Fullstendig transkripsjon av Professional Services Tech Talks Episode 6: How We Automated Our CRM with AI

this task manually again? 00:05:31.839 --> 00:05:39.519 Or, you know, maybe I'll go for a walk and come back once it's back up and then uh still carry on that way. 00:05:39.680 --> 00:05:41.199 But yeah, it's interesting. 00:05:41.519 --> 00:05:41.759 Right. 00:05:41.920 --> 00:05:43.920 Shall we get into it for today then? 00:05:44.319 --> 00:05:44.959 Of course. 00:05:45.279 --> 00:05:48.240 Yeah, so we're we're gonna talk about our CRM workflow. 00:05:48.319 --> 00:05:53.120 And is CRM first of all stands for customer relationship management platform? 00:05:53.360 --> 00:06:17.759 So that's um how a business, either a business-to-business B2B or a business-to-consumer B2C organization, will track um their sales opportunities, their people they're dealing with, partners they're dealing with, clients, um, and effectively it maps out the sales and opportunity network of a particular business. 00:06:18.079 --> 00:06:45.680 So it captures who we know, um, you know, people and firms that we work with, potentially multiple people at multiple companies, what their roles are, um, how we know them, how recently we've been in contact with them, things like that, um, you know, which is obviously very useful to understand and analyze our lead generation type activity, for example. 00:06:45.920 --> 00:06:58.879 It also feeds our mailing list, so who in that contact list we send mail outs to, um, or things like invitations for drinks events or breakfasts or dinners that we run periodically. 00:06:59.199 --> 00:07:13.519 Um, so people can unsubscribe and therefore we in an automated way, we respect that and and effectively adhere to GDPR um by therefore not mailing people and things like that. 00:07:13.839 --> 00:07:18.160 And arguably most importantly, we use it for tracking our opportunities. 00:07:18.319 --> 00:07:31.040 So when someone we're talking to, a contact um potentially has a lead and that turns into a formal opportunity, we start tracking that opportunity to various opportunity stages. 00:07:31.199 --> 00:07:39.519 Typically, you go to early stage, then you go to qualification, and then it's um proposal, negotiation, and then won or lost. 00:07:39.759 --> 00:07:41.519 That's those are some typical stages. 00:07:41.759 --> 00:07:59.519 So effectively, most businesses are using it to understand the opportunity space where their existing in-flight opportunities are, but also further upstream who they're talking to, which um contacts might become leads and therefore opportunities in the future. 00:07:59.680 --> 00:08:13.839 Um, and therefore it's a very important platform for things like providing to the market confidence in your sales pipeline, um, and also having that confidence in yourself based on various metrics. 00:08:14.079 --> 00:08:20.160 We need to have this number of conversations with leads to generate this number of early stage opportunities. 00:08:20.319 --> 00:08:25.759 Those sorts of metrics are quite important in you know many types

of businesses. 00:08:26.160 --> 00:08:29.759 So that's effectively what our CRM is. 00:08:29.920 --> 00:08:38.559 And with that being defined, let's go into how we manage to keep our CRM up to date, how we did before and how we do it now. 00:08:38.960 --> 00:08:42.720 Yeah, so before it was quite an ad hoc task. 00:08:42.879 --> 00:08:47.120 It certainly wasn't something that people looked forward to doing. 00:08:47.279 --> 00:09:08.639 Uh, it was more when they had the time, they would go through their emails, you know, who have they had meetings with, and manually update the CRM so that we have an accurate account of you know where potential clients are in the pipeline and you know who have we spoken to, what are their contacts, all those kind of things. 00:09:08.879 --> 00:09:24.399 As a result of this, it was quite easy to miss certain people uh as it relied on each individual going through regularly and making sure that the details entered into the CRM were accurate. 00:09:25.039 --> 00:09:27.039 Okay, that makes sense. 00:09:27.200 --> 00:09:31.200 Um and listen, I was as guilty as this as anyone else. 00:09:31.440 --> 00:09:44.159 Um and I've been on you know in the in the Revenue Ops world and chasing sales and account people to get their things up to date, get their contacts up to date, their activities up to date. 00:09:44.320 --> 00:09:54.080 But um, you know, when when I'm due to do it, I realize that it is it's it's kind of harder than you think if it's not drilled into the fabric of your day, into your routine. 00:09:54.320 --> 00:10:00.639 So um, yeah, I think the ad hoc by hand, as we'll show, was probably not as effective as it needed to be. 00:10:00.960 --> 00:10:12.000 So we began exploring ways to automate the writing of contacts and people and firms that we'd spoken to in the week to our CRM tool. 00:10:12.159 --> 00:10:17.519 Now, we've touched on in another episode the AI infrastructure that we have. 00:10:17.759 --> 00:10:24.639 We've got Claude Team as our main large language model, and we've since rolled out ChatGPT business as well. 00:10:24.799 --> 00:10:29.840 So we can use either of these tools for any Garwood-related work. 00:10:30.799 --> 00:10:35.600 Fueling that with context is the Microsoft 365 connector. 00:10:35.679 --> 00:10:40.000 So we have access to SharePoint, OneDrive, Outlook, and Teams. 00:10:40.320 --> 00:10:48.639 We also have a meet and transcription tool in Fellow that provides verbatim context to either of these LLMs. 00:10:48.960 --> 00:10:55.279 We also have Xero, our accounting system, and BigIn, which is the CRM system that we use. 00:10:55.440 --> 00:11:04.240 Now, arguably, the most important part of assessing whether automating this process is feasible is looking at the Big In connector itself. 00:11:04.480 --> 00:11:07.600 Big in has both read and write functionality. 00:11:07.759 --> 00:11:25.440 So not only can we query our CRM within Claude or Chat GPT, but we can also write directly back to the CRM, which unlocks a number of options and workflows we can build to remove a lot of the manual work from this process. 00:11:25.840 --> 00:11:36.720 So, George, if if somebody's looking at a different CRM and they have a different CRM, and that CRM has, first of all, they need to have an MCP connector, am I right, to make this work? 00:11:36.879 --> 00:11:40.720 And um they need it to be able to both read and write. 00:11:41.039 --> 00:11:41.600 That's correct. 00:11:41.840 --> 00:11:42.639 Yes, that's correct. 00:11:43.279 --> 00:11:54.399 So once we confirmed that we have the tools available to us to automate this process, we needed to identify what steps do we need to make this happen. 00:11:54.799 --> 00:11:56.639 The first is gathering context. 00:11:56.799 --> 00:12:04.799 So we've already identified that we have a number of third-party systems that provide that context to our LLM. 00:12:04.960 --> 00:12:11.440 So email, calendar, meet and notes, we've got all of that, and that all already feeds in. 00:12:11.840 --> 00:12:23.600 So we run a scrape which goes through all of that information to identify people that we've spoken to and firms we've been in contact with, and it checks that against the CRM. 00:12:23.759 --> 00:12:29.519 So who's missing, who has gaps, um, and who do we need to fill. 00:12:29.919 --> 00:12:32.000 It then fills those gaps. 00:12:32.320 --> 00:12:37.919 Job titles and contact numbers, they can be inferred from email signatures. 00:12:38.080 --> 00:12:44.799 And where it's not sure, we can actually cross-reference with company websites to see if that information is stored there. 00:12:44.879 --> 00:12:50.240 Uh, all of this information is collated in an Excel spreadsheet, which is checked by a human. 00:12:50.399 --> 00:12:59.120 A lot of our team have communication with people outside of the organization who we wouldn't necessarily want written into our CRM. 00:12:59.200 --> 00:13:07.840 So this human in the loop stage is really important to make sure that we're only marketing to individuals that we want to. 00:13:08.080 --> 00:13:12.480 Once we've approved it, it's then automatically written to the CRM. 00:13:12.799 --> 00:13:17.279 Yeah, and that's really um that's really interesting, George. 00:13:17.440 --> 00:13:20.960 So the person checking, how do they check that? 00:13:21.120 --> 00:13:22.240 Is that an app? 00:13:22.399 --> 00:13:23.600 Is it a spreadsheet? 00:13:23.759 --> 00:13:24.879 How does that work? 00:13:25.200 --> 00:13:34.159 So our main priority was making sure that this process didn't have a lot of manual work associated with it. 00:13:34.320 --> 00:13:40.480 And any sort of friction in this new process would limit the likelihood of it being completed. 00:13:40.799 --> 00:13:55.200 So the way that we've structured it is um this the initial scrape can run independently within the cloud, and then you're sent an email that confirms that a spreadsheet is ready to look at. 00:13:55.440

--> 00:14:04.960 And in that spreadsheet, it will contain all of the people that you've spoken to over the last week with their names, contact numbers, just everything to check over. 00:14:05.120 --> 00:14:14.960 All you need to change to have that automatically written back into the CRM is just one cell in the spreadsheet uh that changes checked from no to yes. 00:14:15.200 --> 00:14:20.080 As soon as that's done, it's picked up in the next scheduled task run. 00:14:20.639 --> 00:14:23.759 If it says yes, then it will continue with the rewrite. 00:14:23.840 --> 00:14:26.879 If no, it might prompt you to go and have a look. 00:14:26.960 --> 00:14:32.159 And then as soon as it identifies that you've changed that to yes, it'll write it to the CRM. 00:14:32.720 --> 00:14:33.360 Got it. 00:14:33.840 --> 00:14:47.840 And I I personally don't mind you know going to into a spreadsheet and just checking the lines are accurate and you know reliable and in line with what I want added to the CRM. 00:14:48.000 --> 00:14:51.279 So I think that's a good that's a good interface for now. 00:14:51.600 --> 00:14:54.559 We did have some issues when we first tried to roll this out. 00:14:54.879 --> 00:15:04.639 So the plan was to have this as a weekly task, but the first time that we did it, there was quite an extensive backlog that needed to tackle it. 00:15:05.120 --> 00:15:17.759 And that consisted of months of mail and meetings which the system couldn't handle in the way that it was originally set up, so it broke it. 00:15:17.919 --> 00:15:20.799 Essentially, we had one agent running this task. 00:15:21.039 --> 00:15:30.000 There was just too much for one individual agent to go through, so we needed to re-evaluate how we address that initial backlog. 00:15:30.559 --> 00:15:32.080 There were workarounds. 00:15:32.320 --> 00:15:40.720 Initially, we found that a lot of this context was being taken up by fellow and our meeting transcription notes. 00:15:41.120 --> 00:15:50.720 So we removed that from the initial backlog and just looked at Outlook export that had no email bodies, it just looked at who and when we spoke to them. 00:15:50.879 --> 00:16:07.600 And that got us through the initial backlog, and then moving forward for backlogs for other people within Garwood, we restructured the system to have multiple subagents that had a limit on what they could tackle in any longer. 00:16:08.080 --> 00:16:25.120 So, you know, we we can spawn multiple subagents, but each one only deals with a manageable chunk of work to avoid overloading them, and then we can kind of regroup once each individual subagent's completed their task and then pull that through into a spreadsheet. 00:16:25.440 --> 00:16:30.799 Yeah, so George, with these sub-agents, what what actually happens logically? 00:16:30.960 --> 00:16:50.159 Does the LLM kind of initially say, okay, this is a this is a scrape of size X, and therefore I know I know I need to chunk it up into you know Y number of chunks, and then create a subagent to go through each chunk. 00:16:50.320 --> 00:16:52.159 Is that how it works logically? 00:16:52.559 --> 00:16:53.759 Well, almost. 00:16:54.000 --> 00:17:01.679 So actually scanning that full kind of the full breadth of that context could overload an agent because there's so much. 00:17:01.919 --> 00:17:10.880 So essentially we spawn the first subagent, the first agent deals with their portion, um, and then once they've reached that, they know that there's more. 00:17:11.119 --> 00:17:25.759 So they call the next agent, they deal with theirs, there's still more, they call the next agent, and that kind of goes on until the overall task is completed, and then there's another agent that kind of groups all of that together and pulls that through into a spreadsheet. 00:17:26.079 --> 00:17:26.480 Got it. 00:17:26.559 --> 00:17:35.440 So the first agent is created automatically, and then it decides whether there's another agent who's needed to continue the task, and so on and so forth. 00:17:35.599 --> 00:17:36.720 Okay, interesting. 00:17:37.039 --> 00:17:46.720 Another obstacle that we needed to overcome was making sure that this could run without the need for someone's computer to be on. 00:17:47.039 --> 00:18:00.640 And that wasn't possible across the whole task, but we were able to segment it into smaller tasks to ensure that only a small portion requires that computer being on. 00:18:00.960 --> 00:18:06.559 So the weekly sweep itself all runs through third-party connectors. 00:18:06.880 --> 00:18:11.519 So that can run in the cloud independent of someone's computer. 00:18:11.839 --> 00:18:21.519 The issue that we had was that we needed that computer to be on when we were writing a spreadsheet to their system. 00:18:22.160 --> 00:18:32.079 So the first part of the check runs independently, it sends an email and tells uh the individual that it's ready to check. 00:18:32.240 --> 00:18:46.400 Um they then click that link, it writes the spreadsheet to their PC, they then check that, set it to yes, and then the next scheduled task picks up that update and then writes it to the CRE. 00:18:46.799 --> 00:18:52.720 So the initial check can run in the cloud, but the actual write requires their machine to be on. 00:18:53.599 --> 00:18:55.440 Okay, makes sense. 00:18:55.759 --> 00:19:02.480 And let's say I'm traveling, my laptop's off until 10am, you know, it's the check around at 8:30. 00:19:02.559 --> 00:19:08.400 What then happens is it wait until the next time I'm I'm online at 8.30, or would it run again that day? 00:19:08.799 --> 00:19:11.200 So it should wait for the next run. 00:19:11.680 --> 00:19:19.680 However, we've set that up in a way that's it's frequent enough that it will be picked up shortly after. 00:19:20.079 --> 00:19:20.400 Okay. 00:19:21.359 --> 00:19:21.680 All right. 00:19:21.839 --> 00:19:40.160 And if there's any additional contacts or um, you know, firms that w I've been in contact with since yesterday's

sweep, that will get picked up in the next sweep and therefore added to the to the spreadsheet that that then gets created on my laptop. 00:19:40.480 --> 00:19:40.960 That's right. 00:19:41.119 --> 00:19:50.880 So involved in the check is an assessment of who you've spoken to over the past week, and it cross-references that against who we've got in our CRM system. 00:19:51.119 --> 00:20:01.680 So if they've already been picked up in a previous sweep, but they haven't been written into the CRM, that gap's still there, so they'll be picked up and put in this in this next list. 00:20:02.160 --> 00:20:05.920 So what was the impact of automating this process? 00:20:06.240 --> 00:20:20.960 Well, we implemented this in September, and since then we've seen 236 new contacts uploaded to the CRM and 44 new companies from everyone's contacts within their email, fellow, or teams. 00:20:21.359 --> 00:20:29.599 I mean, that's just brilliant, and there's almost certainly paid for itself, you know, within the first week of running. 00:20:29.759 --> 00:20:36.319 Um, that amount of people who we've been talking to who weren't captured on the CRM is really significant. 00:20:36.720 --> 00:20:37.519 Yeah, definitely. 00:20:37.599 --> 00:20:40.880 And we expect to pick up more week by week. 00:20:41.039 --> 00:20:46.480 Um, but yeah, to remove that manual work has been really impactful for us. 00:20:46.799 --> 00:20:54.720 George, if you're a professional services business thinking of doing this or something similar, what do you need to do and what needs to be in place? 00:20:55.119 --> 00:20:58.400 So the first would be to check your stack. 00:20:58.880 --> 00:21:04.480 Um, does the CRM that you use have a first-party MCP connector? 00:21:04.640 --> 00:21:08.160 Does that have both read and write functionality? 00:21:08.480 --> 00:21:12.000 If so, you're headed in the right direction. 00:21:12.319 --> 00:21:15.599 The second is a company-wide AI tool. 00:21:15.759 --> 00:21:23.279 So we mentioned that this process relied very heavily on the wider AI infrastructure that we've set up. 00:21:23.359 --> 00:21:28.319 So being able to access Outlook, Teams, a meet and transcription tool. 00:21:28.559 --> 00:21:34.400 That's all dependent on having those systems built into your wider operating model. 00:21:34.640 --> 00:21:52.799 And in order to do so safely, you need a company-wide approved AI platform that doesn't train on inputted data, that you have a data processing agreement, and all of those kind of safety and security pieces that are really important for using AI safely within the business. 00:21:53.359 --> 00:21:56.319 The third would be a human in the loop. 00:21:56.559 --> 00:21:59.759 So you need to identify a clear step. 00:22:00.319 --> 00:22:09.599 That a human can check the process to avoid having people written into the CRM who you wouldn't have written in manually. 00:22:10.559 --> 00:22:13.759 The fourth is managing the workload. 00:22:14.000 --> 00:22:21.119 So we're not necessarily finding this to be a problem on the standard weekly scheduled suite. 00:22:21.440 --> 00:22:30.960 However, with the initial backlog, there's likely to be quite a difference in the number of contacts that different people have spoken to. 00:22:31.680 --> 00:22:38.400 And it is possible that this will break the process as it did for us. 00:22:39.440 --> 00:22:46.240 So make sure you're using subagent so you can split up that work and manage it suitably. 00:22:46.480 --> 00:22:51.839 And the fifth would be clear rules for who stays out of the CRM. 00:22:52.480 --> 00:23:03.759 So there's a number of like engagement groups, for example, where you know people have committed to not marketing to contacts that are within that engagement group. 00:23:03.920 --> 00:23:17.039 You can set rules like that up so that if it identifies a contact that belongs to a certain group, they're not picked up in that scan and subsequently not written to the CRM and marketed to. 00:23:17.359 --> 00:23:18.720 Yeah, it makes absolute sense. 00:23:18.880 --> 00:23:28.079 And, you know, I know, you know, a few people, myself included, will be in networking groups that are, you know, groups for like-minded professionals. 00:23:28.240 --> 00:23:34.079 We you know may use our work emails, we may be on group, you know, group calendar events and things like that. 00:23:34.240 --> 00:23:46.000 But it's important to be able to keep those separate from genuine um contacts with whom we're talking and and who might want to use Garward services at some point in the future. 00:23:46.400 --> 00:23:49.440 So, George, we're at the end of our six-part series. 00:23:49.680 --> 00:23:55.440 Um, do you want to run us through what we what we've covered in the series in case people want to go back and listen to other episodes? 00:23:55.839 --> 00:23:56.480 Yeah, of course. 00:23:56.640 --> 00:23:59.039 So, episode one was choosing our stack. 00:23:59.200 --> 00:24:05.599 So, how did we select the large language model that we wanted to use within Garwood? 00:24:05.759 --> 00:24:17.359 Uh the second was governance, so what security processes and kind of what oversight did we need to use that system safely and effectively? 00:24:18.079 --> 00:24:32.799 The third started to look into uh how we can build skills to remove some of the manual processes, such as branding documents and having our output sound more like us. 00:24:33.440 --> 00:24:37.680 The fourth was focused on meeting transcription. 00:24:37.839 --> 00:24:45.839 So, how do we select our meeting transcription tool to remove the need to take manual notes in meetings? 00:24:46.079 --> 00:24:57.119 Uh, the fifth explored how we build new systems within Garwood, turning our existing IP into a self-service lead magnet. 00:24:57.519 --> 00:25:00.720 And the last episode was focused on our CRM workflow. 00:25:00.799 --> 00:25:11.200 So, how did we use the tools that we've set up um to automate and remove the manual work required to write contacts into ASVRM? 00:25:12.480 --> 00:25:13.759 Yeah, thanks, George. 00:25:13.839 --> 00:25:17.759 And as you kind of walk back through, I can see a real trend here. 00:25:17.839 --> 00:25:21.200 So I think episodes one to four were very much the foundation. 00:25:21.519 --> 00:25:26.160 You know, how do we set those foundations to move forward? 00:25:26.480 --> 00:25:31.680 Um, episode five, how to build a lead magnet, was very much about opportunity AI. 00:25:31.839 --> 00:25:36.240 How can we use AI to create more opportunities for ourselves as a business? 00:25:36.480 --> 00:25:42.720 And number six, episode six, which was around the CRM workflow, was much more about efficiency. 00:25:42.880 --> 00:25:49.279 How do we make things more efficient, more effective within the business in processes that otherwise aren't automated? 00:25:49.440 --> 00:26:01.279 So I think that's a really good um, you know, point to end on, a really good foundation for how we use AI within Garwood and how we think other professional services businesses could use AI. 00:26:01.359 --> 00:26:04.880 Um, I've really enjoyed our time together on these podcasts, George. 00:26:05.039 --> 00:26:11.519 Um, and I want to give a big shout out to you because it's your work that's made all this possible. 00:26:11.680 --> 00:26:13.519 So um a massive thank you. 00:26:13.680 --> 00:26:17.119 Um and I know the our other Garwood colleagues would say the same. 00:26:17.359 --> 00:26:28.880 Um and I guess the only thing left to say is that if you enjoyed this podcast, do check out our Unlocking Value podcast, which is um can be found on your podcast platform of choice. 00:26:28.960 --> 00:26:37.200 That's Unlocking Value, where our colleague John Howard talks to um leaders and investors in professional services businesses. 00:26:37.359 --> 00:26:41.440 And other than that, until next time, until next series, it's bye from me. 00:26:42.720 --> 00:26:43.759 Bye from me.