Customer sentiment analysis

What it does

Sentiment analysis reads free-text customer feedback attached to unified profiles and returns two things per customer: a sentiment score on a −5 to +5 scale, and the business aspects the feedback is about — pricing, delivery, support and so on. It's an out-of-box text model, not a general-purpose LLM prompt over your data.

Key facts

Still a preview feature. Microsoft's own wording is that preview features aren't meant for production use and may have restricted functionality. Treat any client commitment accordingly.
Created from Insights > Predictions > Create, on the Customer sentiment analysis (preview) tile. Contributor permissions or higher.
English only at present. Feedback in other languages still returns results, but Microsoft states those results won't be accurate — which is worse than no result, because nothing on screen flags them.
Up to 10 million feedback records per model run.
Each comment is analysed to a maximum of 128 words. Longer comments are truncated and only the first 128 words are considered.
Feedback must be unified and mapped to the Feedback semantic activity type, with a feedback ID, a timestamp, the text field and a customer ID that matches unified profiles.
Only one feedback table can be configured.
You name two output tables during setup: one for sentiment scores and one for business aspects.
Business aspects come from a fixed predefined list, including account management, checkout and payment, customer support, in-store pickup, packaging and shipping, pre-ordering, price, privacy and security, promotions and rewards, receipt and warranty, returns and cancellations, fulfilment accuracy, and website or app quality. You cannot add your own.
Two NLP models sit behind it, trained on public data spanning social media, retail, restaurant, consumer products and automotive. That training mix is a reasonable proxy for whether it will suit your client's vocabulary.
Results include average sentiment distribution, customers grouped as negative, neutral or positive, sentiment trend over time, sentiment split by business aspect, influential words, and colour-coded sample feedback showing which words drove the score.
Microsoft flags potential bias and privacy compliance as considerations for this feature specifically.

When to use / skip

Use it where a client already collects a decent volume of English free-text feedback — post-purchase surveys, support case notes, review feeds — and wants it attached to customer profiles rather than sitting in a survey tool. The business-aspect breakdown is the genuinely useful part: it turns "our NPS verbatims are bad" into "our verbatims are bad about delivery and returns", which is actionable.

Skip it on any multilingual client base unless English is overwhelmingly dominant. A UK-only retailer is fine; a European or Middle East rollout is not, and the model won't tell you it's producing rubbish. Skip it too where the fixed aspect taxonomy doesn't match the business — the list is retail and consumer shaped, and a B2B industrial client will find half of it irrelevant with no way to extend it.

And be careful about preview status in regulated or high-visibility deployments. If the client's governance process requires GA-only features, this doesn't pass, and finding that out at the security review is avoidable.

Configuration decisions

Which feedback source becomes the single configured table. Survey verbatims, review text and case notes all behave differently, and you only get one.
Whether to consolidate several feedback sources upstream into one table, and how you'd tell them apart in the output afterwards.
Whether to filter out non-English records before ingestion rather than letting the model score them badly.
How to handle long-form feedback given the 128-word cut-off — truncate at source, split into rows, or accept the loss.
Naming for the two output tables, since both are user-defined and both need to be recognisable to whoever consumes them later.
Whether sentiment scores feed segmentation or stay reporting-only, given the preview caveat.

Gotchas

The language limitation fails silently. Non-English rows produce scores that look exactly like valid ones. If your client's feedback has any multilingual tail, you need a filter, not a caveat in a slide.
The 128-word truncation hits detailed complaint text hardest, which is exactly the feedback with the most business value in it.
The aspect list is fixed. Clients ask to add their own categories almost immediately, and the answer is no.
One feedback table only, so a client with survey data and support notes has to make a choice or do the consolidation work upstream.
Preview means the feature and its output shape can change. Anything you build on top of those two output tables carries that risk.
Sentiment is computed per feedback record and surfaced per customer. A customer with one furious comment and twenty neutral ones may not read the way the account team expects.

Consultant notes

State the preview status in writing, in the design document, not just verbally in a workshop. It's the single most likely thing to be raised later by someone who wasn't in the room.
Demo the colour-coded feedback samples. Showing which words drove a score is what converts a sceptical CX team, far more than the aggregate charts.
Sanity-check the aspect distribution against what the support team already knows. If the model says price and they say delivery, look at your data mapping before you look at the model.
Push back on using sentiment as a hard trigger for customer-facing action while it's in preview. Reporting and prioritisation, yes; automated apology emails, not yet.
Run a language audit on the feedback source during discovery. It takes ten minutes and it decides whether this workstream is viable at all.

Worth revisiting once this leaves preview, or when Microsoft adds languages beyond English.