Some of your customers read email. Many do not. If your tool can only send email, your score only tells you about the people who read email. Look for email, SMS and WhatsApp. Also look for a link that can be opened during a phone call, one per customer, if you still call people.

You will want to change a question, add a language, or skip a question for some people. Check whether you can do that yourself. If every change is a request to the vendor, you will stop making changes.

Your first draft is never right. You will rewrite questions. Ask what happens to your trend when you do. Some tools tie your history to the exact words of the question, so changing a word starts a new line and your old data no longer joins up. Ask about the open comment box too, not just the score.

Scores like NPS and effort scores have a set way of being worked out. It is not an average. A tool that averages a 0 to 10 question is giving you a number you cannot compare to anything, including your own last quarter. Also ask if the method stays fixed once you pick it.

If you run three surveys, one person can get all three. That is the fastest way to kill your response rate. Look for a limit on how often one customer can be contacted, counted across all your surveys and not just one.

Ask your own security team first. Many banks and insurers say no. If yours does, you need a tool that keeps names, phone numbers and account numbers on your own servers. It swaps them for codes before anything is sent out, and only turns them back into real names inside your own network.

Tick yes if the tool can do that, or if your security team is happy for the data to leave. Check this now. Otherwise you can get all the way to signing and then be stopped.

Thousands of comments are useless as a pile. You want them sorted into your own groups, with big groups broken into smaller ones. Check whether you build those groups or whether the tool ships a fixed list. No two companies have the same problems, so a fixed list will not fit yours.

It will get some wrong. What matters is what happens next. Look for a way to check a batch, fix the mistakes, and have the tool learn from your fixes. If you cannot correct it, you will not trust it, and if you do not trust it you will not use it.

"The app crashed and then nobody answered the phone" is two problems in one sentence. If the tool files it under one heading, you lose half of it. This has a side effect worth knowing: your category percentages will add up to more than 100. That is correct, and a good tool will not hide it.

Same trap as 1.3, one level down. You will reorganise your categories as you learn. Ask whether the trend for a category survives that.

You want a list of what to fix first. Two things decide if that order is real. First, ask whether the things on the list have passed a test, or are just the ones that came out highest. Second, ask whether it counts people or counts answers. One customer who picks three problems is one person, not three, and tools that get this wrong will overstate whatever gets picked alongside other things.

If the answer to "which branches?" is a spreadsheet someone has to prepare, the analysis gets done once, by one person, and then never again.

Small groups produce wild results. A good tool hides a chart when too few people are behind it, rather than drawing it anyway. Ask what its limit is and whether you can see it.

Boards ask for target versus actual. Some tools build every chart only from survey answers, so there is nowhere to attach a target line or a range you set yourself. If your board pack needs this, find out before you buy, not after.

Your branch managers will not learn a reporting tool. They should be able to type a plain question and get an answer. Check that the answer shows its working: from the score, down to the causes, down to the actual comments customers wrote. An answer you cannot trace back is a guess with a nice font.

Giving decisions to more people means giving data to more people. You need to control who sees what, create your own roles, and connect it to the sign-in system you already use. Without this, your risk team will stop the rollout.

Two ways to solve this, and either one counts as a yes.

If you took the route in 1.6, the details never reach the tool at all, and you are done.

If they do reach the tool, you should be able to mark a field as personal with a switch, then choose which roles can see it. A service rep working a case may need the customer's name. Everyone else does not. Check that this applies to downloads as well as screens. Hiding a name on screen and then letting anyone export it is not hiding it.

Ask whether an alert fires on one customer's answer, on its own, straight away. Some tools collect them up and send them overnight. By the morning, the customer has already told somebody else.

A low score is the obvious trigger. It is not the only one. Look for the ability to set different limits for different customers or regions, to fire on certain words in a comment whatever the score was, and to fire on the tone of a comment. A score-only rule misses the customer who gives you an 8 and then describes something serious.

The case should land with the branch, the dealer or the region it came from, worked out from the data itself. If everything lands in one central queue, somebody has to read and forward all of it, and that job is what kills these programmes.

You should be able to build your own steps, as many as you need. Two, five, twenty. Each one marked as open or closed. Some needing a note before anyone can move on. Some locked so a case cannot be passed to someone else.

Here is why it matters. A useful default separates "waiting on us" from "waiting on the customer." Most systems roll those into one. When they are rolled into one, a case sitting for four days because your team has not touched it looks exactly like a case sitting for four days because the customer has not replied. Only one of those is your fault, and you cannot see which.

This is the one people forget to ask. You want to say: if a case has been open three days, tell the manager. If it is still sitting in one step after a week, tell their boss. Look for a way to build those rules yourself, choosing the wait, the condition and who gets told. If it can only be set up by the vendor, every change is a support ticket.

Due dates. A flag when something is late. Live counts of what is open and what is overdue. And a time-stamped record of every change, every handover and every contact. If you cannot show it, you did not close the loop. You discussed it.

Adding a note, sending an email, logging a call, attaching a file, sending a follow-up survey, changing how the case is filed. Every jump to another piece of software is a place the work stops.

In some markets you must have permission before contacting someone about their feedback. The place to ask for it is the survey. If the customer says no, no case should be created.

This is a survey design choice, made before you launch, not something you sort out later. It also means fewer customers can be contacted than complained, which changes how many staff you need. Rules differ by market, so take your own advice on what applies to you.

Every step that customer met, on one line, in the order it happened. With their opinion attached at the steps where they answered a survey, and nothing invented at the steps where they did not.

Ask what this costs to set up. It should need no new data feed and no integration work, because the event that says a customer came back is the same event that sends them their next survey. You are already sending it.

This is the money question. You want a number that says: customers who rated this well came back more often than customers who rated it badly.

Then ask how that number is made. If it is a count of what your own customers actually did, you can put it in front of your finance director. If it is a model, you will spend the meeting defending the model instead of the finding.

Four things to check about that number. Each is its own yes or no.

Four sub-questions follow. Each one is its own yes or no, and each one counts.

Your big yearly or half-yearly score arrives too late to do anything about. The useful trick is to connect the small surveys you run all the time to that big score, so the small ones warn you early.

Two things to check. The delay between them should be measured from your own data, not guessed. And ask honestly how far ahead it can see. It can only reach as far as the shortest delay in the chain, because after that the early warnings have not happened yet.

At this level you need more than a screen. You need someone who can look at your data, work out why one group of customers comes back and another does not, and help you set a target you can defend.

A target should come from your own data and from how hard your problem actually is to fix. Not from a benchmark, and not from a rule about how far a score can move in a year. Ask who does this work, whether they are included, and what they have done before.