Transcript: Datasite MCP for ChatGPT

Doug Cullen (00:00)

Good morning, good afternoon, and good evening to those joining us around the world. My name is Doug Cullen. I'm Chief Strategy Officer here at Datasite, and it's my pleasure to welcome you to the Datasite MCP for ChatGPT webinar. This is the second in our series, where we walk through the opportunities now available because we've enabled and connected Datasite, the world's leading VDR, to various frontier model platforms via MCP. Today we're going to talk about ChatGPT, which is a fantastic platform.

Before we get going, I just want to remind everyone of a few things. Okay, first of all, we want to hear from you. Please ask questions in the Q&A panel throughout the session. We have reserved a little bit of time toward the end to get to those questions and answers, so please ask them. We also have a survey, and we'd love to know what you think. There are additional resources around the console as well, including some PDFs, documentation, and frequently asked questions.

Please check there for the resources. This session will be made available on demand, so if you have colleagues who weren't able to attend live, or if you want to replay it to understand some of the things that we demonstrated, please feel free to do so. And the last one is: who doesn't love a good legal disclaimer? I will read this one out here, just so you know: all opinions expressed are our own and not those of Datasite or Blueflame AI. So with that, Alice, thank you so much for joining. I'd like to introduce everyone to Alice Esmerian.

She's a product strategist at Blueflame AI, and I know you've been here for a few months. If you don't mind, please give the audience a little background on what you're doing at Blueflame AI and what you did prior to joining us.

Alice Esmerian (01:53)

Sure. Thanks for having me today. Yes, I joined Blueflame six months ago on the product strategy team. My background is in investment banking and private equity, where I spent almost eight years. It's great for me to have the opportunity to work on products that really enable investment bankers and private equity professionals to leverage AI in their workflows. I'm super excited to be here today and talk about a few workflows around the MCP product release.

Doug Cullen (02:25)

Awesome. We want to get into this; it is definitely show, not tell. There is a real step-change moment happening in dealmaking. I've been in dealmaking for almost 20 years now and have been watching a lot of the transformation across how deals get done. I run corporate development, and we've done nine acquisitions in the last couple of years. I think this is the greatest time to be a dealmaker, when you think about what we can take advantage of with some of these new technologies.

But before that, one of the things we wanted to do is talk about how we launched MCP at DealMAX. This was a big deal for us here at Datasite, allowing the first-ever connectivity from a data room, our Datasite Diligence platform, and our Acquire platform to MCP. We'll be demonstrating enterprise ChatGPT, one of our favorite platforms, here today. It's a new age, as we're talking about. Raj introduced the concept of a prompt-first dealmaker, which we think is amazing.

One of the things we also focus a lot on here is making sure this is done in a proper way. This is a serious business, right? We want to make sure that it's offered in a compliant, resilient, and effective way because that matters a lot in deals.

Alice Esmerian (03:45)

Yeah, totally. We're seeing it a lot today with our clients at Blueflame. When I joined six months ago, I could see that people were still experimenting a bit with AI. They were trying to understand where it sat in their organization, and which workflows were applicable or not.

Today, things have changed. Every organization is really trying to deploy AI consistently across the firm and across workflows. It is really important that we enable them to power AI while staying totally safe and maintaining all compliance and governance that are essential to executing a deal safely. The data they are dealing with is so important and valuable that we really empower them to do that. It is very exciting to leverage this MCP connector to bring this safe AI connectivity into Datasite for the first time.

Doug Cullen (04:36)

Yeah, it's an incredible time. Let's transition here quickly. We want to think about how AI is actually used. We want to demonstrate some of that and talk about it securely inside the data room. This is one of the most important things. We launched the Blueflame AI Assistant within Diligence a few weeks ago, and that is the vehicle we use to connect MCP. We just wanted to ground some of the things we're talking about for the audience because this came up earlier this week.

I wanted to make sure there are a couple of core concepts. One is MCP. That's what we're talking about here. This is the connection. We've built an MCP server and published it via ChatGPT through our great partners there, which makes an available connection for any enterprise ChatGPT user to connect to and enable their data room. This is always going to respect the role you have, whether you are an admin or a reviewer, and certain things are going to be enabled via that MCP connection.

The other thing that really makes the MCP connection even more valuable is that the content within the data room has been AI-enabled. You're seeing here the Blueflame AI Assistant. If you have not seen it yet, within Diligence, our core platform, you have your standard Diligence platform with a small right panel and an icon in the upper right. You click that, it slides out, and it gives you the full agentic experience of Blueflame within the trusted environment of Datasite. So you've got the Datasite project, you've got the Blueflame AI Assistant running within the Datasite project, and then that project itself can be connected to via MCP in enterprise ChatGPT.

I just wanted to anchor those concepts. We'll be showcasing the MCP connection today because that is the purpose and the new innovation. But to reinforce and underscore it, that is powered by the Diligence project that has the Blueflame AI Assistant enabled. Then there is Blueflame, the enterprise company, which is used by private equity firms, investment banks, and corporate development executives that have a separate Blueflame subscription. These are the three dimensions we want to talk through. This came up on our first webinar, and we wanted to give a little more grounding in terms of what we're going into.

With that, we can probably shift into the show, not tell, aspect of this and highlight and talk through some of the powerful use cases that a platform like ChatGPT brings to bear. One of the most challenging things when you get going is that you have all this content and are trying to figure out a way to set up the data room. You've done it many times in your life. What we're going to showcase here is probably the most profoundly efficient and effective way, as an administrator, to create and set up a VDR.

So why don't you go through a little bit of what we can do with ChatGPT?

Alice Esmerian (08:11)

There are plenty of things we can do today. We're going to run through four workflows that are essential to the admin work within the data room. There are plenty more workflows that are applicable. If you have any questions, ideas, or suggestions, feel free to add them to the Q&A section. We'll look at them afterward. Today we'll look at four different workflows. On top of delivering this MCP connector and Blueflame enablement, we've built eight skills, which are essentially, for those not too familiar with this concept, almost like eight chef's recipes that help the model be even more empowered with additional guidance on how to understand queries around data room-specific workflows.

These are going to be available at the GitHub link that will be shared with you in the resources for this webinar. I just wanted to preface that because these skills are going to power some of the workflows we're going to look at today. The first one here is a simple one. When we go back to this dealmaking prompt, you can see that we're not starting with anything sophisticated. We're essentially asking the data room, "Can you start this project for me?" and providing it with a bit of context.

Models tend to operate better when they get a bit of context around the task you're asking them to carry out. Here, we're just setting the stage with a bit of information around the company, its industry, and its size. Immediately after asking for that, the model is going to proactively share an index suggestion, which is powered by our indexing skill. Here, I'm getting a detailed index with subfolders as well, and I can immediately provide some feedback around it. You can see how a task that would have taken a lot of time, whether putting together a spreadsheet index or just reorganizing folders and naming them while making sure everything was clean and had zero typos, is done from a simple prompt in a few minutes.

Doug Cullen (10:16)

Yeah. This is something where dealmakers are very often baptized by fire. They may or may not have ever set up an index. What does a good index look like? That is so paramount to the success of a deal because one of the things that I believe really reflects the quality of the asset and the quality of the advisors that people have hired is the way the information is reflected about the underlying company. It's that adage about a tidy house: you really want to make sure everything is tidy, easy to find, easy to locate, and follows a very logical setup.

This is just a super powerful capability that AI can help us with, and ChatGPT in this instance can just give us a great, clean index.

Alice Esmerian (10:59)

Yeah, absolutely. You enter, and you automatically have a great first impression of what you're going to experience within this data room. The index is set up. Here, I made some suggestions for editing given the particular context of that company. The model automatically understands it, and I can then just ask it to push it. The model comes back in 53 seconds.

You can see on the screen that my 14 folders and all the subfolders have been pushed to Datasite. That is a massive time save. That then gives me a lot more time to do everything that is more valuable for me as an admin creating a data room.

Doug Cullen (11:40)

Yeah, I mean, it looks kind of magical. It is kind of magical to get that content and put it in that very logical place with a logical order. Now I think we're going to go into maybe some file naming conventions.

Alice Esmerian (11:55)

Yeah. Typically, when you start your data room, you have your index set up, you can push your documents, and then you just want to make sure that every document is clearly labeled. There are two ways you're going to do that. One is looking at the existing name and cleaning it up, which is totally doable with just the MCP before you enable Blueflame, because the MCP connector has access to every single file name. For example, if the company we're working on here is called Voxel Matter, and it sees capital letters, no capital letters, VM instead of Voxel Matter, or just clunky names, it will be able to clean that up, clean the typos, etc.

What it's not going to be able to do until you enable Blueflame AI is actually access the content of the document and leverage that content to inform how to rename a document. A good example here is that I wanted all my contracts to have the year in which those contracts had been signed in the name so that I could clearly orchestrate the content. To do that, you obviously need to enter a document, read it, and pull the year into the name, which the model was able to do because this data room is empowered with Blueflame AI.

Doug Cullen (13:10)

Yeah. A couple of steps here, just to rewind the tape. One is to get the name of the document consistent, which in and of itself doesn't always happen. In many a data room I've gone into, I've seen the famed Book 007.xls. What the heck is that? Who knows? Is it a financial model? Is it part of a customer queue? We have no idea. So getting the name into a logical place is super important.

As I'm clicking and doing the diligence, either in preparation before opening it up to the buy side or ultimately for my buyers, prospective buyers, lenders, or investors, I'm able to see what that document is before looking at it. Then we take it up a notch because we're actually able to look at the content of the file itself and read that content. In your great example, it's, okay, I know this is a contract. Okay, great. It is a contract. Well, what type of contract is it?

We're using the ability to look at when the contract itself was executed and apply that as part of the naming convention.

Alice Esmerian (14:23)

Again, as someone working in a data room, it's very time-consuming to do this, and it's not highly rewarding because you're really just searching for information and spending a lot of time checking for typos or inconsistencies. This is something you can automate so that you can spend more time on more valuable and informative tasks to get your process ready and prepared in the best form possible.

One thing we potentially wanted to highlight for people watching is that when we say Blueflame enables you to access information within the document, there is absolutely zero information retention.

This is why Blueflame is particularly powerful and safe in a deal context, because highly confidential information will never be retained by Blueflame.

Doug Cullen (15:12)

Yeah, no retention and no training, right? In the early days, people were really worried about this highly sensitive content being put into LLMs to allow the LLMs to be trained. There is nothing like that going on here. We are a financial services powerhouse, and we understand that the content is highly confidential. So there is no training on the content and absolutely zero retention of any content that is made available via this platform.

Alice Esmerian (15:49)

Yeah. Moving on to another workflow that honestly cost me a few hours and nights back in the day is a Q&A process. I think this is a great addition and a great way to leverage access to the content of your documents. In my example here, I attached a list of Q&A questions. This could happen early in your process, when your VDD providers start asking questions and making information requests around the asset. Later on, when you start receiving plenty of questions from different buyers, their advisors, and everything comes to you in an unstructured manner.

You can then just drop in the questions. Here, you can see I attached an Excel file with unstructured data. The questions are not organized either by theme or by advisor or buyer. Then I'm just asking the model to prepopulate draft answers to all of these questions. This is also powered by one of the skills that we're providing. Something I'm very adamant about in everything we do with Blueflame for our clients is making sure we have zero hallucination. This skill prompts your agent to really look at the context and the content and make sure it provides an honest opinion about its ability to answer the question, as well as a very detailed VDR link and the section of the document that it used to prepopulate the draft answer.

To me, again, this is a very good way of trusting AI safely and having a partner that works with you hand in hand, helping you save time and direct yourself in the VDR while having all the guardrails that enable you to trust and feel safe while using it.

Doug Cullen (17:35)

I actually had the benefit of showcasing this to a managing director here in New York for a very large investment bank, and he was reflecting, "This used to crush my weekends." The availability of taking questions, essentially uploading a set of questions and allowing them to be answered by the content of the data room, is powerful.

Questions are being asked, they're being uploaded, and they're being answered with cited content. We are literally putting the citation to the document that supported it, as well as, just to reiterate what Alice said, a bit of a confidence aspect as well: a confidence score. For some questions, the model is going to be highly confident in its response based on the document and the framing of the content. For others, it will not be as confident. We're able to discern these two categories to narrow in, from a preliminary standpoint, before the asset goes live. You want to understand what questions people are likely to ask and make sure you have the appropriate information in the data room in order to answer them.

And then, of course, when you're live, the orchestration of Q&A is probably one of the most complicated tasks we have when doing a deal. You've got various questions, various stakeholders, and each person has their own series of experts: HR, legal, compliance. We need to orchestrate that question from the buy side to the sell side, to the expert, and back.

Alice Esmerian (19:16)

Totally. You can see here on the screen that the model took all the questions and then batched them by theme. Here on the screen, I'm showing you the commercial section. For each question and topic, it can provide a status, whether complete or partial.

For every draft response, you have the document or documents that it used to address it, as well as the page reference. This is then all aggregated by theme into a single spreadsheet that you will see here on the screen in a few seconds. It is organized nicely by workflow with the number of questions and their level of completeness. Then there is the total Q&A tracker, which, for each single question, includes a draft response, a status, the source reference, a citation, and a follow-up question required or additional data needed.

To me, as a banker, this is a great way to really start the Q&A process and accelerate it while keeping it valuable and high quality.

Doug Cullen (20:17)

Yeah, this frames the problem. It is super valuable for the person helping from the advisory perspective, whether that's a law firm, a bank, or another financial advisor, but it's also super valuable for the corporate. Having been involved in selling the company a couple of different times and selling different assets, as a senior executive, you're super curious about how the process is going and what types of questions are coming up. Being on top of it is the sign of a great advisor.

Being able to give this type of report to someone like me or Rusty, our CEO, is something I really can't overstate in terms of how valuable that is to the underlying corporate. That's the person behind the deal.

Alice Esmerian (21:02)

Yeah. Accelerating reporting and communication with your client, as you mentioned, is something that is now possible with this MCP connection because you can extract activity from the data room. One more thing I wanted to show today that ties to that is something I think a lot of people here will have some sensitivity to.

Publishing a data room for the first time, when you add buyers in, always comes with this little moment of anxiety when you wonder, "I really hope I didn't leave anything contradictory that doesn't tie, or just a document that won't open properly. Are there any risks? What are the risks potential buyers are going to flag immediately?" That's why we put together this skill called the Prelaunch Readiness Report. It's going to focus on confirming three main areas of your data room. The first is that it's going to run a gap analysis.

It will look at all of your sections, keep in mind the context it has around your deal, industry, and size, and provide feedback on how complete each section is. Are there any areas where you might be lacking substance or context? It will provide feedback around that. It will then also look at the quality of the document. It will run through whether any document is blank, any contract is unsigned, or any PII information has not been redacted. That could create a lot of risk around your data room.

Finally, it's going to run through all the potential risks that we provided in this skill around every single area of tax, finance, operations, IT, and IP. It will cross-check all of those and give you feedback on where your largest risks are and where buyers are going to pay the most attention. As an admin, that really gives you time to anticipate what you need to prepare in your defensibility approach and how you communicate this asset to potential buyers. All of that is then consolidated into a single Word document that you can download and share with your client and team to make sure everything is ready.

Doug Cullen (23:17)

I remember hearing a story from a managing director who described to me that every time before she launched a deal, she had to literally go through every single document. That would take hours because ultimately she was the one representing the asset and taking it out. Think about the value from a managing director perspective, a senior banker, or maybe even your analyst or associate who has to do this. I'm sure Alice's pre-readiness report was fantastic back in the day, but this is something that is just available out of the gates, both within the data room from the Blueflame AI perspective and in ChatGPT.

I mean, what a powerful tool for both, really. Everyone across the board, whether you're the corporate, a senior advisor, or someone on the deal team itself, can benefit. It is fantastic.

Alice Esmerian (24:08)

You have this extra pair of eyes, as if someone were really working with you while you're doing something else and reviewing the content for you. That's very exciting, and I really believe that everyone working on sell-side data rooms will find great value from using this.

Doug Cullen (24:25)

Yeah. I think this is changing a lot.

I was talking to someone just yesterday, a very senior banker, and he joked, "I actually may log into a data room again for the first time in 10 or 15 years." This has been a ritual among advisors where the junior teams tend to interact with this. But now, with natural language querying and the availability of these capabilities within the data room, you can log in, ask a few questions, and get that answer in your preferred platform. You can log into the data room or the Diligence project. You can log directly into ChatGPT, and you get that same immediate availability of key information and context. It’s incredible.

Alice Esmerian (25:06)

It is great. Although I think everyone is starting to get used to the speed of AI and ChatGPT, everyone is so concerned about accuracy and getting confidence around the content. This is really a great way to get in there.

Doug Cullen (25:22)

Yeah, this is an incredible capability. We're having a lot of fun. We're in the early days with this. Just to reinforce, this capability launched about three or four weeks ago. The feedback has been tremendous.

We're looking forward to a few questions here, which I think are great. One way to think about this, too, from a broader perspective, is that this is the worst the models will ever be at answering these types of questions.

Alice Esmerian (25:47)

Yeah, it's changing every day. We are keeping up with all the innovation here and making sure that our prompt-enabled dealmakers are getting the best of the tools.

Doug Cullen (25:57)

Yeah, you were saying earlier that you were going to leave today and test one of the latest models from ChatGPT. OpenAI has been great with us in terms of sharing its latest models, and we usually get our hands on those a bit early in a preview session. This is like your Christmas Day, right? You get to open up new presents.

Alice Esmerian (26:19)

Now that happens more than once a year, so it's great.

Doug Cullen (26:23)

Hey, let's take a look at some of the questions. We have a few here that I'm going to try to fly through.

I saw one earlier: what kind of audit trail is available on these actions? The answer is a full audit trail. We're able to capture everything that's been done, all the queries, and all the information, and that will become part of the reporting package that is available in Datasite. Let's see. Another one: how strong do you need to be with LLM prompting to use something like ChatGPT and Datasite?

Alice Esmerian (27:04)

It's going to depend, but honestly, the way we've built it is, as you've seen from all my prompts on the slide, fairly straightforward. As long as you're able to convey it the same way you would ask someone on your team to do a task, you can really just use that, and the model is going to do it for you.

One piece of advice I would give is that iteration is very powerful when it comes to handling AI products. Try something. If it doesn't quite work, refine your prompt. Nothing has been built in this MCP connection that requires any particular prompting knowledge. Any investor with no prior AI knowledge is able to leverage the most of it.

Doug Cullen (27:45)

Yeah, I mean, I think like everything with these LLMs, you just have to jump off the curb and try it. Your framing is really good: as you would speak with a colleague, as you would seek advice from someone else, seek advice and see what types of responses you get.

Then you have the ability to do things with follow-up questions, which are super important. Let's see. Is it possible to give buyer advisors, for example, limited access to ChatGPT output if you do not want to share full access? These are dynamics that we're thinking through right now. To reinforce, a lot of the capabilities you're seeing right now are really for administrators of the projects only. Do we intend to bring some of these capabilities to reviewers and buyers? The answer is yes. We've launched this, as I said, about three or four weeks ago.

We're still in the early days. These are administrative capabilities that we're demonstrating right here. There is a list of capabilities that you have no matter what your role is, such as asking questions like, "How many projects do I have?" They are all very well defined and articulated. Honestly, you can just ask ChatGPT, "What am I able to do with Project Dragon? What am I able to do with Project X-ray?" and it gives you an amazing response directly in there. Okay, last one here, and this is certainly worth reinforcing. My question came after reviewing the material fact sheets and FAQs you shared. It seems that the client's data does not leave the four walls of the Datasite environment. Exactly.

We took meticulous care to make the client content stay within that. I would describe this as the secure boundary and barrier that is Datasite, and all of this is still within that secure operating environment. The extension out to MCP also is governed and needs to operate by those same sacred rules. That is how it works. Anything else to add from that perspective?

Alice Esmerian (30:11)

No. Again, it's a semantic search that goes into Datasite without pulling anything from it.

This is how you operate in the safest way while leveraging AI in your data room with Datasite.

Doug Cullen (30:23)

Yeah. We've gone through all the security reviews, etc., as you might imagine, and we've taken meticulous care in bringing this capability into the incredibly secure operating environment that is Datasite. You trust us. We want to make sure we're handling this in the appropriate way, and we'd like to think that we are. With that, we're at time. Alice, thank you so much. It was amazing.

Thank you, ChatGPT, and our great partners at OpenAI. We do have a few other upcoming webinars. We're taking a little bit of a break as we approach Memorial Day here in the U.S. and probably bank holidays all around Europe and the U.K., but we'll be back with our next series. Pay attention to those. A reminder: this will be made available. Thank you for your Q&A. We did not get a chance to answer all the questions, but we will wrap those into some post-webinar follow-ups in the form of blogs as well as FAQs.

Look, it's never been a more exciting time to be a dealmaker. The capabilities available via large language models and, of course, via incredibly focused, almost vertical-specific capabilities from someone like Blueflame, brought into the trusted environment of Diligence, are really changing and transforming the way deals get done. We're getting to prompt-first dealmaking, and ultimately we want to make sure that we're delivering these results in a way that people can trust and depend upon. It is AI, so you have to double-check things before making them available.

Alice Esmerian (32:06)

It goes without saying: human in the loop, always. But it's really a way to accelerate your dealmaking.

Doug Cullen (32:13)

Thank you once again, Alice. Thank you to the hundreds of people around the world who attended this webinar series. We're so pleased to be able to bring not only the innovative technology, but also this practitioner perspective to help you explore the world of AI. With that, I will thank you and wish you a wonderful day. Thanks, everyone.