The platform

Feedback lands from any channel and is normalised into one canonical format: one row, one piece of customer feedback. It passes into a governed layer where your metrics, your glossary and your rules live, approved by a human before any agent can read them. Specialist agents then do the analysis in sequence, and a final agent challenges their work before you see it. Every output carries its provenance: where the number came from, which query produced it, how fresh the data was.

Four channels, each properly built. Web surveys by email invitation. DIGI for on-site intercepts, pop-ups and embeds, with a chat mode for step by step surveys and matrix questions for grids that share a scale. SMS through your gateway. IVR with keypad capture and recorded voice answers.

Speech is not a bolt-on. Open answers on IVR are recorded and transcribed, and VOICE is a first-class column type in the data model, not an attachment. Speech to text has shipped since February 2021.

90+ connectors, and four ways in. Scheduled SFTP import on your own timetable, down to every thirty minutes. Manual file import. Typed upload into Store, where a column header marked TEXT, NPS, DATE or CHOICE tells the platform what it is looking at. And a JWT authenticated REST API for everything else.

250+ languages, and the mechanism matters. Verbatims in any language supported by the translation layer are normalised into one base language, and from there the whole analytics stack applies: lemmatisation, topic extraction, sentiment, the lot. Your German, Italian and Thai feedback ends up in the same analysis, compared on the same terms.

Contact policy built in. Quarantine and blacklist rules stop you over-surveying the same customer. Anonymity settings and GDPR handling are configuration, not a services engagement.

43 analysis types. Sentiment, themes, pain points, touchpoints, journey stages, NPS CSAT and CES, churn risk, segmentation, competitive comparison, operational metrics, aspect-based sentiment and complaint risk. Each one is a senior analyst's procedure, encoded so it runs the same way every time.

Aspect-based sentiment, against your taxonomy. Sentiment scored per aspect from minus one to plus one, with a confidence level on every score. The model can only assign categories that exist in your approved taxonomy. It never invents one. That is why wave two is comparable to wave one: Pricing still means Pricing.

You own the taxonomy. TopicAI reads the feedback and proposes topics from what customers actually said. You select, group, name and apply them, in the interface, without a consultant and without a change request. Change it later and old data can be re-scored against the new version.

Complaint risk. Every complaint scored on likelihood of escalation and severity of impact, zero to one, and placed in a matrix that tells you what to engage, escalate, monitor and contain. No training data required. It runs on any complaint dataset from day one.

An agent with 104 tools, not a chatbot with a prompt. Ten tool packs: analysis, topics and risks, datasets, dashboards, library, capture, stories, memory, SQL and live agent monitoring. It does not describe what could be done. It calls tools that write to real tables.

Ask it to build a survey and it builds the survey. Ask for a dashboard and it creates the dashboard and the widgets on it. Ask it to extend the taxonomy and the taxonomy changes. This is the difference between an assistant that drafts you a ticket and a colleague who does the work.

And it shows its working. Every answer carries the source it drew from, the query it ran, the row counts and the freshness of the data. If it had to guess, it says so.

Analysis runs as a pipeline, not a prompt. Nine specialists in sequence: data quality, domain context, sentiment, themes, trends, segments, risk, then synthesis. The ninth is a QA agent whose only job is to challenge the other eight. It re-reads their SQL, checks their assumptions against your definitions, flags findings that lack evidence and adjusts the confidence scores down when they are not earned.

The principle it enforces has a name inside the product: do not self-certify. A stage that fails review does not quietly pass downstream.

Plug a language model into a data warehouse with no semantic layer and you get confident, plausible answers that are wrong in ways only your own analyst would catch. The Knowledge Pack is what stops that.

It holds your approved metrics with the exact formula behind each, your glossary of what terms mean in your organisation, the rules the agent must follow, the documented traps in your data, golden question and query pairs, and per-dataset reference docs.

The model can draft all of it. A human approves it. Nothing enters the agent's context until someone has said yes. Raw exploration is the discouraged fallback, and when it is used the answer is tagged as such so nobody forwards it to a board pack by accident.

Dashboards and widgets you build yourself or ask Scoty to assemble from a finished analysis. Narrative stories for the people who will never open a dashboard. Alerts on the cases that need a human today, routed into ticketing and tracked to closure, so the loop actually closes instead of ending in a report.

SANDSIV is Swiss, with no US parent company. Run sandsiv+ on European public cloud, private cloud, in your own data centre, or fully on premise with the language model inside your walls. The US CLOUD Act reaches US-headquartered providers wherever their servers sit. We are not one, so a European region here is a change of jurisdiction and not just a change of address. Nothing you put in ever trains a shared model.

ISO 27001, ISO 27701, ISO 27017, ISO 27018, SOC 2 Type II, CSA STAR Level 1, GDPR native. Open architecture, full exportability, and your data stays yours including on the way out.