Hands-On AI: Turn Gainsight & Staircase AI into a CS Workflow Engine

[Speaker 1]
Okay, welcome everyone, and thank you for joining the session. First of all, I just want to introduce myself. Again, my name's Matt Smith.

I work at Accepta, and also my colleague Cal here. He works in operation via Accepta. He's my beautiful assistant today.

He'll walk around and provide the microphone if anyone has any questions, okay? So, quick housekeeping before we get started. This is a workshop, so I want it to be more practical.

So, I'm gonna try and get through the initial slides as quickly as possible, so we can actually get down and get into actual the practical side of it. Myself and Cal will come around, answer any questions, and try and help people along as we go. But feel free to stick your hand up for any questions.

I'll put it on the slider for any questions, and we can answer at the end, okay? So, you've got the setup here. We've got the paper there for anyone that hasn't got GameSite MCP setup.

So, you can just take this as a takeaway for when it's set up later on. But if you have already get it set up for your laptop, please just use that as is right now. That'd be great.

So, before we go forward, I just want a quick read into the room. So, we've got the first Slido poll coming out now, which is, how are you currently using AI in your CS workflows? So, single sort of answer on this one.

So, if you take a bit of time to go through that, and vote on it. Okay, it looks like most people are in the middle of the spectrum. Not at the very beginning, where it's like not done anything, and not at the very end, where everything is automated, and it's a perfect utopia.

So, that's good. Just wanted to say that this is more to start, really. This is more to the business users that are getting started, and sort of hints and tips to get started.

So, if you're in the last session, that was a bit more advanced for more of the operations users. This is more geared towards, say, CSMs and business users that want to start using this, and say, Claude. So, most CS teams I know of are more of the lower end of the scale.

They're at this part here, where they're using AI as a chatbot. There's nothing wrong with that. You do get value from using AI as a chatbot, and it's a perfectly reasonable starting point, but you have a ceiling on that.

So, every time you start a new conversation, you're starting from zero. There's no memory on that. You're manually copying health scores out of game site, or whatever it is you're copying, pasting them back in, asking a question, copying about.

It's still a lot of manual work, regardless. It's useful, but it doesn't really scale. Where we want to be, really, is on that right-hand side, as AI as a workflow engine.

So, it has access to your game site data, staircase, if you have it, that data, the relationship signals, account context, and it can act on it, not just describe it, and that shift towards what MCP makes possible. Now, there's that word MCP. What is it, really?

So, again, if you're not aware, it stands for Model Context Protocol, but you don't really need to remember that. What you need to understand is what that actually means in practice. So, instead of acting as a middleman between Claude, which is basically the AI of choice that we use, but it could be anything, and game site, MCP creates a direct connection.

So, Claude queries your game site. It reads the health scores, pulls CTAs, checks your timeline entries, et cetera, and when you've got the right setup, it can also write it back into game site. So, it's not just read-only.

You're writing the results back in as well. Staircase also has an MCP connector, but just something I wanted to mention today for when we're doing our practical tests. Staircase, it just connects into prod, and for confidentiality, et cetera, we're not doing that today, so it's just, I'm going to talk about it in theory, if that's okay.

So, again, that there pulls relationship intelligence, sentiment, stakeholder engagement, risk signals. Think of that as more of the sort of sentiment-based stuff where game site is sort of the hard facts and data, and it works together to give you a bigger picture. So, what I want to do now is just sort of show you just very briefly what a basic skill can do.

One of the things that the CSMs can ask for is give me a prequel briefing for X. I've just called it test account here for the sandbox data we're just about to see. It's got realistic but invented data, so nothing I'm going to show is confidential.

It's stuff I made up. It's going to have some names, like my name and Alyssa's name on it, but the point is here, it's just going to land. It's not specific output.

It's just a mechanism, okay? Use one prompt, get the data, and it pulls it from the CS platform. You don't have to switch tabs.

You don't have to do any manual copy-paste. All start from scratch, and it remembers the context, okay? So, with that, I'm just going to swap over to my Claude instance.

So, on this, what you'll see here is what I made earlier, as it were, because I wanted to make sure that if I run this now, it could take a few minutes to go through. I can do a live demonstration again in a bit, but all I asked at the very top, can you read that okay there? Do I need to zoom in?

Okay. And again, for a caveat, I had to, at the very beginning, give it a context to say, use the sandbox connector. So, forget the first part of that prompt.

It's literally just the end, which is, give me a prequel briefing for test account. So, I put that in, and then it goes through there, using the skills, the setup I've given it, and then it pulls back very useful data that can then be not only give you an idea of what's going on, but what is then actionable, okay? And that's the important factor here.

So, on here, it gives a very big, it gives a brief summary at the top, okay? It's almost like an executive briefing. It then visually tells you more about the current health score.

Lots of ambers and red on there. That's not good, but you know. And then important information about key people, recent history, what commits we got, and suggested outcomes, really.

So, from one skill, with a little bit of input, you have this consistent output that the CSMs can then just do very plain English prompt at the top, and then get actionable insights from. And like I was saying as well, if you allow it to then write back, you could then correct it, go, oh, I had a meeting with X on such date. Go check my emails.

Oh, yes, I see that. Would you like me to create a timeline tree for that? Yes, please.

Here is the additional context. So, you can do that as well, okay? Any questions at the moment before I move on?

No? Cool. Okay.

So, again, I mentioned it very, very quickly. What is a clawed skill, really? So, because doing that and showing you that just as it is is one thing, but then practically getting some integrity and trying it out yourself is something completely different.

So, what you want is something tangible that you can take away, that your whole team can run and run repeatedly. So, without needing to know what and right skill prompts from scratch, let's tell you, go through what it is. So, there are really four parts to a clawed skill.

It's essentially a safe set of instructions that it loads automatically when you use a keyword or phrase to invoke it. Think of it like a macro, but instead of logging keystrokes, it's automating the whole reasoning and data retrieval workflow process. And it has four parts, like I said.

The trigger is the phrase that fires it, and you can go through variations of what a CSM might write up to trigger that. On here, I've just said, give me a pre-pull briefing four, but it can be variations on that. And Claude is clever enough to know that it doesn't have to be exact language.

Anything around that, it will then fire that same skill. Second bit is the instructions. So, that's where it recognizes the pattern and loads the right, sorry, not that part, I won't remember that.

Instructions tell Claude exactly what to do. So, just like in the last session, you can keep it quite open-ended, go here and get data, go here, or you can say, only pull this subset of data from game site. So, that's the balance between how open you want to be, the more open you are, the more, less rigid it is, the more different answer you can get, but it can also be more token costly.

Whereas if you really sort of filter it down, you get less sort of inferred insights, but at the same time, it's something that's very repeatable, and you know that if you put it in three times, you're more likely to get the same answer three times. And then the calls there, the MCP calls are basically, it's just behind the scenes of how it goes into game site or wherever you're connecting to to pull the data back. And then finally, this is something that's important that most people forget, is the output format.

Unless you tell it the format you want it in, it's going to just infer how you want it. As someone said in the last session, they're more visual, so it says, please send me this in more graphs or visuals if you prefer it to be in a Word document. Say, output this in my company letter-headed Word document, please.

So you can tie skills together, okay? Okay, so having a look at this, this is just sort of a stripped-back version of an actual pre-call briefing skill that I just ran. And if you look at the agenda that I've got, I've got a link where you can download it and get a copy of that if you wanted to use that as a framework to build on as well.

As you can see, it's all in plain English. The trigger there is giving the pre-call briefing for add account. The instructions, what they do is they resolve the account.

They make sure they check the name is correct in game site. They pull the health score, the CTAs, the timeline, and then they also query Staircase AI for any sentiment and risk around that account as well. And then finally, I said to output it as structured briefing, context, actions, and talking points.

I'm telling it that's what I want, and I always want the output in that format. So, this is an important thing to remember. You don't need to be a developer to write these.

You can do it in plain English. They keep talking about prompt engineering as though you need to be an engineer. Anyone can do it.

It's just syntax and making sure that you're keeping things standard. But the hard part isn't the syntax. It's really thinking in a slightly different way from what you would normally do.

What workflow am I trying to automate? What does the data need to look like? And what does the output need to look like?

So, you're always going to be iteration here. You're never going to get it right first time, but the more you do it, the more you practice. Just like with any other skill, it becomes easier to do, and you get more effective at it.

So, without further ado, now it's your turn now. We're going to go over to the worksheets at your desk, or if you've got your laptop with you. It'll be sort of just discussing in your teams.

Think of, first of all, it's like, well, think of a problem you're trying to solve for your CSMs with the data you're trying to get in your organization. It's just, okay, well, there's a problem. How are we going to deal with that?

So, just to help with that, I'll leave this up here so you can have a look at that to help you along. So, pick a workflow that's manual that you keep doing you think is something that you have to do over and over again. It's laborious.

Oh, that seems like it could be automated. Use that as a starting point. Use those sort of guidelines on the sheet just to give you a bit of a framework.

Again, if you feel adventurous enough or if you've done this before, you can actually just go a bit off-piste, as it were, but just try to keep within the framework because keeping that structure means that the skill can be repeatable. And, as I said, if you have your laptop, go ahead and try and do it real-time. See how you get on.

Okay, if there isn't any other questions, I think Cal and myself are around. We'll pop around, see how you're getting on, ask a few questions of what you're working on, and you can ask questions of us. Okay, thank you.

Okay, we've got a few minutes left, so we've been wandering around the table seeing how people are getting on, and there's a mixture of where people are in their journeys and, more importantly, what IT let them do, really, with these skills and whether or not they can put these into production, et cetera. So, just wanted to sort of check through, and the next thing is, like, asking about what did you build. I think we've already gone through that, so we won't sort of go up the top here, but there seem to be a lot of common themes or problems that people had.

Most of it might have been, like, the actual scenario, what they're trying to fix, but, secondly, the other common problem was how do we actually deploy these skills that everyone can use, and that seemed to be a common theme, as well, that wasn't really part of this, but it seemed to be a key thing that we should think about when taking it away, really. Okay, just because of interest and time, I'm just going to go to the final Slido poll that we've got closed here. So, this had some canned sort of answers where I think what would be most useful.

So, which workflow would you find most useful for a pre-built skill? So, again, there's a Slido poll, and if we do have a lot of that something else that pops up, what I might say is we just get people just to raise their hand very quickly or shout out what that might be. Just for the interest of time, we've just got a little bit of time left.

Okay, I think I'll jump on then. It's like, it's a healthy mix. There is a couple of, like, for anyone that said there's something else, though, do they just want to shout out quickly what that something else was if they're happy to do so?

Cool, okay. So, very, very quickly, the key takeaway here is for a lot of people that are starting out, MCP's the missing link. Skills are good for making something repeatable and standardized.

You get the same sort of answers when you ask it several times. It can be user-developed, so you don't need a developer. It can be citizen developers.

It can be CSMs that can create this and always start small and work up from there. So, again, you've got physical materials you can take away today. But finally, just landing on any questions anyone might have.

I know we're at timer, unfortunately, so it might be that you might have to just come up and find me afterwards, but thank you. Thank you. Thank you.

Thank you, Matt. That was amazing. I hope everyone had a lot of nice interaction.