Anygraph AI Demo

Hi, I'm Bryan from Anygraph and we're a team of Harvard, Cambridge and Princeton engineers building the best AI application for audit and advisory services.

If you're watching this, you're most likely an auditor and you completely understand the pain and volume of work it takes to complete a client engagement.

This typically looks like thousands of client files with millions of rows of data, that all sum entirely to zero across hundreds of working papers.

Existing software solutions today have you consolidating and sanitizing data, to load it up into cloud platforms that are 10, 15 years old at this point that all see they have some kind of AI function but where none of them just do work.

And so introducing Anygraph here in the far left, we have your existing chat threads.

Here in the middle, this is where your agent lives and that's how it's going to do its work with you.

And on the far right, this will be your file system.

And as with any engagement, you typically start with your client files.

And so here you'd click choose folder and you'd sort of scroll and select your specific files.

yep, as you click through this, it completely initializes here on the right.

this typically starts with something like a general ledger cleaning.

it's as simple as dragging it over here as an example, clean this.

And I've got two different samples here we can take a look at while this starts to run.

One is a PDF format.

So here I have a couple hundred of pages of PDF.

There's no table structure and you typically have to upload this into some cloud software to process.

Similarly, I have an Excel version Yep.

And it's rendered here you can typically see this is what Excels look like out of SAP or ERP systems.

And here, you've got lots of additional rows at the top.

A lot of cells are sort of merged, descriptions are sort of double line, also merged.

And blank rows in between.

this goes on and on and on.

And it's so large it can't even fit on one sheet.

This goes on for multiple different sheets.

agent here as it's still working are a couple of interesting things to call out in the product and the model.

We've trained and developed very specific audit skills so that they can sort of process and understand these types of documents.

as this is going on on the back end, the model is actually extracting the data from the PDF via OCR and then using AI vision as well to cross compare and looking at the file structure to then try to rebuild all of this in an Excel.

and the benefit of this is because it is an AI agent, it's not limited by a sheer number of physical humans.

And so I can actually click new agent here and I can drag a completely different Excel and I can say clean this.

We've designed our application in the same way you'd expect a really senior auditor to work.

And just like how you wouldn't spend hours of your time explaining how to do a very basic task to a junior, the AI agent reading up on exactly what the files are, what's needed, and understanding that process.

And just like how you wouldn't want someone to go off and disappear for a week and produce something that's completely wrong, the AI agent so too will come back and say, hey, based on the understanding it has, and based on the context of the documents, this is what it wants to do.

And based on that, if all of this is okay.

Yep, that looks good.

Go ahead.

And here we have, after about 30 minutes for a PDF or about 15 minutes for an Excel, you can get a completely flat Excel file.

fully extracted, singular rows with all the metadata account description split out, and so on.

It can also produce something like this that has account summaries.

One huge pain point we understand from auditors is that these big clients can have up to 20-30 subsidiaries each with bimonthly general ledgers and that all balloons into 20 times 24.

That's something like 500 ledgers that auditors are spending three to four hours to sanitize per document.

In our product once we understand the file format and template of this one particular client, you can actually spin up 1020, 100 subagents to parallelize and run all of these sanitizations all at once.

and so at a baseline just from a data sanitization, data manipulation, extraction and analysis across general ledgers, ARs APs and all different kinds of JVs, we see that we can handle any and all mechanical data preparation.

We've also seen tremendous breakthroughs in doing highly technical, highly judgmental audit work.

And so here I have a fantastic example of showing some kind of IFRS disclosure checklist.

So looking at this, I can open up the files and here I have a annual report from a typical listed company.

And as this loads up here, Yep, it has all its financial data, right there at the bottom.

And based on the total financial statements and all its relevant notes, we have to go through and do a disclosure checklist.

And here once again, through using sub agents and parallelizing all this work at once, we can do all the different tabs of the different disclosure, checklist processes all at once.

And this is what we've been able to produce across all the workpapers.

So here we've got all the specific notes and disclosures It's going in and filling in, yes, no, or, not applicable, and extracting the specific sub notes, the pages, and the specific terms that it references to answer the questions.

Another example that we found huge, tremendous value to auditors is in our callover and casting.

So once again, taking something like an annual report, in this case CIMB, it's as simple as dragging this in and say, do call over, and casting for this.

And this is what the output looks like from our AI where it's gone through the whole document process across all the different types of categories and issues that we've classified and identified with the audit firms.

And, Yep, I can say let's look at the ones that are issues here.

Another big elephant in the room is how do we trust and audit the work outputs from AI?

And so I have a fantastic example here of an NRV Net Value Realization workpaper.

So here I have my MUS sample, I have my sample of all my different SKUs that I want to vouch for.

And here I've got my work paper where I need to go through and fill this whole document in.

And if I were to do this manually, I'd have to look at one specific line item and let's say that 8 by 10 screws and I have to go in and look at invoices, GRNs and DOs across all the different branches and client files and I have to dig up and open them.

nope, not here, Nope, not here.

And I'd have to open them one by.

One.

and instead of all of that, the way I interact with AI is as simple as saying drag this in.

These are my samples that I need to vouch for.

This is my NRV work paper template.

My My invoices are here.

As you can see here in the chat thread, it's pulling up all its institutional knowledge of how to do audit work as well as looking at the existing documents and trying to set this up for us.

In this case here it's pre processing all these documents, here it's spinning up multiple subagents four at the same time to list out and process all these different documents Into its raw data, that it'll then map back into the vouching process.

And voila.

So in just about an hour, we've got some work done with this work paper.

So looking at these, we had 100 different samples.

And in this particular data set, we had 50 different invoices, DOS GRNs across 10, 15 different branches.

So now at scale, this is like 1000 over files for AI to process.

Going through each of these, we've managed to find a total of 70 out of the 100, invoices that match against the selling price.

And with that, 60, one mapped with deals as well.

And this is what it looks like.

So it's gone through, filled in the entire working paper and everything here previously was completely blank here.

It's been able to put in the details there, gone through and updated its invoice references.

references.

These documents are also hyperlinked.

So if you click this.

Yep.

Okay.

It opens up the relevant file.

And if you see this, there should be snap on tool 8500.

You can open this up.

And it opens up the exact file as well and references it here.

and it's been able to do this across the entire working paper.

And the key outcome we really want with our auditors is to really feel like this product is closer to emailing a colleague as opposed to using some kind of AI tool.

And so with this, with the work done, I might want to go through and proceed to next steps.

Right?

And I'll say which ones are cost higher than nrv.

Give this to me in a, table.

For the ones without invoices or dos, give it to me in an Excel file.

I'll request this from my client.

And this as it goes along, will be able to process and do this next step as well.

In a nutshell, Anygraph is an AI agent that can jump into any client file with you and I'll go through understand immediately the client's scope of work and the nature of business and if necessary, ask intelligent questions.

As it goes through all its work, it's auditable every step of the way, down to each exact line of code and specific rows of data it's collecting its inputs from.

It processes all of this at, speed and at scale by parallelizing and running hundreds of sub agents to do multiple different specialized tasks all at the same time while you do your work as well.

and we've built multiple layers of AI coherence checks into the product by using coding scripts, using Excel formulas and AI vision to visually parse and understand the file formats, to triple check all of its work.

Today, as it stands, the product can jump in and handle about 80% of the audit engagement with you across all the different scopes of work as listed here.

that's essentially our product Anygraph.

Thank you.