So What? Marketing Analytics and Insights Live
airs every Thursday at 1 pm EST.
You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!
AI can update the outdated contacts in your CRM, so you don’t have to keep paying for expensive data lists. Set a few simple guardrails and your team can use the same tools to keep the data from becoming CRM decay. A clean database makes it easier to find the right buyers and close more deals.
Watch the video here:
Can’t see anything? Watch it on YouTube here.
In this episode you’ll learn:
- What CRM decay looks like and why it’s so bad
- The two components of CRM decay
- How to fix CRM decay with AI
Transcript:
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Katie Robbert – 00:00
Well, hey everyone. Happy Thursday. Welcome to So What, the marketing analytics and insights live show. I am Katie, joined by Chris and John. Howdy fellows.
Christopher Penn – 00:47
Hello.
Katie Robbert – 00:48
Hello. Hello. This week we are diving into B2B AI use cases, specifically CRM decay and data rot, which sounds just uncomfortable to say. When Chris and I were at the MarketingProfs B2B Forum a couple of weeks ago, we were talking through on the podcast, which you can get at TrustInsights.ai/TI podcast, the different topics that are still prevalent today for marketers. These are the things that they are struggling with but that do not make it onto the event schedule. They are not breakout sessions because they are not the new shiny thing.
When Chris and I were talking, it struck both of us that maybe the problem isn’t the new shiny thing, but the way that you can approach the solutioning is part of the new shiny thing, which as of today everybody knows is generative AI. How can we apply AI to fix CRM decay and data rot? Chris and I earlier on this week on the podcast were talking about what you can do internally with your own data. Chris, I think today we are talking about more external sources. Where would you like to start?
Christopher Penn – 02:02
You nailed it exactly, Katie. There are two forms of data rot. There is internal, which is, “Hey Chris, could you update the CRM today, anytime today, please, with the things that only live in your head?” That is very much a people and process issue. Katie, as you are fond of saying, new tech doesn’t solve old problems. That is the oldest of problems.
Katie Robbert – 02:26
It is true. New tech does not solve old problems.
Christopher Penn – 02:30
The other form of data rot is when people change jobs, leave, experience layoffs, or your CEO resigns to an affair. The data itself goes bad. You have a record for Christopher Penn that used to be at Shift Communications. Christopher Penn works at Trust Insights now and has for almost a decade. If your CRM doesn’t update that data, then you are making potentially bad decisions.
This is a platform and data problem. This is not just a people and process problem. In the past, there was a legion of companies that offered to sell you appended data at exorbitant fees. John, what are some of the data append things that you have seen people charge enormous sums of money for?
John Wall – 03:38
It used to be you could buy these lists for 10 or 20,000 dollars. They weren’t garbage years ago. Everything was so much more wide open in the past. You could actually get data like attendees from events or customer dumps from certain pieces of software. All that stuff has gone extinct.
Pretty much anybody trying to sell that stuff now is selling snake oil or complete lies and fantasies. It used to be you could buy your way out of trouble, especially in enterprise software, but that is no longer the case.
Christopher Penn – 04:18
Yet the data is out there, and you do not have to pay exorbitant sums of money for it. We do have to apply a big warning here. I don’t have a big warning sign. Maybe Katie, you can find the red flags.
Katie Robbert – 04:35
Ready?
Christopher Penn – 04:36
The big warning is this. Personally identifying information, like who you are and who you work for, may be restricted in terms of how you use it depending on where you do business. In Europe, for example, GDPR governs PII, and you may not be able to use even publicly available data legally in the EU or for people who are physically located in the EU under those restrictions. Please, we are going to talk about technology solutions today, but you have to consult with your legal counsel to make sure that you are compliant with all national, regional, and local laws.
Katie Robbert – 05:19
To be clear, consulting with legal about what to do correctly is not the red flag. The red flag is not doing those things.
Christopher Penn – 05:27
Exactly. It is saying, “I am just going to scrape a whole bunch of data from the internet because if it is out there, I can use it for free.”
Katie Robbert – 05:34
Yeah, don’t do that.
Christopher Penn – 05:37
Here is where AI can help. For anyone who has used Claude, ChatGPT, Microsoft Copilot, or Google Gemini, they can go and do stuff like search the web. It is pretty trivial to search the web for things that are public information, like who is the CEO of Trust Insights. Any search engine is going to say it is Katie Robbert.
You can ask, “What is Katie Robbert’s LinkedIn profile URL?” A search engine can go and find that information. Because it is publicly available, you can use AI to programmatically go out and get it and then pass that data back to your CRM. There are two ways you can do this. There is the extremely expensive, slow, but non-technical way.
This is where you fire up Claude, make sure your CRM connector is on, and ask it to go into our CRM one record at a time. It pulls it out, checks if it is valuable, and then updates it. That works great, or you can do it programmatically.
Katie Robbert – 06:53
Depending on how much data you have in your CRM, going record by record could be a really expensive way to approach it. I do not do it as programmatically as you might do it, Chris, using Claude code. I try to do things in bulk as much as I can by focusing on one area of the CRM at a time. I make sure that all of the closed-lost reasons are in there. I will pull a list of all of the records that have been closed-lost but do not have a reason.
I do not necessarily have Claude go back in and fill it out. That is where human judgment comes in. It can suggest a reason based on the notes, but I am going to punt it back to the team to ask what data belongs in here. Our team, John especially, is always going to have more context and nuance than Claude ever could.
Christopher Penn – 07:53
The programmatic approach is to do exactly this, but have your favorite AI tool build an app in the language of your choice to do it. You are going to need a few things. You are going to need a search engine. DuckDuckGo is probably the best one to use because it is private, they don’t retain search data, and they have very generous limits for AI agents.
Google, on the other hand, is not private. They have been aggressively blocking almost every kind of AI agent. You need a search engine, and you need a way to connect to your CRM. For pretty much every major CRM vendor out there, there are command-line tools that don’t use tokens to talk to your AI. As long as you are authorized by your martech team to use it, you can hand that tool to your AI agent of choice or to a Python script that can grab that information.
The third thing you need is some kind of AI to write the code and also a lightweight, cheap one to effectively look at things like whether this is even a customer we care about. You should have one or more ideal customer profiles. Part of this cleanup would be evaluating if this contact should even be in our CRM. If this person sells clown balloons, they are not going to be a good customer for Trust Insights.
Katie Robbert – 09:39
What is interesting about using the ideal customer profile is that is what is relevant today. Do we run the risk of deleting things that could be relevant tomorrow?
Christopher Penn – 10:03
You definitely run that risk. I do not delete anything, much to my hard drive’s detriment. You can mark things as a priority. John, as the sales team lead, might log into HubSpot one day and see 10 contacts marked high priority.
These 10 people are squarely in ideal customer profile two and have had 82 sessions on the website. They are as interested as they are going to get. Maybe John should reach out to these 10 people first, as opposed to the guys selling clown balloons.
Katie Robbert – 10:43
John, before we get into how we are going to do all of this, what is your thought on using AI to clean up a CRM? Do you see that as putting your job at risk, or are you relieved that it handles the admin stuff you never get to? I am curious about your thoughts on where we are headed with this.
John Wall – 11:03
I am completely on the side that everything AI can do is fantastic. The last thing I want to be doing is trying to figure out where some guy who talked to us three years ago is today and what he is doing, especially when we have already made a reasonable effort to find him. I have come to this weird point with CRM where the value is not the same as it was five years ago because things are moving so fast and changing so much. Unless you are in a little niche business where your customers are buying from you every month or more frequently, it is different.
If it is a huge organization with lots of people and opportunity, then you can go all in on CRM. In a situation like ours with a very narrow product and a fast-churning market regarding employment and technology, I can only really trust data that is about three months old. We have had cases where we were talking to somebody for four or five weeks, and then suddenly they were laid off or reshuffled. This happens in real time, so if AI can constantly be watching, tracking changes, and updating things, I am all in.
Christopher Penn – 12:23
Regardless of which approach you take, you need to have one or more ideal customer profiles written down. You should have your company, your products and services, and what is for sale written down. Part of CRM cleanup could involve leaving notes that a specific company or person would be a good candidate for the Trust Insights Academy. Maybe they do not have the budget for our consulting line but do have the budget for some of our courses.
The inverse could also be true. If a person bought a course, completed it, and scored 100, they might be a good candidate for having us do some enablement. If a person used our free AI view tool, maybe they are a candidate for using our AI Pro Consult pre-packaged audit. You want to have that stuff written down.
If you want to take the low-technology approach, you would have to know what is in your CRM. You could pull a list of contacts from our CRM that have had 25 or more web sessions in the last three months or companies that have had one or more download engagements and are not already in the pipeline. You could give it to Claude as a Markdown or YAML file, limiting it to 10 at a time because you will be doing a lot of updates. You could instruct a Python library like DDGS to web search these companies and people.
You would check the company size, their estimated annual revenue, make sure they are still in business, and check the company’s website URL. You would check the contact’s name, see if they are still in the same position, and find their LinkedIn profile URL. After that batch of 10, you would send it back to the CRM and update those 10 records.
Katie Robbert – 14:38
How is that different from some of the AI tools built into the CRM? HubSpot has Breeze, Salesforce has their version, and other CRMs have their versions. How is what you are describing different from what is built into these systems?
Christopher Penn – 14:56
They charge you more. They all have similar capabilities, and they are all trying to sell you an agent like Agentforce to do exactly this. This is a homebrew version that utilizes your existing Claude subscription so you are not paying more on top of what you are already paying.
Katie Robbert – 15:25
This goes back to the build-or-buy slide you often present. The pro of buying more from your CRM is that you do not have to build it. The con is that it costs you more, and you have less control over what you are getting back. If you are building it, the pro is you have control and it will not cost you more, but the con is you have to build it. As soon as you mentioned Python, I was hesitant.
Christopher Penn – 16:04
Here is where agent skills and Claude’s skill-building capabilities come in handy because it can run Python.
Katie Robbert – 16:14
Good.
Christopher Penn – 16:15
You would tell it to use a specific library in the Python interpreter, and it will build that into a skill. If you give the skill a schedule, you could check your top contacts every 30 days. One of the things we say in our 5P Framework keynote is that a skill plus a schedule plus a plan is an agent.
If you build the skill to refresh your CRM, put it on a schedule, and give it a simple prompt to run every Monday morning, it will do that. You will have a cleaning robot continuously fixing your CRM over time, constrained only by how much usage you buy from Anthropic.
Katie Robbert – 17:16
How do we do that? Where do we get started?
Christopher Penn – 17:25
We just said it.
Katie Robbert – 17:27
Okay, live stream over. Bye.
Christopher Penn – 17:29
Let’s talk about the advanced programmatic version. I see in the chat a question from Amanda asking how you know you want Python over something else. There are three languages that tools like Claude and ChatGPT support out of the box in their environments: Python, JavaScript, and Go. Of those three, Python has the most robust third-party libraries that can run without constantly reinventing the wheel.
We generally recommend Python for agent skills and working inside these tools. Python is the language that large language models have seen the most and are able to code in best.
Katie Robbert – 18:40
If you are someone like me, you do not automatically default to knowing there is a Python library for this. I am going to ask Claude to help me clean up my CRM on a regular basis. Is it going to suggest Python? I am not going to suggest it because it is not top of mind for me.
Christopher Penn – 19:12
It may suggest it, but you should embed it into your system instructions so that it knows to do it. As a first principle in our system instructions, I tell it to never reinvent the wheel, to ask the user, and to use web search tools to ensure it is not making something that already exists. If you are using the Trust Insights client services Claude account, it will know to search the web rather than build something net new.
Katie Robbert – 20:11
That doesn’t necessarily say it is going to suggest a Python library for it.
Christopher Penn – 20:19
No, but if you are using Opus in Claude, Astra in ChatGPT, or Argon or Carbon in Google Gemini, those advanced models will automatically infer it. They can make that logical leap. The lesser models like Sonnet or Gemini Flash will not make that leap, and you will have to specify it.
Katie Robbert – 20:51
Most of the things I am working on in Claude do not explicitly include a Python script. If it does, I am not aware of it. The average user is not going to instruct it to run a Python script. I am basically saying what I want it to do, and it just does it.
Christopher Penn – 21:33
Depending on the task you are working on, Python isn’t always the best fit. The smart models know that. If you are working on something that works best in TypeScript, Node, or React, Claude will do that behind the scenes. You just see the little thinking box.
If you use the skill fixture we have in the Trust Insights Academy, you can ask Claude to turn your conversation into a reusable skill. It will review what it did and package the scripts into the skill folder. You will see that it wrote multiple scripts on the back end that you didn’t even know were happening. These tools burn thousands of tokens at a time behind the scenes.
Katie Robbert – 22:54
Okay.
Christopher Penn – 22:55
To build out the actual heavy-duty tool, you first have to start with the 5P Framework by Trust Insights.
Katie Robbert – 23:09
For those of you who don’t know, the 5P Framework by Trust Insights is purpose, people, process, platform, and performance. Purpose is what we are doing. People addresses who will use it and who is involved. Process is how we are doing it, platform is which tools we are using, and performance is how we measure success.
People often skip over people, process, purpose, and performance and go straight to platform, asking Claude to do the task. That means you are missing four key areas of information that will make your solution rock solid. If you want to learn more, you can go to TrustInsights.ai/5P framework.
Christopher Penn – 23:48
This overall system audit details all of that. When I build using the software development lifecycle, the 5P Framework is the place I start. The purpose here is that I need to keep my CRM fresh and up to date as a CRM manager. As a user story, I need my CRM data fresh so that my sales team always has the most up-to-date information for the most valuable prospects, enabling them to win more deals.
The system architecture requires the data for my CRM, a way to search the web to check accuracy, a place to remember where the data lives, and the ideal customer profiles. I also need a language model that can process language. You could use a lightweight model like Claude Haiku or GPT-4o Mini. I am using Alibaba’s Qwen running locally on my laptop because I do not want to pay for it.
The use case compares whether a person’s job title falls within a general bucket. A language model can recognize that a director of PPC is still a marketing person, even if the word marketing is absent.
Katie Robbert – 25:28
To recap, a user story is a three-part sentence that we recommend starting with because it informs the 5P Framework. The format is, “As a persona, I want to do this, so that…” The persona answers the people aspect, the “want to” covers process and platform, and the “so that” dictates purpose and performance. Chris started with his user story, which filled in the pieces his large language model would need to satisfy the 5P Framework requirements.
Christopher Penn – 26:05
I want this running locally on my computer primarily because I am cheap and do not want to pay for cloud compute services. You could absolutely run this on Google Cloud if you like spending money for no reason. The people involved are me as the user, the CRM tool, and the AI assistant.
The process dictates that the AI looks at a company or person in the CRM and checks the job title and organization. It decides if this is an organization we want to target. It filters out irrelevant contacts, like someone selling AI software, placing them in the no pile. If it is a company we do want, it evaluates the country, company size, and job title.
If there is not enough information, it goes to get it, keeps digging, and eventually scores the record. This requires an interactive conversation with a tool like Claude, ChatGPT, or Gemini to build the product requirements document. You provide sample data from your CRM so the AI understands how to manipulate and change the data, ultimately producing the product requirements document, the technical specification, and the work plan.
Katie Robbert – 28:51
Is the other benefit of running it locally better data protection?
Christopher Penn – 28:58
It depends. If your infrastructure is safe, then yes. If your infrastructure is not safe, you are vulnerable to the weakest link in the chain, such as not having a password on your laptop. For Trust Insights, running locally on our infrastructure is the best solution because we have total control over it.
The sensitive data is processed locally without ever going out to a cloud provider where it could be at risk. I strongly encourage people to explore options for local AI hosting or local models for sensitive data. Your CRM represents the crown jewels of the company, and you do not want to hand that off to someone else.
Katie Robbert – 30:10
Right.
Christopher Penn – 30:13
What comes out is an assessment showing the ideal customer profile match, the percentage score fit, and an explanation of whether this person is a good fit. Humans must inspect this initial data to ensure the scoring mechanism is working. We might look at a contact and know they are a good fit, which means the scoring rubric needs revision if the AI scored them poorly.
After a few cycles of adjusting the mechanism, you will feel confident enough to allow it to start making updates directly to the CRM.
Katie Robbert – 31:25
You just highlighted a critical step that we as humans often skip over: true quality assurance. Starting with a smaller sample ensures that the AI is performing the task correctly. I recommend not immediately writing the extracted data back to the database. Inspect it, pause, think about it, and make sure it aligns with your expectations.
This is an essential part of software development that frequently gets missed. Starting with the 5P Framework gives you those expectations to fall back on. John, did you have anything to add to that?
John Wall – 32:53
It is the classic scenario where a script runs once and someone declares it finished. In reality, you have about 90% of the work ahead of you to test and ensure it works reliably all the time.
Christopher Penn – 33:08
Exactly. The system assesses contacts and companies and tells you where it got the information from, such as domain names, web search data, and tooling profiles. It provides its reasoning. When using smaller language models, you want the AI to document its thinking so the human in the loop can audit the decisions.
If most records state there is not enough data to judge fit, your web search functionality might be broken. You are looking for obvious errors, like a contact showing a high fit score despite the reasoning explicitly stating they are not a fit. This human-in-the-loop auditing is crucial.
Katie Robbert – 34:19
I am trying to look at the screen, but it is so tiny.
Christopher Penn – 34:26
It is not super important because this is synthetic data. I really should add a zoom feature to this application. I created this application based on a tool called Table Tool, which is great for quickly looking at large spreadsheets, but it lacks a zoom function.
Katie Robbert – 34:56
We are basically looking at the initial analysis to verify if it is performing the requested job. So far, the system hasn’t changed anything in the CRM. You are checking if it accurately identifies who best matches the ideal customer profile and provides solid reasoning. You are looking for errors, such as claims of insufficient data when the required data actually exists, or scoring issues based on CRM notes like “do not contact.”
Christopher Penn – 35:54
Ideally, you would export 10 or 20 records into a spreadsheet and hand it to someone like John with an approval column. John would mark whether the AI’s assessment is approved or wrong. After a few cycles, you feed that data back to the machine to tune the model until you start getting consistent approvals down the line.
Katie Robbert – 36:26
Have you been spying on what John and I are doing? That is exactly what we do every other day.
Christopher Penn – 36:33
I have not been. Should I be?
Katie Robbert – 36:35
No. Not unless you want to be really bored by CRM cleanup.
John Wall – 36:40
Everybody wants the inside dirt on CRM updating.
Christopher Penn – 36:45
This Python application is something we built in less than 24 hours for this live stream to demonstrate how we could use AI combined with deterministic web search to fix our CRM data. I could pay services like ZoomInfo 20,000 dollars a month, but I prefer having a machine build the same infrastructure so I can keep the money.
Katie Robbert – 37:39
When discussing CRM decay and data rot, would this process identify if a business is still functioning, if there are new owners, or if email addresses are still valid? Ensuring that contacts are up to date is one of the biggest challenges. Data appending services handle this, but it costs money.
Christopher Penn – 38:52
You do not have to pay money for that. DuckDuckGo has a Python library that allows you to search for specific individuals. If I search for John Wall using DuckDuckGo, it retrieves the relevant LinkedIn profile. We can run this search programmatically for each person in our HubSpot database to verify their current employer and job title based on their LinkedIn profile URL.
Katie Robbert – 40:10
This comes with the caveat that the human owning the LinkedIn profile has kept their profile up to date.
Christopher Penn – 40:19
Exactly, and that circles back to what John mentioned earlier about services selling snake oil because they rely on profiles that people fail to update. The software relies on search engines crawling websites to gather this information.
Regarding Amanda’s question about planning, using Claude on the terminal version unlocks a hidden model called Opus Plan. You can designate your planning model, your doing model, and your reading model. We advise using the large, smart model for planning, the mid-range model for execution, and the lightweight model for reading and summarizing web searches.
Katie Robbert – 42:42
Brian asked if we would use AI to draft a drip campaign and personalize emails. You can use AI to write a first draft or outline the campaign, but you should leave the final personalization to a human. AI can research how to best personalize an email based on job roles or CRM notes, but the human touch is essential to establish trust. We can tell when an email is personalized by an AI that doesn’t actually know us.
I recently reached out to a few older contacts and asked both of you for context on past interactions so I could include genuine personalization. Use AI to draft and research, but the human still needs to finalize it.
Christopher Penn – 45:14
No matter what you are doing, this process is not going to be cheap regarding token usage, which is why I strongly suggest using local models where possible. In 18 hours of testing without writing back to the CRM, the system consumed close to 100 million tokens. Because it ran locally on my laptop, it did not incur a massive API bill from Anthropic. Local models save a tremendous amount of money and are far more sustainable since they do not require a massive data center to power them.
Katie Robbert – 46:37
Plan it out using the 5P Framework beforehand, and run a small proof of concept with a human in the loop to ensure it understands the assignment before letting it make changes to your CRM.
Christopher Penn – 46:58
I will publish a generic version of the 5P Framework prompt I used to create this Python agent in our free Slack group, Analytics for Marketers. You can find it at TrustInsights.ai/analytics for marketers. It serves as a starter kit, though you will need to spend time customizing it for your specific needs. Any final thoughts?
Katie Robbert – 47:55
Keep your CRM up to date.
John Wall – 47:57
Just update the damn records.
Katie Robbert – 48:02
If you incorporate it into your daily process, it doesn’t get unwieldy. New tech doesn’t solve old problems; if you do not have a good process today, AI isn’t going to fix it.
Christopher Penn – 48:41
Next week on So What, Katie and John will be discussing martech sprawl. Thanks for tuning in, folks. Subscribe to our show, check out the Trust Insights podcast at TrustInsights.ai/TI podcast, and join our Slack group. See you next time.
|
Need help with your marketing AI and analytics? |
You might also enjoy:
|
|
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, INBOX INSIGHTS. Subscribe now for free; new issues every Wednesday! |
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new episodes every Wednesday. |
Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.
Newsletter
Podcast
