WEBVTT
00:00:00.000 --> 00:00:02.720
Hi, I'm George, Senior Consultant at Garwood Growth.
00:00:02.959 --> 00:00:05.599
And I'm Adam, managing consultant at Garwood Growth.
00:00:06.000 --> 00:00:16.239
Welcome to my new podcast, where we talk tech in professional services businesses, in particular, how to select, implement, and operate key technology platforms based on our own experiences.
00:00:16.559 --> 00:00:24.879
We'll talk about the good, the bad, and the ugly of tech in a PS business, and share our learnings and experiences along the way, warts and all.
00:00:25.199 --> 00:00:30.559
To kick it off, we're going to be running a series on a little known and rarely discussed topic in tech.
00:00:30.719 --> 00:00:32.479
You guessed it, AI.
00:00:32.799 --> 00:00:36.320
George, can you tell our listeners and viewers what to expect from this podcast?
00:00:36.640 --> 00:00:42.159
This podcast is for leaders through to fiendas in boutique to medium-sized professional services firms.
00:00:42.320 --> 00:00:51.039
It's really focused on people who are interested in getting beyond the headlines and finding out about the practicalities of making technology work in a professional services firm.
00:00:51.280 --> 00:00:56.399
In this first series of the podcast, we'll be talking about how to choose, set up and run your AI platform.
00:00:56.640 --> 00:01:09.760
We'll cover a range of topics across these episodes, which includes, but isn't limited to choosing your stack, governance and data, having it sound like you, connecting third-party systems and putting it to work.
00:01:10.159 --> 00:01:16.319
And George, you know, you and I, we're consultants, we like PowerPoint as much as the next consultant, don't we?
00:01:16.560 --> 00:01:22.239
But just to reassure our listeners and viewers that this series is not going to be deaf by PowerPoint.
00:01:22.400 --> 00:01:28.319
We're going to get very much into the tools and platforms that you would be using day to day when you implement AI.
00:01:28.560 --> 00:01:29.359
Yeah, absolutely.
00:01:29.519 --> 00:01:32.560
That's something we want to definitely stay clear of.
00:01:33.359 --> 00:01:54.319
So if we go to the challenge that we have overall, we know that AI is being used in pockets across the business, but they're individual tools on individual accounts, which can be very effective in maximizing individual efficiency, but they don't really carry that compounding organizational impact that institutional AI has.
00:01:54.719 --> 00:01:56.239
Yeah, absolutely, George.
00:01:56.400 --> 00:02:01.680
And I think we are seeing now the acknowledgement that AI has made people really efficient.
00:02:01.760 --> 00:02:05.359
You know, individual AI is commensurate with efficiency AI.
00:02:05.519 --> 00:02:09.919
And what we need to move towards is really driving forward opportunity AI.
00:02:10.000 --> 00:02:13.120
And I think that's very strongly aligned with institutional AI.
00:02:13.199 --> 00:02:25.759
Once we institutionalize the benefits that you can get from AI in terms of how to fundamentally change the way you do business, I think that will drive forward growth in a lot of professional services businesses.
00:02:26.000 --> 00:02:26.960
Yeah, absolutely.
00:02:27.199 --> 00:02:27.840
Absolutely.
00:02:28.000 --> 00:02:32.719
So for our first episode, we're looking at how we chose our AI stack.
00:02:32.879 --> 00:02:35.919
We split this into four key areas.
00:02:36.240 --> 00:02:38.960
The first was sending out the questionnaire.
00:02:39.120 --> 00:02:44.400
So we needed to understand how people use AI in the business currently and to what level.
00:02:44.639 --> 00:02:46.960
We would then move to a findings report.
00:02:47.120 --> 00:02:48.560
So what have people told us?
00:02:48.719 --> 00:02:53.919
What are the themes that emerge from those questionnaire responses before moving to vendor research?
00:02:54.080 --> 00:02:57.599
So which vendors meet those success criteria?
00:02:57.759 --> 00:03:04.400
And then we can make a recommendation on which large language model solution do we think is most applicable to us here at Galwood?
00:03:04.639 --> 00:03:06.479
Yeah, George, all sounds really logical.
00:03:06.719 --> 00:03:09.599
Just out of interest, how did you run the actual questionnaire?
00:03:09.840 --> 00:03:19.360
We ran it through Microsoft Forms, mainly because people know how to use it and it gives us a structured Excel output that we could then synthesize and move forward with.
00:03:19.520 --> 00:03:20.319
Yeah, it makes sense.
00:03:20.560 --> 00:03:22.800
The first step was sending out the questionnaire.
00:03:22.879 --> 00:03:27.120
Now, the questionnaire was loosely based around these six key areas.
00:03:27.280 --> 00:03:30.240
The first being who you are, you role and current AI use.
00:03:30.479 --> 00:03:32.479
The second is how you use AI.
00:03:32.560 --> 00:03:37.360
So what frequency do you use these tools at, and how sophisticated are you in using them?
00:03:37.520 --> 00:03:40.319
The third is how much you trust those tools.
00:03:40.560 --> 00:03:50.240
So if you're using them, how much manual work is required to tweak the output into a position that you can then move forward with it?
00:03:50.639 --> 00:03:52.080
Fourth is where your time goes.
00:03:52.159 --> 00:03:59.680
So this slightly comes away from the AI specific side of things, but where are the bottlenecks in your week?
00:03:59.840 --> 00:04:05.599
What do you spend the most amount of time on and where can we potentially target an AI solution to give you back some of that time?
00:04:05.840 --> 00:04:07.280
The fifth is what you handle.
00:04:07.439 --> 00:04:11.280
So data security, what kind of data are you handling?
00:04:11.360 --> 00:04:19.759
And that will inform our decision on how secure these systems have to be to make sure that that data is dealt with appropriately.
00:04:20.000 --> 00:04:22.879
And the sixth is appetite and constraints.
00:04:23.120 --> 00:04:33.759
So more of a an attitude question around how open you are to changing the way that you work and what do we need to get right to make sure that this project is a success.
00:04:34.000 --> 00:04:36.399
Yeah, it's all sensible stuff, isn't it, George?
00:04:36.560 --> 00:04:41.360
I wondered, was there anything else that you wanted to capture but but didn't?
00:04:41.759 --> 00:04:52.959
I think we've covered this across the board, but one of our primary concerns when rolling this out was we didn't want everyone to spend a lot of time answering the questionnaires.
00:04:53.120 --> 00:04:57.600
We wanted the key consulting-based questions, which is what we included.
00:04:57.839 --> 00:05:01.759
We're not bothered about what AI you use to kind of record your shopping list, that kind of thing.
00:05:01.839 --> 00:05:02.480
That's irrelevant.
00:05:02.560 --> 00:05:09.759
We wanted to keep it as concise as possible and stick to between 10 and 20 minutes at the responses overall.
00:05:10.000 --> 00:05:10.079
Yeah.
00:05:10.319 --> 00:05:12.399
AI for your shopping list, that sounds brilliant.
00:05:12.720 --> 00:05:17.040
Yeah, I'll have to say I need some after the podcast, I need some details on that, please, George.
00:05:17.199 --> 00:05:17.759
Yeah, I'm sure.
00:05:17.839 --> 00:05:18.319
No problem.
00:05:18.480 --> 00:05:19.040
No problem.
00:05:19.519 --> 00:05:23.920
Questionnaire came back with a range of answers, but the key themes we've labelled here.
00:05:24.079 --> 00:05:34.560
Everyone's using AI tools at work currently, and there is an appetite or at least an openness to change the way that people work, to feel the benefit of these tools.
00:05:34.879 --> 00:05:38.000
60% are using AI either weekly or daily.
00:05:38.160 --> 00:05:40.319
I'm sure Adam, you you understand that.
00:05:40.480 --> 00:05:45.199
One of the most important parts with any sort of tech implementation is adoption.
00:05:45.360 --> 00:05:53.839
And starting from a position where people already use these tools certainly helps get them on board with using what we're trying to put forward.
00:05:54.079 --> 00:05:54.879
Yeah, absolutely.
00:05:54.959 --> 00:06:02.160
You know, if if you're already using these types of tools, as many people in professional services and consulting are, you're pushing against an open door.
00:06:02.319 --> 00:06:03.360
Yeah, absolutely.
00:06:03.600 --> 00:06:08.560
Something that surprised me coming back was the fact that five chatbot brands are currently in use.
00:06:08.800 --> 00:06:23.519
Admittedly, I was expecting it to come back saying that people have dabbled with ChatGPT but but haven't really explored some of the alternatives, which was a pleasant surprise because it means that our consultants aren't necessarily just going with the mainstream option.
00:06:23.680 --> 00:06:26.319
They are choosing the tools that work for them.
00:06:26.480 --> 00:06:30.480
One of the main questions that we asked was where time is being lost.
00:06:30.720 --> 00:06:36.879
What came out on top was turning meeting notes into findings and meeting prep notes and follow-ups.
00:06:37.120 --> 00:06:42.319
A hundred percent of the team hit at least one of these five tasks weekly.
00:06:42.480 --> 00:06:47.040
So it was clear what the team's success criteria were for using a large language model.
00:06:47.439 --> 00:06:48.480
Yeah, that's that's really good.
00:06:48.560 --> 00:06:51.759
And I see the top one there is turning meeting notes into findings.
00:06:51.839 --> 00:06:57.360
I know as a consultant I've spent many hours scribbling meeting notes and then trying to make sense of them afterwards.
00:06:57.439 --> 00:06:59.920
So I can see that's very much a key requirement.
00:07:00.160 --> 00:07:00.879
Yeah, absolutely.
00:07:00.959 --> 00:07:03.680
And and it can often take you out of the meeting as well.
00:07:03.839 --> 00:07:06.959
You want to be focusing on the conversation, not recording it.
00:07:07.040 --> 00:07:12.319
And your notes are rarely to the level of depth that the conversation actually has.
00:07:12.800 --> 00:07:16.000
So a tool that automated this was a key priority.
00:07:16.240 --> 00:07:17.120
Yeah, makes sense.
00:07:17.360 --> 00:07:19.360
One of the other key findings was risk.
00:07:19.519 --> 00:07:22.879
So everyone in the business handles sensitive material every week.
00:07:23.040 --> 00:07:32.319
And that stems from client confidential material to NDA protected material, deal material, and then Garwood internal protected information.
00:07:33.040 --> 00:07:36.720
Yeah, I mean, it's really important, isn't it, your risk posture as a business.
00:07:36.959 --> 00:07:46.879
A lot of people don't realize that when you're using personal free AI platforms, those platforms are taking the data and information that you're uploading and using that to train their models.
00:07:47.040 --> 00:07:56.079
So what people don't realize is if you're using the free versions of things like ChatGPT, um, the model is by default training on the data that you put in.
00:07:56.160 --> 00:08:06.959
And it's not beyond the realm of possibility that a bad actor might try and invert the model, which means using very specific prompts to try and get that data back out of the large language model.
00:08:07.120 --> 00:08:09.279
So you do really need to be careful with this stuff.
00:08:09.519 --> 00:08:10.160
Yeah, absolutely.
00:08:10.240 --> 00:08:14.720
And it's not just free accounts either, even some individual paid accounts.
00:08:14.879 --> 00:08:27.040
While there is a toggle to turn off the fact that the model is trained on input data, you do have to be proactive about setting up your account so that it doesn't train the model.
00:08:27.199 --> 00:08:29.279
So it's it's certainly one to be aware of.
00:08:29.519 --> 00:08:34.720
Once we've received the questionnaire feedback, we identified five key themes emerging from it.
00:08:34.960 --> 00:08:42.720
The first was meetings into findings, second was knowledge retrieval, proposal production, tool consolidation, and governance.
00:08:42.960 --> 00:08:43.120
Yeah.
00:08:43.279 --> 00:08:47.519
George, we've got a combination of qualitative and quantitative inputs here.
00:08:47.679 --> 00:08:52.639
How did he use the free text inputs into the questionnaire to get these outcomes?
00:08:53.039 --> 00:09:03.039
A mix of qualitative and quantitative information was important to give us the fullest picture of how people are experiencing AI currently.
00:09:03.519 --> 00:09:17.039
We included open text boxes, particularly around the constraints and things we need to get right to make this project successful, because we didn't want to force people to choose from a predetermined set of options.
00:09:17.200 --> 00:09:22.080
We wanted to understand what people's success criteria are for the project overall.
00:09:22.720 --> 00:09:24.960
Synthesizing it was fairly easy.
00:09:25.120 --> 00:09:33.279
We were able to use a clawed project, and any large language model is good at taking unstructured input and structuring it.
00:09:33.679 --> 00:09:39.200
And we were able to use that in conjunction with the structured output that we got from the task ratings.
00:09:39.840 --> 00:09:47.440
In order to determine how we went out to market, we needed to convert those themes into specific success criteria.
00:09:47.759 --> 00:10:03.679
We came up with seven: enterprise grade data handling, Microsoft 365 access, workspace and access control, long document and transcript analysis, meeting transcription, persistent workspaces or projects, and a credible integration path.
00:10:03.919 --> 00:10:06.240
Yeah, I mean these all look pretty critical.
00:10:06.399 --> 00:10:12.159
What would you say if you had to pick out two or three as really, really critical, which ones would you pick?
00:10:12.480 --> 00:10:21.360
Enterprise grade data handling is certainly one of the most important due to the nature of the information that we use on a day-to-day basis.
00:10:22.000 --> 00:10:28.159
We needed to make sure whichever large language model we opt for is UK GDPR compliant.
00:10:28.320 --> 00:10:36.240
And we needed a signed data processing agreement that confirms that the LLM won't train on inputted data.
00:10:36.639 --> 00:10:41.039
Another important one, perhaps rather selfishly, is meeting transcription.
00:10:41.120 --> 00:10:46.639
So I was taking a lot of notes in meetings prior to rolling out our transcription tool.
00:10:46.879 --> 00:10:56.240
And I would often leave meetings with smoke coming off my pen and not fully capturing the context that was discussed in the meeting.
00:10:56.480 --> 00:11:01.679
So that was something that was incredibly important in whatever solution we decided on.
00:11:02.000 --> 00:11:04.080
But also Microsoft 365 access.
00:11:04.240 --> 00:11:09.360
One of the barriers to adoption would be completely changing the way that people work.
00:11:09.679 --> 00:11:13.120
People are working in Microsoft 365, so the tool has to as well.
00:11:13.679 --> 00:11:13.919
Great.
00:11:14.480 --> 00:11:21.279
Once we had the success criteria, we needed to identify how we're going to assess vendors against it.
00:11:22.000 --> 00:11:32.639
We adopted a structure of starting with a Claude project, which took the scorings, took the free text, and synthesized it into a prompt that we could then go out to market with.
00:11:33.039 --> 00:11:44.639
We then pushed that through perplexity to identify the vendors that met those success criteria, which ones they kind of met but not fully, and which that they didn't hit at all.
00:11:44.799 --> 00:11:50.240
We then came up with a short list which showed three credible providers, which we assessed in a little bit more detail.
00:11:50.799 --> 00:11:59.360
George, just um help me realize because I'm a bit of a cynic, if Claude was involved in the research, how do we know it didn't just select itself?
00:11:59.759 --> 00:12:01.600
Yeah, it's it's a good question.
00:12:01.759 --> 00:12:06.159
So the Claude project that we used was to synthesize the findings.
00:12:06.320 --> 00:12:11.519
We didn't use Claude itself to go out to market and recommend vendors.
00:12:11.759 --> 00:12:16.559
We don't want any sort of kind of favoring of one large language model over another.
00:12:16.799 --> 00:12:18.000
So we use perplexity.
00:12:18.240 --> 00:12:21.919
The benefit of using perplexity is that it's model agnostic.
00:12:22.080 --> 00:12:30.320
Its deep research function uses a range of models from different providers, collates the responses, and gives us the most objective view.
00:12:30.559 --> 00:12:37.279
Once this was completed, we had a report which confirmed three platforms cleared the bar.
00:12:37.519 --> 00:12:47.200
We assessed them specifically against those success criteria, and it's clear that some performed better than others, and that fueled our decision on which LLM we would choose.
00:12:47.440 --> 00:12:54.960
I think it's important to say, isn't it, George, that this was done probably late May, early June 2026 for those listening later on.
00:12:55.200 --> 00:12:59.600
The world has changed hugely since that because things move so fast in the world of AI.
00:12:59.679 --> 00:13:06.399
And for example, ChatGPT didn't have those Microsoft platform plugins when we did this assessment, but they do now.
00:13:06.639 --> 00:13:09.440
So it's important to know that this is a point in time.
00:13:09.519 --> 00:13:19.120
And it's probably, given this pace of change within AI, worth doing in a light version of this exercise every quarter or so to make sure that the platform you select meets your needs.
00:13:19.440 --> 00:13:25.759
And you open the opportunity to do that with the infrastructure that you set up as well, which we'll touch on later.
00:13:26.000 --> 00:13:31.600
If you build your AI infrastructure well, you can switch the model that is being used.
00:13:31.679 --> 00:13:35.279
And it's always good to have that approach as models change all the time.
00:13:35.440 --> 00:13:40.320
And at the moment, it feels like ChatGPT and Claude are constantly one-up in each other with every model.
00:13:40.879 --> 00:13:42.799
So it's good to have an open mind on that through.
00:13:43.039 --> 00:13:47.200
But at the time we conducted the research, Claude team did come out on top.
00:13:47.440 --> 00:13:50.399
And there are three main reasons for why that's the case.
00:13:50.559 --> 00:13:56.159
So it's best where work is hardest, meetings to findings, proposal production, knowledge retrieval.
00:13:56.320 --> 00:14:04.799
Its integrations are more developed, and we're able to tap into the information that we need to to create draft deliverables most efficiently.
00:14:05.679 --> 00:14:08.720
Similarly, it reads Microsoft 365 natively.
00:14:08.799 --> 00:14:18.799
So it has a first-party MCP connector which reaches SharePoint OneDrive Outlook Teams, which provides Claude with the context it needs to move forward with Glide.
00:14:19.120 --> 00:14:20.960
And also it keeps the firm portable.
00:14:21.120 --> 00:14:34.399
So Floud released the model context protocol in 2024, which is an open standard that third-party integrations can adhere to to make sure that their apps can connect directly with Claude.
00:14:34.799 --> 00:14:39.519
Okay, so that's the MCP acronym I hear being banded around all the time.
00:14:39.840 --> 00:14:41.679
That means model context protocol.
00:14:41.919 --> 00:14:42.320
That's it.
00:14:42.480 --> 00:14:44.879
Okay, like an API for your AI platform.
00:14:45.200 --> 00:14:53.440
You can still use APIs for AI platforms, but yes, it's an open standard that third parties can adhere to make sure that they're able to connect with Clock.
00:14:53.759 --> 00:14:54.000
Right.
00:14:54.240 --> 00:14:58.399
We were then able to flesh out our AI infrastructure overall.
00:14:58.559 --> 00:15:07.840
So you can see that we've selected Claude team, and above it sits the Microsoft 365 content that fuels the bulk of the context that we're feeding into Clock.
00:15:08.320 --> 00:15:11.759
This covers SharePoint, OneDrive, Outlook, and Teams.
00:15:11.919 --> 00:15:17.919
And we also selected a meeting transcription tool to give us even more context from meetings.
00:15:18.159 --> 00:15:20.799
I know a lot of our viewers and listeners will be interested.
00:15:21.039 --> 00:15:24.080
Fellow is our meeting transcription tool.
00:15:24.320 --> 00:15:27.039
There's lots of other meeting transcription tools out there.
00:15:27.200 --> 00:15:34.399
Of course, Teams Native, things like Fathom, Otter, Fireflies, Granola is one that I hear coming up a lot.
00:15:34.559 --> 00:15:36.080
What made us choose Fellow?
00:15:36.960 --> 00:15:50.960
We did analyze whether we could use the inbuilt Teams transcription and have Claude read that directly, but it would require an automation that pushes that transcription from Teams into SharePoint, which was quite a lot of work.
00:15:51.120 --> 00:15:57.600
We then started evaluating third-party options, and Fellow came out on top for a couple of key reasons.
00:15:57.919 --> 00:16:02.960
The first is it offers both bot and botless recording.
00:16:03.200 --> 00:16:09.840
And for anyone who doesn't know, a botted recording is when a bot joins your call.
00:16:09.919 --> 00:16:17.840
And that's particularly effective because the quality of the transcription a bot can do is much better than a botless approach.
00:16:18.240 --> 00:16:32.639
Within Teams, it's recording everyone's audio individually, so it can attribute speakers to their speech and it allows it to not get bogged down by crossover speech as a botless approach would.
00:16:32.799 --> 00:16:36.000
Now that's not to say a botless approach doesn't come in handy.
00:16:36.240 --> 00:16:45.840
If we're conducting in-person workshops, for example, we would need to put a microphone in the middle of the room, and it would be more difficult to attribute speakers.
00:16:46.080 --> 00:16:52.799
The fellow gives us an option no matter the circumstances and allows us to store all of that meeting context in the same place.
00:16:54.720 --> 00:16:55.679
What does it unlock?
00:16:55.919 --> 00:17:01.840
We've already spoken about the fact that our AI infrastructure is essentially a context store.
00:17:02.000 --> 00:17:10.240
Every meeting, every document, email and chat, and the firm's IP are all available and can be referenced by Claude when creating documents.
00:17:10.559 --> 00:17:11.599
What are the actual outputs?
00:17:11.759 --> 00:17:13.519
We're able to find information faster.
00:17:13.759 --> 00:17:17.200
We're standardized in draft outputs and quality assurance.
00:17:17.519 --> 00:17:21.519
And we're giving ourselves the time to really focus on the higher value work.
00:17:21.680 --> 00:17:26.160
Within consulting generally, the value proposition is our expertise.
00:17:26.319 --> 00:17:32.480
And a lot of the time we can be bogged down in the manual elements of the role that AI can now solve.
00:17:33.759 --> 00:17:36.000
Yeah, I totally agree with that, George.
00:17:36.079 --> 00:17:45.599
And in the relatively short time I've been at Garwood this second time around, thanks to the infrastructure put in place, I really have utilized AI in a way that I hadn't before.
00:17:45.759 --> 00:17:47.440
For example, writing decks.
00:17:47.599 --> 00:17:50.880
I've not created a deck from scratch since I've been at Garwood.
00:17:51.039 --> 00:17:58.160
My time is better spent synthesizing and inserting the context into the platform.
00:17:58.319 --> 00:18:04.799
Claude pushes out perfectly branded documents from first instance, so I don't have to worry about the latest marketing template.
00:18:04.880 --> 00:18:10.799
So it really saves that sort of time that people don't necessarily think about, but was a time suck in days gone by.
00:18:11.599 --> 00:18:13.680
What are we going to talk about in the next episode?
00:18:14.000 --> 00:18:17.200
Next time we're going to be looking at the boring but essential half.
00:18:17.359 --> 00:18:19.920
So governance, how do we manage these systems?
00:18:20.160 --> 00:18:24.640
Acceptable use, what policies do we need to have in place to protect the business?
00:18:24.799 --> 00:18:26.480
And training individuals.
00:18:26.640 --> 00:18:30.400
How do we make sure that everyone can get the most out of their IOTs?
00:18:30.799 --> 00:18:31.279
Brilliant.
00:18:31.440 --> 00:18:32.319
Thanks, George.
00:18:32.640 --> 00:18:36.160
Just to our viewers and listeners, I've been Adam Mays of Garwood Growth.
00:18:36.400 --> 00:18:37.680
I've been George White.
00:18:38.160 --> 00:18:39.599
See you next time.
00:00:00.000 --> 00:00:02.720
Hi, I'm George, Senior Consultant at Garwood Growth.
00:00:02.959 --> 00:00:05.599
And I'm Adam, managing consultant at Garwood Growth.
00:00:06.000 --> 00:00:16.239
Welcome to my new podcast, where we talk tech in professional services businesses, in particular, how to select, implement, and operate key technology platforms based on our own experiences.
00:00:16.559 --> 00:00:24.879
We'll talk about the good, the bad, and the ugly of tech in a PS business, and share our learnings and experiences along the way, warts and all.
00:00:25.199 --> 00:00:30.559
To kick it off, we're going to be running a series on a little known and rarely discussed topic in tech.
00:00:30.719 --> 00:00:32.479
You guessed it, AI.
00:00:32.799 --> 00:00:36.320
George, can you tell our listeners and viewers what to expect from this podcast?
00:00:36.640 --> 00:00:42.159
This podcast is for leaders through to fiendas in boutique to medium-sized professional services firms.
00:00:42.320 --> 00:00:51.039
It's really focused on people who are interested in getting beyond the headlines and finding out about the practicalities of making technology work in a professional services firm.
00:00:51.280 --> 00:00:56.399
In this first series of the podcast, we'll be talking about how to choose, set up and run your AI platform.
00:00:56.640 --> 00:01:09.760
We'll cover a range of topics across these episodes, which includes, but isn't limited to choosing your stack, governance and data, having it sound like you, connecting third-party systems and putting it to work.
00:01:10.159 --> 00:01:16.319
And George, you know, you and I, we're consultants, we like PowerPoint as much as the next consultant, don't we?
00:01:16.560 --> 00:01:22.239
But just to reassure our listeners and viewers that this series is not going to be deaf by PowerPoint.
00:01:22.400 --> 00:01:28.319
We're going to get very much into the tools and platforms that you would be using day to day when you implement AI.
00:01:28.560 --> 00:01:29.359
Yeah, absolutely.
00:01:29.519 --> 00:01:32.560
That's something we want to definitely stay clear of.
00:01:33.359 --> 00:01:54.319
So if we go to the challenge that we have overall, we know that AI is being used in pockets across the business, but they're individual tools on individual accounts, which can be very effective in maximizing individual efficiency, but they don't really carry that compounding organizational impact that institutional AI has.
00:01:54.719 --> 00:01:56.239
Yeah, absolutely, George.
00:01:56.400 --> 00:02:01.680
And I think we are seeing now the acknowledgement that AI has made people really efficient.
00:02:01.760 --> 00:02:05.359
You know, individual AI is commensurate with efficiency AI.
00:02:05.519 --> 00:02:09.919
And what we need to move towards is really driving forward opportunity AI.
00:02:10.000 --> 00:02:13.120
And I think that's very strongly aligned with institutional AI.
00:02:13.199 --> 00:02:25.759
Once we institutionalize the benefits that you can get from AI in terms of how to fundamentally change the way you do business, I think that will drive forward growth in a lot of professional services businesses.
00:02:26.000 --> 00:02:26.960
Yeah, absolutely.
00:02:27.199 --> 00:02:27.840
Absolutely.
00:02:28.000 --> 00:02:32.719
So for our first episode, we're looking at how we chose our AI stack.
00:02:32.879 --> 00:02:35.919
We split this into four key areas.
00:02:36.240 --> 00:02:38.960
The first was sending out the questionnaire.
00:02:39.120 --> 00:02:44.400
So we needed to understand how people use AI in the business currently and to what level.
00:02:44.639 --> 00:02:46.960
We would then move to a findings report.
00:02:47.120 --> 00:02:48.560
So what have people told us?
00:02:48.719 --> 00:02:53.919
What are the themes that emerge from those questionnaire responses before moving to vendor research?
00:02:54.080 --> 00:02:57.599
So which vendors meet those success criteria?
00:02:57.759 --> 00:03:04.400
And then we can make a recommendation on which large language model solution do we think is most applicable to us here at Galwood?
00:03:04.639 --> 00:03:06.479
Yeah, George, all sounds really logical.
00:03:06.719 --> 00:03:09.599
Just out of interest, how did you run the actual questionnaire?
00:03:09.840 --> 00:03:19.360
We ran it through Microsoft Forms, mainly because people know how to use it and it gives us a structured Excel output that we could then synthesize and move forward with.
00:03:19.520 --> 00:03:20.319
Yeah, it makes sense.
00:03:20.560 --> 00:03:22.800
The first step was sending out the questionnaire.
00:03:22.879 --> 00:03:27.120
Now, the questionnaire was loosely based around these six key areas.
00:03:27.280 --> 00:03:30.240
The first being who you are, you role and current AI use.
00:03:30.479 --> 00:03:32.479
The second is how you use AI.
00:03:32.560 --> 00:03:37.360
So what frequency do you use these tools at, and how sophisticated are you in using them?
00:03:37.520 --> 00:03:40.319
The third is how much you trust those tools.
00:03:40.560 --> 00:03:50.240
So if you're using them, how much manual work is required to tweak the output into a position that you can then move forward with it?
00:03:50.639 --> 00:03:52.080
Fourth is where your time goes.
00:03:52.159 --> 00:03:59.680
So this slightly comes away from the AI specific side of things, but where are the bottlenecks in your week?
00:03:59.840 --> 00:04:05.599
What do you spend the most amount of time on and where can we potentially target an AI solution to give you back some of that time?
00:04:05.840 --> 00:04:07.280
The fifth is what you handle.
00:04:07.439 --> 00:04:11.280
So data security, what kind of data are you handling?
00:04:11.360 --> 00:04:19.759
And that will inform our decision on how secure these systems have to be to make sure that that data is dealt with appropriately.
00:04:20.000 --> 00:04:22.879
And the sixth is appetite and constraints.
00:04:23.120 --> 00:04:33.759
So more of a an attitude question around how open you are to changing the way that you work and what do we need to get right to make sure that this project is a success.
00:04:34.000 --> 00:04:36.399
Yeah, it's all sensible stuff, isn't it, George?
00:04:36.560 --> 00:04:41.360
I wondered, was there anything else that you wanted to capture but but didn't?
00:04:41.759 --> 00:04:52.959
I think we've covered this across the board, but one of our primary concerns when rolling this out was we didn't want everyone to spend a lot of time answering the questionnaires.
00:04:53.120 --> 00:04:57.600
We wanted the key consulting-based questions, which is what we included.
00:04:57.839 --> 00:05:01.759
We're not bothered about what AI you use to kind of record your shopping list, that kind of thing.
00:05:01.839 --> 00:05:02.480
That's irrelevant.
00:05:02.560 --> 00:05:09.759
We wanted to keep it as concise as possible and stick to between 10 and 20 minutes at the responses overall.
00:05:10.000 --> 00:05:10.079
Yeah.
00:05:10.319 --> 00:05:12.399
AI for your shopping list, that sounds brilliant.
00:05:12.720 --> 00:05:17.040
Yeah, I'll have to say I need some after the podcast, I need some details on that, please, George.
00:05:17.199 --> 00:05:17.759
Yeah, I'm sure.
00:05:17.839 --> 00:05:18.319
No problem.
00:05:18.480 --> 00:05:19.040
No problem.
00:05:19.519 --> 00:05:23.920
Questionnaire came back with a range of answers, but the key themes we've labelled here.
00:05:24.079 --> 00:05:34.560
Everyone's using AI tools at work currently, and there is an appetite or at least an openness to change the way that people work, to feel the benefit of these tools.
00:05:34.879 --> 00:05:38.000
60% are using AI either weekly or daily.
00:05:38.160 --> 00:05:40.319
I'm sure Adam, you you understand that.
00:05:40.480 --> 00:05:45.199
One of the most important parts with any sort of tech implementation is adoption.
00:05:45.360 --> 00:05:53.839
And starting from a position where people already use these tools certainly helps get them on board with using what we're trying to put forward.
00:05:54.079 --> 00:05:54.879
Yeah, absolutely.
00:05:54.959 --> 00:06:02.160
You know, if if you're already using these types of tools, as many people in professional services and consulting are, you're pushing against an open door.
00:06:02.319 --> 00:06:03.360
Yeah, absolutely.
00:06:03.600 --> 00:06:08.560
Something that surprised me coming back was the fact that five chatbot brands are currently in use.
00:06:08.800 --> 00:06:23.519
Admittedly, I was expecting it to come back saying that people have dabbled with ChatGPT but but haven't really explored some of the alternatives, which was a pleasant surprise because it means that our consultants aren't necessarily just going with the mainstream option.
00:06:23.680 --> 00:06:26.319
They are choosing the tools that work for them.
00:06:26.480 --> 00:06:30.480
One of the main questions that we asked was where time is being lost.
00:06:30.720 --> 00:06:36.879
What came out on top was turning meeting notes into findings and meeting prep notes and follow-ups.
00:06:37.120 --> 00:06:42.319
A hundred percent of the team hit at least one of these five tasks weekly.
00:06:42.480 --> 00:06:47.040
So it was clear what the team's success criteria were for using a large language model.
00:06:47.439 --> 00:06:48.480
Yeah, that's that's really good.
00:06:48.560 --> 00:06:51.759
And I see the top one there is turning meeting notes into findings.
00:06:51.839 --> 00:06:57.360
I know as a consultant I've spent many hours scribbling meeting notes and then trying to make sense of them afterwards.
00:06:57.439 --> 00:06:59.920
So I can see that's very much a key requirement.
00:07:00.160 --> 00:07:00.879
Yeah, absolutely.
00:07:00.959 --> 00:07:03.680
And and it can often take you out of the meeting as well.
00:07:03.839 --> 00:07:06.959
You want to be focusing on the conversation, not recording it.
00:07:07.040 --> 00:07:12.319
And your notes are rarely to the level of depth that the conversation actually has.
00:07:12.800 --> 00:07:16.000
So a tool that automated this was a key priority.
00:07:16.240 --> 00:07:17.120
Yeah, makes sense.
00:07:17.360 --> 00:07:19.360
One of the other key findings was risk.
00:07:19.519 --> 00:07:22.879
So everyone in the business handles sensitive material every week.
00:07:23.040 --> 00:07:32.319
And that stems from client confidential material to NDA protected material, deal material, and then Garwood internal protected information.
00:07:33.040 --> 00:07:36.720
Yeah, I mean, it's really important, isn't it, your risk posture as a business.
00:07:36.959 --> 00:07:46.879
A lot of people don't realize that when you're using personal free AI platforms, those platforms are taking the data and information that you're uploading and using that to train their models.
00:07:47.040 --> 00:07:56.079
So what people don't realize is if you're using the free versions of things like ChatGPT, um, the model is by default training on the data that you put in.
00:07:56.160 --> 00:08:06.959
And it's not beyond the realm of possibility that a bad actor might try and invert the model, which means using very specific prompts to try and get that data back out of the large language model.
00:08:07.120 --> 00:08:09.279
So you do really need to be careful with this stuff.
00:08:09.519 --> 00:08:10.160
Yeah, absolutely.
00:08:10.240 --> 00:08:14.720
And it's not just free accounts either, even some individual paid accounts.
00:08:14.879 --> 00:08:27.040
While there is a toggle to turn off the fact that the model is trained on input data, you do have to be proactive about setting up your account so that it doesn't train the model.
00:08:27.199 --> 00:08:29.279
So it's it's certainly one to be aware of.
00:08:29.519 --> 00:08:34.720
Once we've received the questionnaire feedback, we identified five key themes emerging from it.
00:08:34.960 --> 00:08:42.720
The first was meetings into findings, second was knowledge retrieval, proposal production, tool consolidation, and governance.
00:08:42.960 --> 00:08:43.120
Yeah.
00:08:43.279 --> 00:08:47.519
George, we've got a combination of qualitative and quantitative inputs here.
00:08:47.679 --> 00:08:52.639
How did he use the free text inputs into the questionnaire to get these outcomes?
00:08:53.039 --> 00:09:03.039
A mix of qualitative and quantitative information was important to give us the fullest picture of how people are experiencing AI currently.
00:09:03.519 --> 00:09:17.039
We included open text boxes, particularly around the constraints and things we need to get right to make this project successful, because we didn't want to force people to choose from a predetermined set of options.
00:09:17.200 --> 00:09:22.080
We wanted to understand what people's success criteria are for the project overall.
00:09:22.720 --> 00:09:24.960
Synthesizing it was fairly easy.
00:09:25.120 --> 00:09:33.279
We were able to use a clawed project, and any large language model is good at taking unstructured input and structuring it.
00:09:33.679 --> 00:09:39.200
And we were able to use that in conjunction with the structured output that we got from the task ratings.
00:09:39.840 --> 00:09:47.440
In order to determine how we went out to market, we needed to convert those themes into specific success criteria.
00:09:47.759 --> 00:10:03.679
We came up with seven: enterprise grade data handling, Microsoft 365 access, workspace and access control, long document and transcript analysis, meeting transcription, persistent workspaces or projects, and a credible integration path.
00:10:03.919 --> 00:10:06.240
Yeah, I mean these all look pretty critical.
00:10:06.399 --> 00:10:12.159
What would you say if you had to pick out two or three as really, really critical, which ones would you pick?
00:10:12.480 --> 00:10:21.360
Enterprise grade data handling is certainly one of the most important due to the nature of the information that we use on a day-to-day basis.
00:10:22.000 --> 00:10:28.159
We needed to make sure whichever large language model we opt for is UK GDPR compliant.
00:10:28.320 --> 00:10:36.240
And we needed a signed data processing agreement that confirms that the LLM won't train on inputted data.
00:10:36.639 --> 00:10:41.039
Another important one, perhaps rather selfishly, is meeting transcription.
00:10:41.120 --> 00:10:46.639
So I was taking a lot of notes in meetings prior to rolling out our transcription tool.
00:10:46.879 --> 00:10:56.240
And I would often leave meetings with smoke coming off my pen and not fully capturing the context that was discussed in the meeting.
00:10:56.480 --> 00:11:01.679
So that was something that was incredibly important in whatever solution we decided on.
00:11:02.000 --> 00:11:04.080
But also Microsoft 365 access.
00:11:04.240 --> 00:11:09.360
One of the barriers to adoption would be completely changing the way that people work.
00:11:09.679 --> 00:11:13.120
People are working in Microsoft 365, so the tool has to as well.
00:11:13.679 --> 00:11:13.919
Great.
00:11:14.480 --> 00:11:21.279
Once we had the success criteria, we needed to identify how we're going to assess vendors against it.
00:11:22.000 --> 00:11:32.639
We adopted a structure of starting with a Claude project, which took the scorings, took the free text, and synthesized it into a prompt that we could then go out to market with.
00:11:33.039 --> 00:11:44.639
We then pushed that through perplexity to identify the vendors that met those success criteria, which ones they kind of met but not fully, and which that they didn't hit at all.
00:11:44.799 --> 00:11:50.240
We then came up with a short list which showed three credible providers, which we assessed in a little bit more detail.
00:11:50.799 --> 00:11:59.360
George, just um help me realize because I'm a bit of a cynic, if Claude was involved in the research, how do we know it didn't just select itself?
00:11:59.759 --> 00:12:01.600
Yeah, it's it's a good question.
00:12:01.759 --> 00:12:06.159
So the Claude project that we used was to synthesize the findings.
00:12:06.320 --> 00:12:11.519
We didn't use Claude itself to go out to market and recommend vendors.
00:12:11.759 --> 00:12:16.559
We don't want any sort of kind of favoring of one large language model over another.
00:12:16.799 --> 00:12:18.000
So we use perplexity.
00:12:18.240 --> 00:12:21.919
The benefit of using perplexity is that it's model agnostic.
00:12:22.080 --> 00:12:30.320
Its deep research function uses a range of models from different providers, collates the responses, and gives us the most objective view.
00:12:30.559 --> 00:12:37.279
Once this was completed, we had a report which confirmed three platforms cleared the bar.
00:12:37.519 --> 00:12:47.200
We assessed them specifically against those success criteria, and it's clear that some performed better than others, and that fueled our decision on which LLM we would choose.
00:12:47.440 --> 00:12:54.960
I think it's important to say, isn't it, George, that this was done probably late May, early June 2026 for those listening later on.
00:12:55.200 --> 00:12:59.600
The world has changed hugely since that because things move so fast in the world of AI.
00:12:59.679 --> 00:13:06.399
And for example, ChatGPT didn't have those Microsoft platform plugins when we did this assessment, but they do now.
00:13:06.639 --> 00:13:09.440
So it's important to know that this is a point in time.
00:13:09.519 --> 00:13:19.120
And it's probably, given this pace of change within AI, worth doing in a light version of this exercise every quarter or so to make sure that the platform you select meets your needs.
00:13:19.440 --> 00:13:25.759
And you open the opportunity to do that with the infrastructure that you set up as well, which we'll touch on later.
00:13:26.000 --> 00:13:31.600
If you build your AI infrastructure well, you can switch the model that is being used.
00:13:31.679 --> 00:13:35.279
And it's always good to have that approach as models change all the time.
00:13:35.440 --> 00:13:40.320
And at the moment, it feels like ChatGPT and Claude are constantly one-up in each other with every model.
00:13:40.879 --> 00:13:42.799
So it's good to have an open mind on that through.
00:13:43.039 --> 00:13:47.200
But at the time we conducted the research, Claude team did come out on top.
00:13:47.440 --> 00:13:50.399
And there are three main reasons for why that's the case.
00:13:50.559 --> 00:13:56.159
So it's best where work is hardest, meetings to findings, proposal production, knowledge retrieval.
00:13:56.320 --> 00:14:04.799
Its integrations are more developed, and we're able to tap into the information that we need to to create draft deliverables most efficiently.
00:14:05.679 --> 00:14:08.720
Similarly, it reads Microsoft 365 natively.
00:14:08.799 --> 00:14:18.799
So it has a first-party MCP connector which reaches SharePoint OneDrive Outlook Teams, which provides Claude with the context it needs to move forward with Glide.
00:14:19.120 --> 00:14:20.960
And also it keeps the firm portable.
00:14:21.120 --> 00:14:34.399
So Floud released the model context protocol in 2024, which is an open standard that third-party integrations can adhere to to make sure that their apps can connect directly with Claude.
00:14:34.799 --> 00:14:39.519
Okay, so that's the MCP acronym I hear being banded around all the time.
00:14:39.840 --> 00:14:41.679
That means model context protocol.
00:14:41.919 --> 00:14:42.320
That's it.
00:14:42.480 --> 00:14:44.879
Okay, like an API for your AI platform.
00:14:45.200 --> 00:14:53.440
You can still use APIs for AI platforms, but yes, it's an open standard that third parties can adhere to make sure that they're able to connect with Clock.
00:14:53.759 --> 00:14:54.000
Right.
00:14:54.240 --> 00:14:58.399
We were then able to flesh out our AI infrastructure overall.
00:14:58.559 --> 00:15:07.840
So you can see that we've selected Claude team, and above it sits the Microsoft 365 content that fuels the bulk of the context that we're feeding into Clock.
00:15:08.320 --> 00:15:11.759
This covers SharePoint, OneDrive, Outlook, and Teams.
00:15:11.919 --> 00:15:17.919
And we also selected a meeting transcription tool to give us even more context from meetings.
00:15:18.159 --> 00:15:20.799
I know a lot of our viewers and listeners will be interested.
00:15:21.039 --> 00:15:24.080
Fellow is our meeting transcription tool.
00:15:24.320 --> 00:15:27.039
There's lots of other meeting transcription tools out there.
00:15:27.200 --> 00:15:34.399
Of course, Teams Native, things like Fathom, Otter, Fireflies, Granola is one that I hear coming up a lot.
00:15:34.559 --> 00:15:36.080
What made us choose Fellow?
00:15:36.960 --> 00:15:50.960
We did analyze whether we could use the inbuilt Teams transcription and have Claude read that directly, but it would require an automation that pushes that transcription from Teams into SharePoint, which was quite a lot of work.
00:15:51.120 --> 00:15:57.600
We then started evaluating third-party options, and Fellow came out on top for a couple of key reasons.
00:15:57.919 --> 00:16:02.960
The first is it offers both bot and botless recording.
00:16:03.200 --> 00:16:09.840
And for anyone who doesn't know, a botted recording is when a bot joins your call.
00:16:09.919 --> 00:16:17.840
And that's particularly effective because the quality of the transcription a bot can do is much better than a botless approach.
00:16:18.240 --> 00:16:32.639
Within Teams, it's recording everyone's audio individually, so it can attribute speakers to their speech and it allows it to not get bogged down by crossover speech as a botless approach would.
00:16:32.799 --> 00:16:36.000
Now that's not to say a botless approach doesn't come in handy.
00:16:36.240 --> 00:16:45.840
If we're conducting in-person workshops, for example, we would need to put a microphone in the middle of the room, and it would be more difficult to attribute speakers.
00:16:46.080 --> 00:16:52.799
The fellow gives us an option no matter the circumstances and allows us to store all of that meeting context in the same place.
00:16:54.720 --> 00:16:55.679
What does it unlock?
00:16:55.919 --> 00:17:01.840
We've already spoken about the fact that our AI infrastructure is essentially a context store.
00:17:02.000 --> 00:17:10.240
Every meeting, every document, email and chat, and the firm's IP are all available and can be referenced by Claude when creating documents.
00:17:10.559 --> 00:17:11.599
What are the actual outputs?
00:17:11.759 --> 00:17:13.519
We're able to find information faster.
00:17:13.759 --> 00:17:17.200
We're standardized in draft outputs and quality assurance.
00:17:17.519 --> 00:17:21.519
And we're giving ourselves the time to really focus on the higher value work.
00:17:21.680 --> 00:17:26.160
Within consulting generally, the value proposition is our expertise.
00:17:26.319 --> 00:17:32.480
And a lot of the time we can be bogged down in the manual elements of the role that AI can now solve.
00:17:33.759 --> 00:17:36.000
Yeah, I totally agree with that, George.
00:17:36.079 --> 00:17:45.599
And in the relatively short time I've been at Garwood this second time around, thanks to the infrastructure put in place, I really have utilized AI in a way that I hadn't before.
00:17:45.759 --> 00:17:47.440
For example, writing decks.
00:17:47.599 --> 00:17:50.880
I've not created a deck from scratch since I've been at Garwood.
00:17:51.039 --> 00:17:58.160
My time is better spent synthesizing and inserting the context into the platform.
00:17:58.319 --> 00:18:04.799
Claude pushes out perfectly branded documents from first instance, so I don't have to worry about the latest marketing template.
00:18:04.880 --> 00:18:10.799
So it really saves that sort of time that people don't necessarily think about, but was a time suck in days gone by.
00:18:11.599 --> 00:18:13.680
What are we going to talk about in the next episode?
00:18:14.000 --> 00:18:17.200
Next time we're going to be looking at the boring but essential half.
00:18:17.359 --> 00:18:19.920
So governance, how do we manage these systems?
00:18:20.160 --> 00:18:24.640
Acceptable use, what policies do we need to have in place to protect the business?
00:18:24.799 --> 00:18:26.480
And training individuals.
00:18:26.640 --> 00:18:30.400
How do we make sure that everyone can get the most out of their IOTs?
00:18:30.799 --> 00:18:31.279
Brilliant.
00:18:31.440 --> 00:18:32.319
Thanks, George.
00:18:32.640 --> 00:18:36.160
Just to our viewers and listeners, I've been Adam Mays of Garwood Growth.
00:18:36.400 --> 00:18:37.680
I've been George White.
00:18:38.160 --> 00:18:39.599
See you next time.
