What goes into it
Sentiment scoring runs across the text the relationship already generates rather than asking for new input.
- Support tickets and their comment threads, including the messages after the resolution. - Email correspondence with the account team. - Call and meeting transcripts, where tone and hesitation carry more than word choice. - Open-ended survey comments, when they exist.
Each interaction gets scored, then the interactions roll up to an account score weighted by recency and by who was speaking. Both weights matter. A frustrated message from six months ago that was resolved should not drag today's number, and a complaint from a daily user is a different object from the same complaint made by the economic buyer.
Read the change, not the level
Absolute sentiment levels are close to meaningless across accounts. Some customers write curtly and always have. Some industries communicate in a register that a general-purpose model reads as negative. Comparing one account's sentiment level to another's produces a ranking of writing styles.
The signal is movement against the account's own baseline. A customer whose tone has been consistently brisk for two years and turns warm is telling you something. A customer whose sentiment slides across three consecutive months is telling you more, and the drift matters even when the current level still looks acceptable.
The other pattern worth alerting on is the disappearance of text altogether. ORM's view on deal slippage is that the earliest signal is the lack of a signal, meaning no activity, no data changing, and no notes. Retention works the same way. An account that stops writing to you produces no sentiment to score, and that gap is the finding rather than a data quality problem to be filled in with a neutral default.
Traceability is the adoption requirement
A sentiment score gets ignored the first time a CSM cannot see why it moved. ORM's position on AI in revenue work is that the biggest gap is trust and traceability, and that a model needs to point back to the source that drove the number, because validating an unverifiable output costs as much as producing it by hand.
Build the score so every account number expands into the specific messages behind it. That single design choice decides whether the team acts on the score or quietly stops looking at it, which in turn decides whether it does anything for net revenue retention at all.
Frequently Asked Questions
How is a sentiment score calculated?
A language model scores the tone of each support ticket, email thread, and call transcript, then those scores roll up to an account average weighted by recency and by the sender's role. The account-level number is the output. The per-interaction scores are what make it defensible when someone challenges it.
Is sentiment score better than NPS?
It has better coverage. Sentiment scoring reads every interaction the account had, while a survey reads the subset of people who responded. That makes sentiment continuous and unbiased by who bothers to reply. It is also noisier, because tone reflects the urgency of the moment as much as the state of the relationship.
What is the biggest failure mode?
Scoring the wrong person. Sentiment from an end user annoyed by an interface detail carries far less renewal risk than mild dissatisfaction from the executive who signs the contract. Weight by role before rolling up, or the score becomes an average of opinions that do not decide anything.
Should sentiment scores be trusted without seeing the source?
No. ORM's position on AI-generated analysis is that the gap is trust and traceability, and that a model needs to point back to the point of truth behind its numbers. A sentiment score that cannot show the messages that produced it is a number nobody acts on when it matters.
Put these metrics to work
ORM builds custom revenue forecast models that turn concepts like customer sentiment score into prescriptive action for your team.
Schedule a Demo