How do you analyse customer feedback in Power BI?

I'm Sarah Wilson and I work in the account management team here at Wordnerds, and I'll be your host today, so I'll be taking you through the sessions and the agenda. Then we roped in Steph from product, who's going to take us through some of the challenges she sees from clients — we've given Pete a week off, which I'm sure everyone's relieved about. And we've brought Steve back as well. Some of you will remember Steve from the series of webinars he ran last year; he's now heading up a newly formed team at Wordnerds with all things reporting. Welcome back, Steve.

Today's agenda: we've got about 40 minutes planned. We always say this stuff works much better as a dialogue and not a monologue, so we're going to have interactive points throughout. Please do let me know your burning questions in the chat — I want to put these guys to the test. It's Steph's first webinar, so we're looking for some tricky questions for her ideally.

We're going to start with setting the scene — that's me taking you through the problem and the framework. Then Steph's going to jump into the framework in more detail: the different parts of the problem and what you're ultimately looking to achieve. And then Steve's got a really great live interactive demo. We like to test these things and we don't want to do anything that's pre-rehearsed, so it's all live in the platform — he's going to take you through a full journey. Then we've got time for Q&A, which tends to be a really juicy part of the webinar. This is a community of housing people, and we want to make sure we're learning from each other — we like to steal everyone's ideas and pass them off as our own. So definitely give us all your suggestions, comments and questions in the chat.

I'll bring up a poll just to see where you're at with Power BI, how you're feeling about it, where you are in your journey. No wrong answers, as we always say. We're all here to learn, and it doesn't matter how long you've been doing this — there's still plenty to learn with this tech as it continually evolves. We do have time at the end for questions, so we'll get to them all.

So who are we? We are Wordnerds — I can see tons of names I've heard of before, lots of familiar faces, lots of clients, which is really nice. Just in case you're new here: we're a customer feedback analysis specialist. We started out predominantly in the retail space, doing a lot with M&S and Sainsbury's, then we moved into the social housing space with the change in regulations and the TSMs, and we've also been making inroads recently into financial services and travel and hospitality. So they're our main sectors.

Why do they all choose to work with us? They're all super different with very different use cases, but despite that they've all got a really similar desired outcome. They all want to turn their feedback into deep, actual insights that can be easily found, prioritised and shared through the business — and most importantly that they're actually confident in. In our experience of working with clients, Power BI is really seen as this silver bullet. We've been asked about it for a long time now, and that's because it's already embedded in your business — it's already within the BI team and the wider business for all the quant data.

So it seems like the next step to just say, well, just add in the call data in a similar fashion and call it a day. You're probably thinking at this point, why did I sign up at all, this all sounds straightforward. Ultimately, though, we all know it's not quite as straightforward as that. As we always say at Wordnerds, if we could treat words like numbers it'd be really easy — but unfortunately you've got things like regional differences, sarcasm, nuances, loads and loads of misspellings.

So in this webinar we're going to unpack the specific challenges around doing that. We've worked with call data for seven years and figured out the framework to make reporting this verbatim in Power BI possible. And by we, I obviously mean the product and data boffins — I had absolutely nothing to do with this, just to make that clear, much to everyone's relief.

So, introducing the framework. Steph's put together this lovely three-step process to understand your customers in Power BI. You've got to start with the classification piece — getting everything into topics and working out what fits in what category. Then you've got to take that classification and translate it into a language that Power BI understands — that's the semantic model side of things. And then once you've done that, get it into an output you're comfortable with, that you can share through the business and dice as much as you want — that's the visualization piece, making it appealing for people in the business to understand.

So that hands me over to Steph. Steph's going to take you through the framework in a bit more detail — challenges, outcomes. Over to you, Steph.

Stephanie Clish: Thanks, Sarah. Hi everyone. So picking up where Sarah left off, let's dive straight into that first section of classification. And this is where it tends to start for a lot of the organisations that we see. You're doing a great job of gathering customer feedback, but you're left with a whole load of messy, unstructured text data and no good way of gleaning insights from it.

The goal of classification is to create a system where every single sentence you're receiving from your tenants is tagged with a sentiment score, as well as lots of categories that are meaningful to you. This gives us two outcomes: it makes the data easier to get insights from, and it structures it in a way that's more like quant data — which is good, because that's what Power BI is designed for.

As Sarah mentioned, we've been doing this for many years now, seven or eight years. We've spoken to hundreds of organisations and distilled what they want into three things. They want accurate classifications; ones that are transparent, so you can see the workings out; and they want them to be as broad and also as narrow as required, so they can perform deep dives and get to the root cause of their problems. Achieving these three things enables them to trust the classification process and be confident they're making good decisions.

We see housing associations go on a bit of a journey trying to find actionable insights from their qual data. They trial a bunch of different options, most of which end up falling short. It often starts with manually tagging. You definitely don't need me to tell you that this activity is a massive time sink and puts pressure on the team. Generally, unless you've got a large group of people doing this, you just can't keep up with the volumes of data coming in, meaning lots of it's left on the shelf, unclassified. Teams also massively reduce the number of tags they're applying so they can make it an achievable task, but this prevents the ability to do those deep dives. So they get super frustrated — they have so much rich data coming from customers, the customers are spending time giving them this data, and they feel like they can't even scratch the surface with it.

At the same time, we all have generative AI at our fingertips now, and we see lots of people wanting to use the likes of Copilot to surface insights from their qual data. We did a webinar on this a few weeks ago showcasing the major problems with this technique — it's on our website if you didn't catch it. But the TLDR version is that Gen AI are probabilistic systems, so they can't give insights teams the precision they need to confidently surface, size and prioritise their issues. More recently there's been a move from business intelligence tools to offer LLM technology solutions like text analysis, functioning straight out of the box. These can be relatively quick to set up, which makes them super tempting. But at their heart they're what we call black box AI, because they don't understand your context and the nuances between the themes you need, and they don't show you their workings out — so you can't understand and therefore trust how well they're classifying your data.

Hopefully that gives you an understanding of the potential pitfalls. In our experience, to do this well — and I'm a little bit biased — a specialist tool is definitely the best way. But with LLM technology being more accessible these days, we often hear, well, can we just build it ourselves? And the answer is, depending on the tech team you've got, possibly. But at Wordnerds we spend every day building and maintaining a platform solving these sorts of problems — we've got something like 142 different technologies and a team of 30 people that help us do that. It's a full-time job, and it would take a bunch of resourcing and upkeep which might not be suitable and could take your teams off track.

Whichever specialist tool you go with, once you're classifying the data there's just one more step before you can start getting it into Power BI. And I'm going to explain it with this guy. Does anyone know who he is? Pop it in the chat — bonus point if you can get his name. Okay, we've got some people with good film taste. This is Doug from the film Up. Doug is a dog. Dogs can't speak English, but when Doug wears his magic collar it translates dog-speak so humans can understand him. And this is exactly what we need to do for Power BI. If you're using a specialist tool for classification, the data is structured and optimised in a way that's suitable for that platform. The semantic model translates the data from your classification tool into a structure that's more optimised for reporting and analysis in Power BI.

There are three challenges to building that semantic model. First, you need expertise in the classification platform's data structure — and often, if you're doing this in-house, the team building the semantic model aren't the end users of the analysis platform, which makes it a little tricky to learn. Second, we're hearing from lots of organisations that more than ever there's a huge backlog for BI development time, so it might be a long time coming. And finally, if you have different tools classifying different sources of customer feedback, that might require multiple models to be built.

Once you've built this, that set of phases is done and you can move on to the exciting stuff: bringing all of that quant data in with your qual into one place so you can find, prioritise and share your insights through the business — most importantly, using a tool that most of the departments in the business are familiar with. Hopefully that's been helpful. I'll pass back to Sarah now, and if you've got any questions, drop them in the chat and we'll pick them up in the Q&A.

Sarah Wilson: Amazing, thanks Steph. I love that analogy — I think it's fab. Thank you so much for explaining that so succinctly. We're going to hand over to Steve now for the dashboard development side of things and the reporting. I've already teased that it's a great session, Steve, so no pressure at all. Over to you.

Steve Erdal: Thank you so much, Sarah. I made no such promise — I want to make that very clear from the start. And thank you, Steph, that was really fascinating. So you've got your data set up, you've classified it as Steph described so you can group it, and you've translated it into a version that Power BI can understand. You still have a big job ahead of you: how do you then turn that into insights that will allow you and your wider team to actually improve customers' lives — what we tend to call actionable insight? What we're going to do now is look at how we take that foundation and create the actionable insights on top of it. There is gold in your customer feedback data.

The next step after you've got your data in a state Power BI can understand is to start digging down layer by layer to find that actionable insight. First, though, you have to understand the jobs to be done. Having the right question at the top is a bit like knowing where to dig, and those questions vary. In this session we're going to be looking at the first couple — what affects TSMs, and what are the key numbers your senior leadership team need? But you might be looking at the root cause of complaints being escalated to the ombudsman, or pulling out positive things to take to the team, because the feedback we get tends to be more on the constructive-criticism side and positive stuff can be really great for the team. And if you're in insights you'll know the old question of "tell me something I don't know." These are areas that help you know where to drill down to find that actionable insight.

A few other things to consider before we dive into the demo. Think about how you get the most impact from the fewest possible visualizations in BI. If you're like me, you get really into visualizations and love what BI can do — but lots of visualizations can make people who are less comfortable feel overwhelmed. How can you give them the most value with the least number of visualizations? Are you planning to democratise your data — curating it and passing it out to the relevant people, or do you want your frontline teams, management and senior team following their own curiosity through the data? Understanding that is really important, because it changes how you present it. And finally, how do you impose a narrative on this? I got into data because I loved stories. You'll have heard the concept of data stories — it's so important to think about story structure, a beginning, middle and end to the points you're trying to make, because that's the best way to properly chime with people and help them want to act on it.

So with those things in mind, I'm now going to turn to the demo. Always a minorly nerve-wracking moment. As I share my screen — this is synthetic data, created by an AI, so forgive it if the actual verbiage is a little stilted at times. Obviously we couldn't show any actual housing association's data. Right, to the homepage. The first thing to spot is where we're getting into those big numbers for senior teams — volume, sentiment, your overall TSM score. There are obviously challenges here with this particular organisation. We've got that across the different levels of sentiment and over time. On this side we have those classifications Steph was talking about, broken down into different groups of people suffering the same issue. That allows me to compare these by volume, to look at them by sentiment — I can see what people really like, helpful staff, all the way down to what people are really struggling with. If I spot something interesting I can click on it. So damp and mould is looking quite high.

Click on that, and I can see what the TSM score is when you're just looking at damp and mould, what the sentiment is, how that's changing over time. There's clearly been some kind of damp-and-mould-based incident in July. Sorry, Sarah, do you have your hand up?

Sarah Wilson: So sorry, Steve, we just had a couple of questions — it might be helpful to contextualise before we go through the rest of the data. We had a comment that we can't see his screen — can everyone else see it OK? Yeah, OK, great. So just the questions, Steve, sorry to interrupt you mid-flow. Someone was asking about the flow of data: where's the data gone now that it's in Power BI? And also, can you explain the sentiment scale before we move forward — is it out of 100, does it go to minus, how does the sentiment score work?

Steve Erdal: Brilliant, thank you — two great questions. I'll do the second one first. The score is out of 100, so 50 would be average. We break it down into five steps of sentiment, from very negative to very positive, using a large language model. We then average that across all the comments, because some people might be saying positive and negative things within the same comment, and then we average that across the whole data set. So this score here, focusing on damp and mould, is a very low score. The score goes from zero to 100, but it tends to be something of a bell curve around the 50 mark, so we tend to say over 50 is good, under 50 needs work.

In terms of the journey the data has gone on — another great question, Tim. The data has been uploaded to the Wordnerds platform, where that classification piece Steph talked about has been done. The enriched data is then passed through — we use Amazon Redshift — through an aggregator, to allow you to pass it into the semantic model Steph just talked about. The semantic model has then presented this onto Power BI. Hopefully that makes sense — but it's still the same data, just enriched now with the classification Steph showed us earlier.

Sarah Wilson: Really helpful, thank you.

Steve Erdal: Not at all — please do keep those questions coming, and sorry for not spotting them. So this is our top level, our homepage. We've got our big numbers, and hopefully you can already see we're starting to pull layers away. We knew nothing about this housing association 20 seconds ago; we now have a few potential suspects, a few hypotheses we can go and explore in the data.

After this step we tend to look at drivers of TSM — what can we do to improve our TSM scores? The great thing about the classification Steph showed you is that you can turn it into hierarchies, into groups, which lets you look at specific challenges. These are designed to mirror the TSM questions you have to answer — maintenance and repairs, safety, communication, complaint handling, estates. We can now see where our biggest challenges are with regard to TSM areas. Maintenance and repairs, for example, might jump out. I can then see what themes within that data are creating the number at the top — the specific things people are talking about. If something jumps out, for example "the problem is getting worse", people talking about deteriorating issues, I can click on that and it shows me all the themes that overlap with it. And again, our old friend damp and mould is coming up as something where people are talking about it worsening. We've just pulled off another layer.

Sometimes you might want to turn that into a journey process, particularly with things like repairs where there's a linear progression — from making an appointment, through waiting for it, the repair itself, the conduct of the operative, the overall quality of the work, and then follow-up after. Here the y-axis is our sentiment, so we can see it going down and up through different parts of the process, and the volume is represented by the size of the bubble. If I see something interesting I can click on it, and it shows what the key issues are. There's an issue with incomplete work while people are waiting, and doors in particular seem to be an area where customers feel they're waiting a long time for an appointment. Again, pulling back another layer, trying to get to that point where we're finding actionable stuff.

The final level we tend to use is called a cross-table, and this brings us to Aldwin's question, provided before the event, about how you overlap your quant data with this kind of stuff. You'll have loads of metadata about your customers that maybe you don't want to pass in through this system. So here we've overlapped different categories and themes — we've got fixtures and fittings down this side, and then whether the customer felt the contractor was reliable or unreliable. You could use this to look at area, type of property, demographic information, tenure, length of tenure — our customers use all these to add richness and value. We're looking at the intersections between different challenges to get a sense of where to start. In this case you might look at just the contractors customers are finding unreliable. What this gives us is a list of which fixtures and fittings customers are most likely to call unreliable, and as a proportion of overall data in this issue, hot water comes top. So customers are most likely to talk about unreliability of contractor when hot water is involved. That's something you can go and action — talk to your contractor in that area about that issue.

We can continue drilling down to the final level, which is the actual customer. We're looking here at where customers feel people aren't doing what they say, and you can see we've got right down now into the specific verbatim from customers. This is synthetic data, to reiterate. But getting that process from the very top — the big numbers the senior leadership team need to know, tracking them over time — down into the TSM categories, then into the customer journey, finding those intersections, and finally getting into what a customer actually said. That's generally how we set up our dashboards, thinking about that through-line, that narrative, going from the very top right down to an individual customer with this problem. I'm going to hop back onto the deck now.

So, congratulations — we have a dashboard. This is highly exciting news for any data person. The question now is, so what? That dashboard alone will not make things happen for your customers. We think of the top levels of the dashboard as being about hypothesising — the way we found damp and mould earlier, finding potential areas you might want to explore. The cross-table and things like that are more about understanding, finding the key issues at the next level down. You still then need to deliver that to your customers, whether you present it in the dashboard, write reports, or present to your senior leadership team.

There are a few things you need to do to get that holy grail of actionable insight. Now, that's a phrase you've probably heard before — it's ubiquitous in our industry, it's become the name for the thing we produce, but people don't tend to do a very good job of defining it. So I hope you're ready to have your minds blown — I'm going to drop a real truth bomb. An actionable insight is an insight that you can action. I know, hold your applause. Flipping it on its head is quite useful: what makes an insight that you can't action?

For us there are three things. First, you can't action an insight if you don't know what the next step is — if the next step isn't obvious. "Customers don't like trains being late" is not an actionable insight, because there's nothing to be done off the back of it. We have a saying in our insight and innovation team at Wordnerds: keep on going until you find the action. That might be in terms of time and space, how things are changing or what area this is happening in; it might be different overlapping themes; it might be going back and looking at unsupervised data. But until you can say "and therefore we should do this", or "have you considered doing this", or "we should stop doing this" — until that next step is obvious — what you've got is not an actionable insight.

The second thing, and this one's fairly obvious, is that you can't action an insight if you can't measure it. If you can't measure an insight, you've got no way of prioritising which issue is biggest and most important, and when you go and try to fix it you've got no way of knowing whether you succeeded. So you need to be able to show: when I've intervened on this, this is what we expect to change. That might be volume — fewer customers describing an issue; it might be sentiment — people feeling happier about it; it might be a correlation with the TSM score, or with something like escalations to the ombudsman. Without that number, you've got no way of demonstrating the efficacy of the intervention.

Finally — and this is a harder one to talk about — you also can't action an insight if you don't care about it. We find this sometimes, not in housing — I don't think we've ever experienced it in housing — but in other industries we've found a really great insight where the bottom line was they didn't care enough to change it. There's a limited amount an insight professional can do there, but the key thing for me is to give the voice of the actual customer: find a verbatim that elucidates the point. In the same way you can't call an insight actionable if you can't measure it, if you don't have a real person who suffered this issue, I don't think you're providing the best possible level of understanding and persuasion to get people to buy into your narrative.

So you need an obvious next step, you need to be able to measure the issue, and you need to find an individual for whom this is a problem. When you've got all three — whether via the dashboard, a report, or a presentation to the SLT — you have an actionable insight. In our reporting it tends to look something like this, going back to our hot water example: we've got our numbers, showing the different issues; we've got a potential next step — we don't tend to order people "you must do this", we talk in terms of hypotheses, but we have an obvious next step; and finally we have the real customers for whom this has become a problem. That's the bit that makes people care — this is a real person I can help, whose life I can make better. That, in total, is what we'd call an actionable insight: it means you can intervene, you know what you expect to change, and you can understand how you're actually improving your customers' lives. Thank you so much for listening — I'll pass back to Sarah.

Sarah Wilson: Amazing, thank you so much, Steve — that was brilliant. My favourite part is going from the top numbers to then seeing the actual verbatim, and like you say, that data-stories piece, which we know is so important. So on to the juicy bit now. We've had a few questions in the chat — do keep them coming.

We're going to send across a full resource pack and a few things we mentioned in the chat — things like the Copilot webinar and a few different resources around Power BI, including a guide to the different models Steph talked about. You'll get that and the recording after the call. The thing we want most is some feedback, please. We've got a mixture of people on the call — people who want to build it and people who want to buy it, so Power BI analysts and also CX professionals who just want the insights yesterday. Whatever route you're taking, wherever you are on that journey, get in touch with me — we've been there, we've done it, we've made the mistakes, and we can have a chat about where you're at and what you're trying to achieve.

Now for the questions. I can see Steph's starting to sweat. One for Steve — this popped up when you were in the middle of your BI layers, another one from Aldwin, who likes to keep you busy: how many layers and themes would you generally attach to one piece of feedback?

Steve Erdal: Another great question. Within the TSM categories, I think we have 141 different themes that could be attached to a piece of feedback. And obviously, when we talk about pieces of feedback, some feedback is really long-form and goes through multiple points, and some is really short with just one key issue. So as many as it needs is the answer to how many are attached to an individual piece of feedback. Any new data that goes through the project will automatically be tagged into however many of those themes it should be in. That's something that, if you're manually tagging, you can't ask an individual person to go through 141 different aspects of the thing — but because we've trained these themes, it's possible. It's also worth saying that we allow our customers to train their own themes, so if there's a specific thing you care about that isn't in our theme bank, you can train up a small AI model to go through your data and tag into that aspect. So the sky's the limit. That's the level we've found useful with something as big and complex as TSMs — in housing there are so many different types of issue, and things that can go wrong or indeed right, that we need that level.

Sarah Wilson: Really helpful, thank you. To be clear as well, when Steve's talking about the theme bank, that isn't just an out-of-the-box solution. That's something we've developed through the TSMs ourselves. We sat down at the start, when the TSMs came out, and worked out the different categories and trained those ourselves, so it has that human-in-the-loop interaction. And then, like Steve says, we have that ready-made theme bank based on the TSMs — but we know you're going to have your own questions specific to your region, your geography, your questions, and you can tweak it however you'd like and have your own themes as well.

Right, let me go back down the questions. I think this is one for Steph: with the rise of LLMs, are you thinking about using them in Wordnerds to create the action plans, or at least suggest them off the back of the insights from all the customer feedback?

Stephanie Clish: That's a good question, Andrew. We've got lots of threads we're pulling in terms of how we can use upcoming technology to help insights teams and speed everything up — we know everyone's time-poor. But the key thing, which Sarah and Steve mentioned before, is having that context and having a human actually know which things are the right things to do. So for us it's about working with technology, but having that transparency and being able to allow our users to choose the right things.

We're also trying to find new ways of surfacing better insights. Once you've got those first two stages set up and everything on Power BI, we're now doing some discovery with prediction — trying to pinpoint, for example in housing, which complaints might be likely to be escalated to the ombudsman, or which tenancies are likely to become unsuccessful. So we're testing things like that at the minute. There are lots of threads to pull — I could talk a lot more about this, but I don't want to take up any of the time. Hopefully that's answered your question; if not, write back and I'll try my best.

Sarah Wilson: The good thing about this topic is that there are so many sub-topics we can cover, and we are going to do a series of webinars, so don't worry about it being a catch-all. Another question, from Rory — a good one, something I've also noticed, recently back from mat leave: we recently migrated our context data within Wordnerds to the new version and found it's a lot more accurate. How was this achieved? We've got a guy responsible for this in the data science team, but I'm sure Steph or Steve can give a fuller answer. Who wants to take it?

Stephanie Clish: Yeah, I'll take that. One of the most difficult things — maintaining and making sure the platform is working to its best potential — is understanding which technologies to use alongside the bespoke things we've got in place. This was just a case of us building some new technology to make context themes more accurate. We ended up switching out old context themes to that new technology. Steve has a lot more experience of building themes, but they seem to let you do a lot more in terms of what you're trying to classify now — lots of behavioural stuff and things like that.

Sarah Wilson: My favourite part is the accuracy check-in. I'm sure a few of you on the call have seen it. Before, when you trained a theme, you weren't really sure; now you're confident you're going in the right direction. It tells you when you've made enough effort, when you've had enough tries. It's brilliant. Really good question, thank you. One from Jack — local lad: how would I connect our qual data with something like you guys showed, the technicalities around it?

Stephanie Clish: It depends where your qual data is, Jack — I'm not sure whether you're a customer already or not. The idea is you find something to classify it; a lot of our customers use us, and then you go through those stages in the framework that enable you to get it into Power BI to start building those dashboards. If your qual data is a bit everywhere — you've got different survey tools doing different things — it's helpful to pull all of that into one place and use one tool to classify it, so you've got that consistency and can start to compare issues on an even keel. I hope that's answered your question; if not, let me know.

Sarah Wilson: We've got a question from a client — hey, Charlotte, Plymouth, I'm hoping it's you. Is it possible to incorporate platform data into our own data warehouse so we can combine it with service-performance data to show a more rounded picture of how we're doing?

Stephanie Clish: Hi Charlotte. Yeah, it definitely is. The benefit of us getting the data out into the likes of Redshift is that you can connect to that and pull the data and basically send it anywhere you want. We do have some customers who want to put all of this back into one data lake so they can bring it all together and make it more useful for them. That's definitely possible — give us a shout and we can talk you through how to do that.

Sarah Wilson: Fab, and there are a couple more. Would the end-product Power BI report reside in the customer's Microsoft tenant? If a change is needed to it, would we contact you via a help desk?

Stephanie Clish: That's a good question, Stephen. We're in a bit of a beta release with this in terms of customer access. What we're doing at the moment is one user from your organisation would connect to the Redshift data lake, we'd send you the PBIX file which contains that semantic model, and then that user would be in charge of refreshing the data. It would essentially live in your Microsoft account, so you can share those dashboards and reports more widely. And to answer your second part: in terms of changes, it would be one of those manual things, because the PBIX file would sit with you — so we'd be able to make changes with you.

Sarah Wilson: Fab, thanks Steph. And then Steve, I think this could be a good one for you — last question: do you find transactional surveys give better actionable insights than the TSM surveys?

Steve Erdal: Okay, so this is a wonderful and multifaceted question. With TSMs there are specific things you have to ask, and a specific order to ask them in; with transactional surveys you have a bit more leeway, a bit more freedom. We tend to find that, because it's fresher in people's minds in a transactional situation, that can lead to more depth.

What I would say, though — and this is probably my number-one takeaway for anything we do at Wordnerds — is that people are weird. They'll give you really interesting advice about your communication channels in transactional surveys; they'll give you really positive stuff about your staff in the middle of complaints. So while looking at individual surveys is really important and valuable, it can also be useful to have oversight of all this data together, grouping by those classifications, so that everybody talking about, say, damp and mould — whether in a transactional survey, a TSM, a complaint, an email or a call — is grouped together. Often these groups are siloed in housing associations: one team looking at transactional surveys, one at TSMs, and so on. If you can unite those things, everybody gets more value, because whatever you care about, customers will be using all these different sources and channels of contact, and the gold might be in any one of those groups. So there are pros and cons to both. There's probably more depth in transactional surveys, but the key thing for me is to unite these data sets, so that however a customer chooses to express themselves and whatever channel they use, the right team gets access to it and understands the issue.

Sarah Wilson: Thanks so much, Steve, that's really interesting. I think we're leaning into the complaints section now, aren't we — expressions of dissatisfaction, however made. We do tend to find that transactional surveys are on the whole more positive than the TSM, because the TSM is based on someone's perception that's really hard to change, whereas transactional surveys are about a specific experience they've had recently. Really interesting.

I think that's us — we've answered them all, we've really flown through them. They were all brilliant, very thought-provoking. So it looks like that's it. We've finished five minutes early, so you can go and grab your sandwiches. I'd really appreciate feedback, so please email me directly or give me a call — as you can probably tell, I absolutely love to chat. Let me know your feedback, any questions you've got, or anything you'd like to see next, because we are running a series of these, so look out for the next invite. Thanks so much, and thanks Steph and Steve — you did an amazing job. Bye everyone.