AI is Nascent, New, Hard… and Inevitable. The “why” and “what” of AI have largely been defined… now it’s getting to the hard task of “how.” Most enterprises realize that the risk of not investigating and experimenting outweighs the risk of trying and failing, and have accepted AI as an inevitable part of business moving forward. But few have successfully implemented AI technology at scale to reach the promised land.
Enter Steve Rainey, Chief Innovation Officer, Tax at KPMG LLP. Steve has been a strong advocate for continually improving the company’s services through innovative approaches, including working with teams to reinvent how to execute services in the age of Artificial Intelligence.
HFS’ Chief Strategy Officer, Saurabh Gupta had the opportunity to interview Steve on his AI journey at KPMG’s Executive Symposium in Austin earlier this month. Here is an excerpt of their conversation about the scope, current progress, and future direction of an ambitious AI-driven journey for KPMG.
Saurabh: There is still a lot of confusion about ‘what is AI’ in the market. People talk about AI as one big monolith technology whereas in reality AI is an umbrella term for multiple technologies. Do you agree? What does AI mean to you?
Steve: That’s right. The way we think about AI is to take a variety of technologies and integrate them to analyze data, generate hypotheses, create insights, and develop probabilistic outcomes – rather than binary yes or no outcomes, which is similar to what our professionals do for our clients. AI has a lot of parts – not just one technology to it.
Please help us understand what are the specific AI-powered tax offerings at KPMG?
[Laughs] People think we just prepare tax returns for individuals. We actually do much more. A major category of our work involves preparing business tax returns for various businesses, large and small, such as corporations, asset management firms, startups, etc., as well as individual returns. Another large part of our business involves primarily giving advice to clients on business transactions. This involves marrying business understanding with tax laws that are always changing – especially since the U.S. Congress made significant tax law changes last year. Tax laws also constantly change as a result of court case and administrative guidance. There really are no natural tax laws for companies to just follow.
So, in doing the above, we undertake a variety of activities using AI to advise our clients including:
1. Analyzing transactions. This includes transactions summarized in general ledger accounts, such as analyzing expenses, revenues, etc. to make determinations. This includes:
- Should the expenditures be capitalized or expensed?
- How much sales tax or Value Added Tax (VAT) should be applied??
- How should revenue be recognized? Is an expenditure deductible for tax purposes?
- What is the cash flow impact?
2. Analyzing documents. This goes back to the business of giving professional advice. We have tax lawyers and tax accountants who review documents and determine the decisions that have impact on cash flows, etc.
In fact, prior to our discussion today, we received news that we won a proposal with a client that involves analyzing documents that professed one determining factor in selecting us over the competition: KPMG’s use of AI.
Let me shed some light on this. This is a use case where we are using AI to analyze documents to advise clients. The U.S. Congress determined that they want to encourage research and development (R&D) among enterprises by using a subsidy. This subsidy is in the form of a R&D tax credit to reduce federal income taxes. However, if the IRS audits a company, the IRS often contends that the company doesn’t qualify for these R&D credits because of the lack of adequate documentation.
KPMG helps companies obtain the R&D credit, but the biggest challenge with this IRS contention is gathering the documentation to support the R&D effort. There are potentially thousands of documents to review, but it is not cost effective for a company to review them all.
The traditional method is to review a sample sub-set of all potentially relevant documents. One way to select these samples is to parse through the file names of the documents and to interview R&D professionals, such as scientists and engineers, to distill more evidence to support the R&D credit.
CFOs, however, don’t want to hear that their scientists are being aggravated by all these interviews! So we trained AI to review these documents and identify the relevant portions of the documents that inform the qualification for the R&D credit. The AI ingests, analyzes, categorizes, and prioritizes the documents so that KPMG professionals are only reviewing the most relevant information supporting the qualification of the R&D credit. We can have the AI review and prioritize the documents at scale in a number of industries, including manufacturing, financial services, pharmaceuticals, and aerospace. The different industries require specialized knowledge, and we leverage our tax lawyers and tax accountants to train the AI models for the industry specialization. With the AI augmentation, we identify better documentation because all documents were reviewed, rather than just a sample. Using AI also lowers business disruption because it reduces the number of interviews needed with the client’s R&D professionals.
So the value proposition is not just doing things cheaper and faster…
Exactly. We’re not just “paving the cow path” and making things more efficient, but, rather, we’re turning the process on its head and creating a new value proposition for our clients, and, ultimately, changing the business model for KPMG. The value proposition is not just making things cheaper and faster, but also minimizing business disruption for our client by decreasing the number of interviews needed with the R&D professionals. Further, because our documentation is better, the R&D subsidy retained upon IRS audit for our clients is better – that’s the ultimate economic outcome.
What has been the impact so far (both for the clients and internally for KPMG)? Is it as per your expectations? Any major surprises (either pleasant or unpleasant)?
We need tax lawyers and tax accountants, data scientists, computer engineers, change management professionals, UI/UX teams, etc. all working together to make this happen. In our first kickoff meeting, we had some of our younger professional share that they told their parents what they were doing at work [training AI], and their parents were worried. “Why are you doing this, AI will eliminate your job,” they said. The AI is actually doing a lot of the mundane work – the soul crushing work. And it serves up the ability for people to do what they really enjoy – analysis and solving puzzles augmented by AI.
There is a symbiotic relationship between our professionals and the machine; we use supervised learning to train the AI and refine the machine. At the same time, our professionals are actually learning more about the tax law and the requirements for the R&D credit. So, there’s a “human in the loop” component while using AI, and we’re confident that we’ll need that “human in the loop” aspect for the foreseeable future.
So we’ve discussed the why, what and how. Let’s talk a bit about adoption challenges. The three biggest challenges according to our research to scale AI are around data quality, training the algorithms to yield the desired results, and explainability. How are you addressing these challenges?
This comes back to the business process. And how we’re constantly using “human in the loop” to supplement AI. During training, we get a head start with training and annotation by using our experienced tax lawyers and tax accountants. Our teams look at the outputs of the algorithms to ensure we can explain the legal requirements for the analysis by AI. This supervised learning is important in order to explain the analysis to the IRS. Potential changes to the algorithms are curated by our most experienced tax professionals. That’s a learning for us — in one of our solutions, we didn’t have that supervised learning process built in the first time around. Now we continue to experiment with how else we might approach supervised learning.
This is a really good story, but I am sure there must’ve been a number of mis-steps or failures along the way. Can you share 1-2 key learnings that you’ve had along the way?
Sure, we’re constantly learning and improving. We met with various folks when we were starting out. We thought innately that there was something we could do with AI in our business. We came up with little use cases here and there and did some proof of concepts. Then, as we learned more, we saw more practical applications. One lesson is learned is that a POC is not a production build. So, if I’m a consumer of AI services, I’d be careful of someone that sells a POC that will be difficult to scale to a production system. I had to unwind an architecture sign-off because I learned that they were writing rules to get a POC result and not true AI.
On another solution, we focused a lot on solving the most difficult technical problem by using AI. In this case, it was how to classify a product based on its product description. But implementing the AI solution was challenging. We learned that each company had completely different product descriptions! Sourcing the data to apply AI was also difficult, though we have overcome that now. The point is, we shouldn’t just focus on solving the hardest technical problem, but, instead on thinking through the entire end-to-end process including data acquisition.
Lastly, no matter how smart the machine is, the UI/UX is the sizzle that sells it all. It’s hard for non-technical people to distinguish what’s happening behind the scenes and how smart it is or isn’t. The UI drives the buying decision.
Where do you go from here? What’s your roadmap for the net 12-18 months? What other emerging technologies (if any) are you exploring that can make this even more powerful?
We keep opening the aperture now that we have legs and are in production, in order to better understand what’s possible with emerging technologies. We have some experiments with blockchain…a hypothesis we’re excited about in which we will use AI as we’re analyzing transactions. Blockchain can give us transactions, and it will be easy to layer an AI-based analysis on that. This applies to all kinds of taxes—one transaction may have 8-15 possibilities of different tax analyses. Once you say you can use AI for tax, that’s also not the only rule set you can apply; there’s accounting and business rules, etc. that can all be layered in.
Finally, if you could make one wish come true, what would it be?
Hmm. I’d probably wish for the doomsday scenario skeptics of AI and other advanced technologies to begin to appreciate the value in ‘human in the loop’ and embrace technology for all that it can offer in conjunction with humans’ creativity, cognitive ability, and more. It’s amazing how much more we can accomplish when everyone is on the same page.
