysis library + per-client taxonomy
Governance lever
Enterprise bank
Shared Knowledge Pack across 5 markets
MR agency
Per-client taxonomy = consistent code frames
Cycle time win
Enterprise bank
3 weeks → 2 days
MR agency
−70% per project
Margin / value win
Enterprise bank
Faster, actionable risk reporting to branches
MR agency
Margin 22% → 38% via richer deliverables
Best fit for you if…
Enterprise bank
You run VOC across multiple channels or regions
MR agency
You analyze qualitative data for clients at volume
"I have fragmented feedback sources and need one governed view."
Book a demo"I analyze verbatims for clients and coding is eating my margins."
Start free trialCase study · Pending
Your company could be next.
We are actively looking for organizations willing to run InsightNarrator on a real dataset and let us write up the results — with you, transparently, using your actual numbers. Enterprise, mid-market, agency, public sector: if you have customer feedback and a question worth answering, we want to hear from you.
Beta-program participants get hands-on support from our team, a governed workspace configured to your taxonomy, and the full platform for the duration of the engagement. In return, we ask for a public case study once we've found something worth reporting — metrics you're comfortable sharing, in a format you approve.
No obligation. We'll tell you honestly if your dataset isn't a fit.
Reproduce this yourself
Five steps from signup to governed insight.
- 1
Connect a source.
Pick one of 41 data sources: a survey platform connector (Medallia, Qualtrics, SurveyMonkey, Typeform, and 21 more), a review/social source (Trustpilot, Google Maps, Amazon, and 13 more), or a direct CSV/Excel upload. The bank in Case Study #1 connected five; you can start with one.
- 2
Let Scoty recommend an analysis plan.
Scoty, the analysis-recommender agent, inspects your dataset and suggests the analyses most likely to answer your actual question. Accept the plan, edit it, or override it. The agency in Case Study #2 used this to go from 3 analysis types per project to 15+.
- 3
Approve or build a governed taxonomy.
Before sentiment and risk analyses run, confirm the categories and subcategories the agents must use. Pre-built taxonomies exist for common verticals; you can also have one generated and then edit it before approval. This is what makes results comparable across sources and over time.
- 4
Run the multi-agent pipeline.
Nine agents run in sequence — Data, Domain, Sentiment, Themes, Trends, Segments, Risk, Synthesis, QA — each bounded by an iteration budget and each passing findings forward. The QA agent challenges the output before it reaches you. Watch progress in real time.
- 5
Read the output, then act.
Findings land in the Library with provenance attached: which source, which analysis, which confidence. Build a dashboard. Share a story. Export to your reporting stack via API. The bank shipped weekly C-Risk dashboards to branch managers; the agency delivered 15+ analysis types per client report.
Being precise about evidence.
What these scenarios are based on
- 41 data sources — 25 survey connectors and 16 review/social sources — all shipping and connectable today.
- 43 analysis types across 14 categories, including ABSA and C-Risk, all available in the platform.
- A 9-agent pipeline with a QA validation step before delivery.
- Governed taxonomies defined per workspace and enforced across analyses.
- Data stored and processed in the EU; LLM inference in Switzerland under EU adequacy — with audit trail and right to erasure.
What these scenarios do not claim
- That the bank and agency described are real, named organizations. They are composites.
- That every reader will see a 70% reduction in analysis time or a 16-point margin lift. Those numbers are illustrative of plausible outcomes, not guarantees.
- That InsightNarrator replaces the need for analysts. Case Study #2 is explicit: the agency redeployed its analysts, it did not eliminate them.
- That the platform is finished. Capabilities like additional connectors and deeper vertical taxonomies are still building. We'll tell you what's shipping versus building when we talk.
EU + Switzerland data sovereignty
Data stored and processed in the EU; LLM inference in Switzerland under EU adequacy
Go deeperFAQ
Questions about these case studies.
Start your own customer intelligence story.
Connect a dataset. Approve a taxonomy. Let the multi-agent pipeline find what three weeks of manual coding would have missed. The free trial is 14 days, no credit card, and every output carries provenance back to the source.
No credit card required · 14-day free trial · EU & Switzerland data sovereignty · Cancel anytime