Step by step: which parts of NPS follow-up are rules, where AI reads comments and drafts replies, and which conversations and decisions stay with a person.
Partly. Sending an NPS survey and routing the scores is rule-based automation, AI is good at reading open-text comments and drafting replies, and a person should call unhappy customers and decide what to change. The point of the survey is not the number but what happens next, and the slow part is usually the follow-up: comments pile up unread and detractors wait days for an answer. Below is how each step splits between rules, AI and people, what you need in place first, and a small pilot on one customer segment. The aim is not to automate the relationship, but to make sure every response is read, tagged and answered by the right person quickly.
What each step needs
01● Standard automation
Choosing who to survey and when
The first step is a sampling rule, not a model. Decide which customers get the survey (for example, accounts past onboarding, users after a support ticket closes, or a relationship check every six months), how often one person may be asked, and which contacts are excluded, such as open complaints, accounts in a renewal negotiation or people who opted out. A workflow tool reads the CRM or product data and builds the send list the same way every time. This is deliberately deterministic: you need to be able to explain why someone was or was not asked, and a quiet change in who gets surveyed makes your score impossible to compare over time.
⛨Works when the CRM has a clean contact owner, a last-surveyed date and an opt-out flag, and when the sampling rule is written down and changed only on purpose.
02● Standard automation
Sending the survey and the reminder
Sending is plain automation. The tool delivers the one-question survey by email, in-app prompt or text message at the agreed moment, adds the account and contact IDs to the link, sends a single reminder after a few days and then stops. Time zones, language and a maximum number of messages are settings, not decisions. Keep the survey short: the 0 to 10 question and one open box asking why. An AI model adds nothing here except risk, because a rewritten question changes what the score means.
⛨Works when each response is tied back to the right contact and account automatically, and when the sending tool respects consent and unsubscribe rules.
03● AI candidate
Reading and tagging the open-text comments
This is where AI earns its place. People write things like "support was great but the invoice portal keeps logging me out", and a language model can split that into themes (support, billing portal, onboarding, price), a sentiment per theme and a short quote. It handles several languages and misspellings far better than keyword lists. Give it a fixed list of themes so results stay comparable from month to month, and have it return the quote it based each tag on. Comments the model marks as unclear or ambiguous go to a person instead of being forced into a category.
⛨Works when the theme list is fixed and reviewed quarterly, the model must quote the text behind each tag, and a person checks a sample of tags every week at first.
04● Standard automation
Sorting responses by score, account value and risk
Once a response is tagged, rules decide where it goes. A score of 0 to 6 from a key account creates an urgent task for the account owner, a 7 or 8 goes into a monthly summary, and a 9 or 10 triggers a thank-you and, where appropriate, a review or referral request. The thresholds, owners and response times are agreed in advance and applied identically to everyone. Rules are the right tool because the routing must be predictable and auditable, and a model that decides on its own who deserves attention will quietly favour the loudest comments.
⛨Works when every account has a named owner, the thresholds are documented, and urgent tasks have a deadline that someone actually checks.
05● AI candidate
Drafting the follow-up message
A model can write a first draft of the reply using the score, the tagged themes, the customer's own words and what the CRM knows about the account. For promoters it drafts a short thank-you. For passives it asks one specific question. For detractors it drafts an acknowledgement that names the problem without promising a fix. The draft always goes to the account owner for approval, who edits the tone and removes anything that is not true. Never let a draft claim that something has been fixed or will be fixed unless the owner has confirmed it with the team responsible.
⛨Works when drafts stay in approval mode, the model only uses facts from the survey and the CRM, and promises about fixes or credits are written by a person.
06● Human review
Calling detractors and closing the loop
A detractor with an open problem needs a human conversation. Someone who understands the product, the contract and the history calls or writes personally, asks what went wrong, listens and agrees on a next step with a date. They can also decide on a goodwill gesture, bring in an engineer or escalate to management, which are decisions with money and reputation attached. AI can prepare a one-page summary of the account before the call, but the call, the apology and the commitments belong to a person. Record the outcome in the CRM so the loop is visibly closed.
⛨Works when the owner has authority to act, a target time for the first contact exists, and the outcome of each call is recorded as a status.
07● Keep human
Deciding what to change in the product or service
The survey is only useful if something changes. Reviewing the monthly themes, weighing them against cost, strategy and what customers actually pay for, and deciding what to fix first is leadership work. AI can count themes, show trends by segment and pull representative quotes, but it cannot know that the billing portal issue matters because three large renewals are coming. Keep this as a standing monthly meeting with a named owner, a short list of changes and a note to customers when you act on their feedback.
⛨Works when one person owns the feedback review, decisions are written down with a date, and customers hear back when their feedback led to a change.
A sensible first experiment
Pick one segment, for example customers who finished onboarding in the last 90 days. For four weeks, let rules send the survey and route scores, let AI tag comments and draft replies in approval mode, and have the account owner send every reply. Measure three numbers: the share of detractors contacted within two working days, the share of tags a reviewer agrees with in a weekly sample of 20, and the time the owner spends per response. Extend only if the first two numbers hold.
The trap to avoid
The usual failure is automating the thank-you and ignoring the problem. A friendly AI reply to a detractor that does not address what went wrong, or promises a fix nobody agreed, damages trust more than silence. Keep replies to unhappy customers in approval mode and never let a draft claim something was fixed.
Questions teams ask
Can AI send NPS surveys and follow up automatically?
Mostly yes, but sending and routing are rule-based automation, not AI. AI helps with reading comments and drafting replies. A person should approve replies to unhappy customers and make any promise.
Is it safe to let AI read customer survey comments?
It can be, if you check the legal basis and your supplier terms first. Comments can contain personal data, so use a provider with a data processing agreement, remove names and contact details where you can, and keep the data where you have told customers it will be.
Should AI reply to detractors on its own?
No. A wrong or generic reply to an unhappy customer makes things worse. Let AI draft and summarise, but have the account owner approve and send, and call the detractors with serious problems.
How do I know if the automation works?
Track the share of detractors contacted within your target time, the share of comments tagged correctly in a weekly sample, and how many detractors move up at the next survey. Do not judge it by response volume alone.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.