Dirty data doesn't fix itself

About this episode

In this episode of Behind the Data, host Matthew Stibbe talks with Susan Walsh, founder and MD of The Classification Guru and author of Between the Spreadsheets: Classifying and Fixing Dirty Data. Susan explains how a failed women's clothes shop pushed her into a data classification job she had never done before, and how she turned that skill into a business nine years ago. The conversation covers why data problems are really people problems, where AI genuinely helps with classification and where it quietly invents what doesn't exist, why "miscellaneous" is the worst category in any taxonomy, and her COAT framework for keeping data consistent, organized, accurate and trustworthy.

Meet the guest: Susan Walsh

Susan Walsh is the founder and managing director of The Classification Guru, a data classification and taxonomy consultancy she set up nine years ago. She graduated with a degree in commerce and worked in merchandising, telesales and account management at companies including Colgate, Philips and SC Johnson, before opening and closing a women's clothing shop in Guildford. She then spent five years in spend data classification, building and managing a team, before starting her own business. Susan is the author of Between the Spreadsheets: Classifying and Fixing Dirty Data and Optimizing Sales and Marketing Data, creator of the COAT framework for data quality, and an instructor on LinkedIn Learning, where her courses use real, messy datasets rather than sanitized ones. Her company has also built Samification, a tool that standardizes and normalizes company names against a master list of around 700,000.

Key takeaways

• Data problems are people problems. Whoever is entering the information is usually the source of the mess, whether that's your own team or the general public. Simple input rules, like blocking text in a phone number field, prevent more errors than any amount of downstream cleanup.
• "Miscellaneous" is the worst category you can offer. Everything fits somewhere. If it's high value, it deserves its own category, and if it's a one-off, put it in the nearest sensible bucket.
• AI finds patterns, not meaning. Given Walsh Hairdressers, Walsh Lawyers and Walsh Cleaners, it will invent "Walsh Group". Given a supplier called Walsh Lodge, it could file a dental practice under travel. The pattern matching is the feature and the failure mode at the same time.
• A confidence score is not a correctness score. Susan has seen 100% confidence on answers that were wrong. Her team exports every AI-normalized name and checks it manually, because that is still the only way to guarantee accuracy.
• Put a COAT on your data: consistent, organized, accurate, trustworthy. Consistent processes and naming conventions, documented. Categorized so things are findable. An agreed definition of what accurate means within the team. Get those three and trust follows. And keep the coat on all year: monthly or quarterly maintenance, not a one-off clean.

Chapters

• 00:54 From paint merchandiser to a clothes shop in Guildford that nobody came into
• 03:12 The data classification job she had never done before, and was immediately good at
• 05:37 What five years in the trenches teaches you that a data course doesn't
• 06:30 "Your code might be right, but the output isn't" — why classification is subjective
• 07:06 Building GPT agents to clean data, with a caveat up front
• 07:59 Samification, 700,000 company names, and the AI that invented "Walsh Group"
• 09:26 Why she won't use fuzzy matching or confidence scores
• 10:41 Where AI is genuinely fast, and where you still need a human
• 12:11 Why people won't pay to clean data
• 13:01 Day one on a project: what you standardize before you deduplicate
• 15:01 The company with 500 suppliers for stationery
• 19:00 Building a custom taxonomy, level by level
• 21:22 Is "miscellaneous" ever acceptable? Spoiler: no
• 25:18 The interview test file, and the guy who classified LinkedIn as a restaurant
• 28:55 100% confidence and completely wrong
• 29:39 The COAT framework: put a coat on your data, and keep it on all year

Episode transcript

Matthew Stibbe (00:01) Welcome to Behind the Data with CloverDX. I'm your host, Matthew Stibbe, and today I'm talking to Susan Walsh, and I'm delighted to have you on the show. She is the founder and MD of The Classification Guru, and she's the author of two books, available from all good bookshops and Amazon. One is called Between the Spreadsheets: Classifying and Fixing Dirty Data, and the other is Optimizing Sales and Marketing Data, which is a book I should be putting on my Christmas list. Great to have you on the show, Susan.

Susan Walsh (00:32) Thanks so much for having me. It's a pleasure.

Matthew Stibbe (00:35) So we were chatting before about air traffic management controller interviews and I'm interested in how you got into data management. Tell me a little bit about your journey, because I think you must be the only person we've had on this show who came in through retail.

Susan Walsh (00:54) Yeah, and even before that. I mean, to say I fell into it by complete accident really is the truth. I graduated with a degree in commerce, so a business degree. I was a paint merchandiser, then I was in telesales, I worked in account management. At places like Colgate, Philips, SC Johnson. And then at some point, you know, like in my 20s, I was gonna be a sales director, that was my dream. And then at some point I was like, I don't enjoy this, it's not the relationship kind of sales that I thought I was going to be doing. So, what did I do instead? I decided to go completely left field and open a women's clothes shop here in Guildford, where I've lived for over 20 years. It's an affluent area. I had done some research. This was, yeah, 15, 16 years ago when women's workwear was changing in the office from, you know, suits and blouses to more casual. But there weren't many places that you could buy nice outfits and online shopping hadn't really taken off. So I was like, I'm going to open this shop.

Beautiful clothes. I had styled the whole shop and decorated it myself. It was lovely. It was mid-range pricing. So it wasn't, you know, unaffordable or unachievable. And nobody came in. Nobody bought anything. Because what I hadn't anticipated was the brand snobbery that exists in Guildford. So had I had Chanel and Dior, I probably would still be in business today.

But because it was mid-range, I was an unknown, people were very wary. They would kind of eye the shop but not come in. And so I managed to hang on for about seven, eight months, and then I was literally at the point where I was so broke I couldn't afford to file for bankruptcy. I had to save up for six months to file for bankruptcy. And I needed a job and I needed a job fast because I had nothing to pay my bills.

Susan Walsh (03:12) And I went online and I found a data classification job. Had never done it before in my life, but I had worked in large organizations, I knew what they were buying, so I felt like it was something that I could do. Started there. Turns out I was really good at it and really enjoyed it. What started as a part-time job six months later became a full-time role managing a team.

And I ended up staying there for five years, building up a team of 14 people, training them, managing them, managing the projects, and basically started from the bottom and just learned on the job, various different aspects of spend data classification, working with procurement data, working in a tool, doing a little bit of data modeling.

And that company then got bought by a large organization and it pretty much changed overnight and it just... I wasn't happy. And so I was like, okay, I'm not actually a data analyst, I don't code, but I know I'm very good at this one thing. So my options were to either, you know, go and start somewhere at the bottom from scratch again, which I couldn't really afford to do, or set up another business, which was a bit of a risk, but lots less overheads when you work from home rather than having a shop. And that was nine years ago and that is when The Classification Guru was born. The name is because we research so many companies and it's very hard to tell what they do by the name or their website. And so I wanted to lay out exactly what I did in my business.

Matthew Stibbe (04:57) Actually what you do, on the tin.

Susan Walsh (05:04) Yeah. Yep. So you know, people who know, know. And you know, that in itself was a challenge because I'd always been behind the scenes. I had no network, so I didn't know anybody in procurement or data. I just put myself out there, I went to a trade event, exhibited, and really just started growing the business from there.

Matthew Stibbe (05:24) There was a book many years ago called What They Don't Teach You at Harvard Business School. And I'm curious what you learn from five years in the trenches with spreadsheets that they don't teach you at university for data engineering courses.

Susan Walsh (05:37) Well, I can tell you now for a start, you are working with sanitized data on a course. Unless you're doing one of my courses. So I've just had one LinkedIn Learning course come out and I've got a couple more coming. You get real-life messy, dirty data sets to work on with them. So I think that's it. I think there's a lot of theory but maybe not so much practical. They maybe don't teach you like critical thinking or even like business soft skills, communication, communicating with the business. I mean, I can't tell you the number of arguments I've had with data scientists, analysts who are like, well, my code is right. And I'm like, your code might be right, but the output in what I'm seeing is not right. So something's wrong somewhere. And so, you know, all those little things.

Matthew Stibbe (06:27) The world doesn't fit my code's schema.

Susan Walsh (06:30) Yeah, and it's recognizing that people think differently too. I mean, data people are very black and white and so they are not very good at the kind of work that we do, which is spend data classification, because it's subjective, there's a lot of context involved, and there's often more than one right answer. So you just need to be consistent and you really need to... it's not a yes or no, and a lot of data people really struggle with that because they are very black and white. So yeah, I learned a lot.

Matthew Stibbe (07:02) And what are you geeking out about at the moment?

Susan Walsh (07:06) So I am about to film my third LinkedIn Learning course next week, and it's going to be all around ChatGPT work. So I am going to be geeking out on that this week and building some agents to help me clean data. And I'm going to caveat that straight away up front with the point that you cannot just use agents and build agents to clean your data. You still need to know and understand your data first. Know what's right and wrong, know what clean is and what's not, and check what you're getting back from GPT. It should be used as an aid and an assistance, not like it's gonna do everything for me.

Matthew Stibbe (07:51) When people bring dirty data to an AI, what's the thing that breaks first?

Susan Walsh (07:59) So I can tell you we've built a tool called Samification and what it does is it standardizes and normalizes company names. So we see Dell, Dell Computers, Dell Inc., Dell Software, PwC, P.W.C., PricewaterhouseCoopers. And we've got a master list of like 700,000 names now. So anything that matches that gets normalized straight away. Anything that doesn't goes through our AI and our GPT and it's learning from our master list.

But it's still not right. It's looking for patterns and commonalities. So what we've found is that say you had Walsh Hairdressers, Walsh Lawyers, Walsh Cleaners, it actually normalized it to Walsh Group. We didn't tell it to do that, it just saw the name Walsh and decided that it was all the same thing. That's an AI thing.

Matthew Stibbe (08:51) And that was an AI sort of being a bit too clever for its own good.

Susan Walsh (08:58) And even more recently, we have started normalizing BT, British Telecom, to BT Group because that's now their name. So we have that in our master list and there's various versions of BT. But even when a new version of BT comes in, it keeps it as BT, not BT Group. So there's a lot of work to be done, it's not magic.

Matthew Stibbe (09:18) How do you decide the threshold for that and sort of how do you encode or encapsulate rules around that?

Susan Walsh (09:26) So we have like some rules. So at the moment we're working on trying to change... you might get Walsh and Walsh or Walsh & Walsh. So we're trying to put some rules in to change space "and" space to space ampersand space. But again, it can go wrong, so you have to be really careful. So that's what we're working on. But honestly, we don't do confidence scores and things like that. We don't do fuzzy matching, which is like, it almost looks the same so we'll match it. I'm not into that. I'm into accuracy.

Matthew Stibbe (09:59) Because that's sort of the standard engineering data science answer to this, isn't it?

Susan Walsh (10:22) So what will happen is whatever it normalizes, the AI normalizes the name as, when it's exported from Samification, our team then manually goes and checks and cleans it. So that's the best guaranteed way that you can get it right, is to have someone check it afterwards.

Matthew Stibbe (10:24) What's the balance in your experience? I'm talking in general terms, between machines can do this programmatically, machines can do this with AI, and human beings need to come and validate, check or do it.

Susan Walsh (10:41) And I think this is where we have to stop talking about data as one blanket thing, right? Because there are different types of data where AI is really good. So for example, parts and part descriptions and part numbers. AI can go and find and classify that really quickly. Probably faster than a human. But if there's a part description and there's abbreviations and it's not clear, then you're gonna need a lot more human involvement. It's not gonna be a magic solution. But again, once you start building a clean database of this kind of information, AI should be able to learn from it so that you then can minimize these things going forward. It's different for my business because each time we work with a client, it's a different industry, it's a different type of data, different but similar issues.

If you're building it for your own business data, you're gonna see the same things each time. And so it's gonna be more effective than say if we tried to do something.

Matthew Stibbe (11:42) I mean, the thing that strikes me with marketing data is you look at say a list of... HubSpot will say, look, of the 40,000 people in your contact database, these 1,500 are probably duplicates. And then it goes, right, that's your problem to solve. And you have to go through and everybody just, you know, it requires a little bit of domain knowledge, a little bit of database, who are those people? And it just suddenly becomes hard.

Susan Walsh (12:11) Yeah, yeah, it is. And then people don't do it because it is hard. And also honestly, there is definitely... I mean I think it's changing slightly with AI, but over the years I've seen that people don't want to pay to clean data. They think it's something they can do themselves or their team can do. And of course everybody can clean data, but when you do it day in, day out, you've got the right tools, you can do it far quicker, far more efficiently and far more accurately and it can get you there a lot faster.

Matthew Stibbe (12:42) For your business, what is the sort of library of, like you've talked about having a database of companies and some rules about company deduplication. And that's sort of interesting. What do you arrive with on day one versus what do you have to build uniquely for each company?

Susan Walsh (13:01) So, day one can be multiple different file sources, and then we need to standardize the column names to pull all that data together. That's when we would then normalize all the names because that's when you can start to identify potential duplicates. But we would never just deduplicate on a name, for example. We would also look at postcodes, emails, phone numbers, other reference points to look for. I mean, we'll even check things like you might have one email address but two different addresses, postal addresses or names. So, you know, you wanna be able to check and flag all those kinds of things too. And if you've never done that before, then you're not gonna know to look for it. So, yes, you might be able to clean it yourself basically in a short space of time, but, you know, we're looking for things that you don't even know exist.

Matthew Stibbe (13:55) Because you've done it hundreds of times, somebody doing it in-house is like doing it for the first time.

Susan Walsh (13:58) Yeah. And you know, the principles are pretty much the same regardless of whether it's procurement data, finance data, sales data, marketing data. You know, it's the same kind of issues. Ultimately data problems are people problems, right? And it's whoever's putting that information in that's causing the problem. And sometimes that can be internally, it could be your team, sometimes it's the general public, which you then have very little control over. But you can put rules in place, you know, like you can't put text in a phone number column or things like that. You know, there's little things you can do to help minimize those errors.

Matthew Stibbe (14:43) I'm fascinated to explore like a specific project and lessons that you've learned from it. And you talked a little bit before we hit the record button about procurement. Can you give me... illustrate some of the common patterns from a specific project that you worked on?

Susan Walsh (15:01) So one of the things that we see a lot of in the companies that approach us is they don't have visibility on what they're spending in their organization. They have multiple countries, multiple systems, and they don't even know the basics. The issues are that they can't pull that data together. Sometimes they even struggle to get the data from within their own organization, and then they don't have the time or the resources or the tools to do it.

So we will take that for them and we will start by normalizing all their suppliers and then we will start to classify their data, and it will depend on the amount of detail in their description, invoice description. So if you go to Ryman's, for example, and you get an invoice, it might just say office supplies. So you can't really go into too much detail with classification at that point, but your invoice might say notepads, pens, sticky notes, folders. And at that point, if your business needs it, we can go into that level of detail of categorization within like a three or four-level taxonomy. So it's about, actually, the biggest thing I've learned is asking the customers what their problems are, because quite often they'll tell you they need something, but when you ask the question and find out what the problems are, they might actually need something else. And so it's about listening to what they're telling you and then saying, right, this is actually what you need. And then we'll classify, and we prefer to build a customized taxonomy based on the client's data. There are off-the-shelf taxonomies out there like the UNSPSC or ProClass or ProKG, but they're very generic.

And might not give you everything you need. And so we'll advocate for something customized if possible. And then we can give that back. They can then do analytics on that. It's so rewarding when you can go to them and they're like, I had no idea I was spending X million on office supplies or IT, or I had no idea that we had 500 suppliers for stationery. Like we don't need that many suppliers. And that happens.

And so then there's cost saving opportunities, there's supplier rationalization opportunities. We can look at, you know, I was telling you we have a client and they had a number of invoices, and the average price cost of an invoice was 23 pounds. Now processing an invoice costs more than 23 pounds. You know, there's an opportunity there for whoever's ordering their stationery to consolidate and maybe order once a month or once a quarter. You know, you're saving on shipping costs, processing fees, someone's time, you know, entering that information or chasing up or, you know, all those little things.

Matthew Stibbe (18:04) And who doesn't like a well-stocked stationery cupboard, right?

Susan Walsh (18:07) Right, I know. I mean, even I'm here and I have a ridiculous amount of pens, I really do. I don't need all these pens. I do like my little crown though. But I never use it. You know, it's crazy. But it just brings some order and visibility. And again, of course teams can do this themselves. But we see wider than just their organization. They have like a single focus on the data that they work with. We see across industries and companies and so we have experience of looking for more things.

Matthew Stibbe (18:49) How do you go about building a custom taxonomy for a client? How do you balance the too broad versus too narrow debate?

Susan Walsh (19:00) Well, we'll ask what they want first of all. Not every customer wants something detailed. Sometimes they just want two or three levels. They don't want detail. So we'll ask that first. So that should be things like IT, HR, travel, professional services, marketing, facilities, maybe logistics if you're a manufacturer, MRO, which is maintenance, repair and ops. So you start very top level and just get them into the right categories. Then level two, you'd start to go into a little bit more detail. So professional services and then maybe legal services. And then if you wanted a third level, it could be IP law, it could be business law, it could be employment law. Again, if you've got that information, you don't always.

We'll start looking at supplier names and invoice descriptions, doing keyword searches. So we'll be looking for lawyers, solicitors, law, or hotel, inn, lodge, all those kinds of things and that will give us clues in either the supplier name or the description about what it is. But you can get caught out too. So you can't just do a blanket keyword search and paste stuff. You still need to check.

Matthew Stibbe (20:26) You check back with the client, is this taxonomy...

Susan Walsh (20:28) No, no, no, just... I'm trying to think of an example that might be like, it could be Walsh Lodge, right? Which sounds like it's kind of travel, but actually it's a dental practice. So you just need to have a bit of common sense about you and not just blanket, you know. And that's again where you've got to be careful with AI because it is looking for those patterns and words, right? It sees a word and then most of that word is associated with travel. It's just gonna make it travel. You need that human aspect to be able to check it.

Matthew Stibbe (21:02) Yes, you need a certain amount of deterministic behavior for classification. You don't want AI to make the assumption about something...

Susan Walsh (21:11) No, I think having kind of worked in businesses really helps that. Any kind of admin or process role really brings a lot of value to this.

Matthew Stibbe (21:22) Little bit of common sense. One of the things that I've been grappling with personally, this may be a daft question, I don't know. When trying to classify things, trying to do that taxonomy thinking, the phrase that rattles around in my head, and I don't know where I got it from, but it's not mine, of MECE, mutually exclusive but comprehensively exhaustive, right? So your 15 top level things, you should be able to cover every single kind of purchasing in one of those 15 buckets, right? That would be, it fits into one of them and the 15 cover all of them. Fine, mutually exclusive, comprehensively exhaustive. And yet in the real world, when I'm trying to do this kind of thing, there's always the doesn't quite fit. Is category 16 going miscellaneous? Is that ever acceptable? Or is it if you have like law and HR and this thing and this thing and this thing and then there's some miscellaneous law.

Susan Walsh (22:25) No, no, no, no, no, no. No, that is the worst thing because what happens is people will naturally try and go for the shortcut, right? So they'll just put everything under miscellaneous. You might only be able to classify to one level. So whatever that thing is, it will fit in one of those, right? Our kind of default, if we're really not sure, is professional services consultancy. It'll go in that. The other thing is if it's a high value or an important thing, it should have its own category somewhere. But if it's only occurring once and the value's like five pounds, don't worry about it. Just put it in the nearest thing. But please don't put it in just anything. You know, it has to relate to what it is.

Matthew Stibbe (23:13) Years and years ago, I had, I've always run my own business, and I had a bookkeeper, office admin, who used to enter purchases into the accounting software, and anything that she didn't recognize as, you know, stationery or this thing or that thing, she had the brilliant idea that she should put it into the director's loan account. At the end of the year, the auditors came in and they said, Matthew, you owe 8,500 pounds to the business for all this stuff. I'm like, what? We had to spend so much time sort of unpicking it. And auditors get very... back in the day when you had to have accounts audited, whatever your turnover, they get quite picky when you go, no, it wasn't a personal expense. They go, no, no, it probably was if it's been put in there. No, no, no, it really wasn't. It was just my office manager. So, okay, miscellaneous is not a good thing. I think we've just established that. Okay, good.

Susan Walsh (24:12) Think about marketing, right? You want to segment all your customers, your businesses, and everyone just starts putting in miscellaneous. Can't target anyone really, can you? Because it's all sitting... it's the same thing. People will go for the quickest, easiest route so they don't have to think about it. If you give them the lazy option, they will use it.

Matthew Stibbe (24:32) Yes, yes, sadly true. So we were talking just before we started this conversation about interviewing and people. And I was telling you the story about, I'm mentioning this now because I touched on it at the beginning, but how sometimes in air traffic management interviews, you go in and there'll be various things on the desk, piles of paper and pens and things. And one of the things they might look for is whether you sort of start lining them up neatly.

Susan Walsh (25:00) Yeah, that picture on the wall that's not quite right, that really annoys me, yeah.

Matthew Stibbe (25:03) Yeah. So when you're interviewing people, or when you're thinking about people who would be good at the job that you do, how do you judge that? How do you test that? How do you evaluate that behavior in an interview scenario?

Susan Walsh (25:18) I give them a test file and I put some particularly tricky things in there to try and catch them out. But by doing that I can see their approach, where they're coming from, what they've been thinking about. And it doesn't have to be right. I'm looking for, you've tried and there's some logic behind what you've done.

Susan Walsh (25:43) The great example that I use in some speaking is there was a guy and he's a data guy and he did some work for us and LinkedIn was the client or supplier, the description said restaurant, so he classified it as a restaurant because he hadn't looked at the data in context. We all know that LinkedIn is not a restaurant.

So you know, it's things like that that I'm looking out for. Are you using common sense, context and not just like trying to cheat? You know, you probably wrote some kind of rules or something, or maybe a bit of code to try and classify. And that's where like AI still gets caught out. If it's never seen that before and hasn't learned it, it won't know.

Matthew Stibbe (26:31) It's incredibly hard. We wrote an algorithm, an AI prompt that sits in a workflow for a client that reads incoming emails that are requests for quotes. And it's in a very complicated technical field. And so the emails are not information rich, not particularly well structured, but often have all the details in, and then the company has a sort of 12 page manual of working out how to assign requests to different salespeople. That stuff's just hard and you kind of need to have a... I'm interviewing myself.

Susan Walsh (27:07) No, no, I need to know now, you need to tell me.

Matthew Stibbe (27:25) Well, the bit that sort of ended up coming very right up to the surface for this was having the AI express a confidence level of how much information it's got. To say, the data that we've extracted from this, it's this kind of request or it's for this kind of thing, or it's got this kind of urgency. High, medium or low, the AI has no opinion. And it had to be able to justify that with something from the email. We think this is from this, it has this kind of, I don't want to be specific about who it is, this kind of request because the email said this, this, that, or the AI didn't answer it because it couldn't find anything.

Susan Walsh (27:50) Actually, we were helping someone QA some AI classification that they were testing and there were two lines and they weren't sorted in order. So one was up here on the list and one was down here. Same name, same description. The AI had given it two different classifications. Now both were right, because with classification you can have more than one right answer, but it was working line by line. We worked as a whole with context, right? So it didn't matter that it'd seen it before and classified it as something else. It started afresh. And so that's where we're still having issues. And you could have the same problems with that. That same exact query could come in on a different day and you might get a different result.

Matthew Stibbe (28:39) There's almost like a meta level to this kind of work where you're using AI where it kind of has to report on the things it wasn't sure about. So you can go and fine tune the prompts or the algorithm or the thing to like next time you get two answers that might be valid, here's how you choose the right one.

Susan Walsh (28:55) But I mean, I don't want to get too deep on this, but I mean, confidence based on what? What confidence level? I mean, I've seen a hundred percent confidence accurate and it's not accurate. So, you know, you still have to check it and make sure, because just because it's confident doesn't mean it's right.

Matthew Stibbe (29:14) That's a really good takeaway, I think, here. AI is extremely good at confidence. I mean, it can tell you some stuff with absolute conviction. And there it is, and it looks like, and yeah. OK, so we've got time for one more question. You have a methodology called COAT. Can you explain that for listeners who haven't encountered it before? Because I think it would be very useful.

Susan Walsh (29:39) Yeah. Like I said, data problems are people problems. Not everybody that is working with data is a data professional and it might just be that it's part of their job and they do it because they have to. So getting people to input the right data at the point of entry is so important for all the way down the line. But of course to non data people data's not exciting, it's boring, it's you know. So I created the COAT framework.

So, put a COAT on your data. Make sure it's consistent so everyone's using the same processes, the same naming conventions, terminology, like units of measure, date formats, those kinds of things are all really important. So get everybody doing it the same way and document it. Organize your data, so have it categorized so that when you need to look for something it's easy to find. Make sure that it's accurate, but define what that is within your team or, you know, people you're working with to make sure everyone knows what accurate means. And once that's consistent, organized, and accurate, it becomes trustworthy. And we always hear we don't trust the data. And so it means that you can make better business decisions, you can be confident in the decisions you're making as well, and know that you're making the best decisions. So that is the COAT framework. So just tell everybody to put their coat on their data all year round as well. You know, it's got to stay on all year round. So maintenance is just as important. You can't just clean your data and leave it. You have to, you know, maintain it and manage it.

Matthew Stibbe (31:17) That's painting the Forth Bridge, right, when you finish.

Susan Walsh (31:32) You wouldn't want your coat to get dirty, right? So your winter coat, you probably take it to the cleaners once a year. Got to do the same with your data. I mean I would say more than once a year for your data. Monthly or quarterly is normally my recommended frequency, but yeah, it's really important.

Matthew Stibbe (31:36) Fantastic. Well, that's a great thing for people to take away from this. Before we wrap up, Susan, tell us where people can find you online.

Susan Walsh (31:45) The best place to find me is LinkedIn. So Susan Walsh, The Classification Guru. TheClassificationGuru.com is my website and that's where you'll get all the information and resources that you need between those two.

Matthew Stibbe (31:58) Fantastic. It's been great having you on the show, Susan. I've really enjoyed our conversation. Thank you. And everybody listening, that brings this episode to a close. If you'd like to get more practical data insights and learn more about CloverDX, please visit cloverdx.com/behind-the-data. Thank you for listening and goodbye.

Susan Walsh (32:02) Yeah, thanks so much.

Resources

• The Classification Guru - Susan's data classification and taxonomy consultancy, with information and resources
• Susan Walsh on LinkedIn
• Between the Spreadsheets: Classifying and Fixing Dirty Data — Susan's first book
• Optimizing Sales and Marketing Data — publisher or Amazon page) — Susan's second book

Related content

• Master Data Management (MDM) Explained

• Why most data teams overestimate their operational maturity

• When to use LLMs vs SLMs for privacy and governance

More from Behind the Data

• Ethics, alignment & the human side of data with Mirela Mart, CFO at Articulate Marketing

• Automating the mundane, liberating the humans with Nick Liles, IT Manager, Sisk Fulfilment Service