Driving ROI from generative AI in corporate finance

The impressive and easy to understand use of generative AI for most people has been as a replacement for web search. Instead of searching the web for page(s) that answer your question, now we expect AI to find sources and summarize for us, saving us the work of reading non-relevant information. For corporate use of AI, this can be helpful for employee productivity at an individual level.

What about the next set of use cases?  How can companies use generative AI to drive efficiency and ROI for more complicated scenarios?  In addition to developing and deploying AI hardware, Qualcomm is among the growing ranks of thought leaders applying AI to transform internal efficiencies. This blog will detail the steps taken to explore, refine, and use generative AI to improve customer service in a corporate finance scenario.  The patterns outlined here apply to any process workflow that takes a long time or more people than you have to apply to it.

Business challenge

For semiconductor vendors like Qualcomm Technologies, Inc., it is common practice to provide channel incentives for meeting sales targets for various products. One of the challenges of running an incentive program is structuring it so that claims from channel partners can be verified as legitimate according to vendor rules.

The goals of creating an incentive program is to reward actual downstream sales, not inventory loading.  It can apply to specific lists of customers, devices, or end markets and may only apply to certain products for a given time period.  Further complexity arises in verifying across many partners, customers, and markets – including invoices and documentation across a variety of languages.

Generative AI value

For incentive claims at Qualcomm, we can say that the time it takes a human to collect information, validate amounts, compare and reconcile variances, and finally approve an incentive payout takes on average 4 hours.  Multiply that by the number of claims per month to see that it takes a team of finance professionals weeks to accomplish this work.  They also have other responsibilities, further slowing turnaround time.

Using generative AI combined with workflow automation reduces processing time to a few minutes. A finance employee still reviews all of the work summarization and approves final payment, but effectively using AI allows the finance employees to turn around incentive claims as quickly as they come in.  The result is that partners are much happier to receive their payouts faster, and the finance team can focus on higher value work.

Implementation journey

Solving a complex process like incentive claims which depends on multiple data sources both structured and unstructured is not possible with one-shot AI prompting. Ten years ago, one might have proposed using RPA (so-called Robotic Process Automation) to solve this challenge.  RPA works well when the inputs and outputs are static and consistent between the systems that are connected in a chain.  In this case, the wide variability of documentation submitted by channel partners makes generative AI for data extraction a better fit.

At a high level, these are the steps the team needed to take to arrive at a reliable working solution that achieves high levels of quality and follows all the business rules:

1. Teach finance personnel how to use low-code automation tools like n8n.
2. Have the expert design a complete end to end workflow for a process.
3. Steps are often very granular, such as: extract the total from this PDF invoice, validate it is a number, parse and pass on to the next step.
4. Steps may include using AI to write SQL queries using a partner name and numerical amounts extracted from a PDF – to validate from internal databases.
5. Many steps include validation against multiple systems. Those validations are then justified by short English summaries generated by AI to explain variances or agreement.
6. Compile a final recommended decision for payout subject to human approval.
7. Allow humans to approve and take the action to pay out the incentive.
8. Alternatively allow humans to modify the data and action prior to approval or rejection.

Using low code tools allows quick experimentation to get the process right for every granular step.  Once the process works at an acceptable quality and pass rate, it can be time to convert the process to actual code.  That code does all the same things but can be run on demand when a new incentive claim lands.  It also provides a blueprint for other processes in the future.

Final thoughts

Automating complex business workflow takes more effort than simply asking AI questions about corporate data.  Agentic processes can vary depending on inputs and may not provide the kind of consistent performance needed. 

Combining the step-by-step discipline of workflow automation while using generative AI to handle extracting data from arbitrary document formats, summarizing steps taken, and providing commentary on why – is the key to solving complex workflow automation. In addition, using AI as a judge for prior outputs in the process creates additional validity that a human can use to ensure their work is done correctly.

We’d love to hear how you are using AI in workflow automation. Hit us up over on the Qualcomm Developers Discord channel to ask questions or connect. We also host generative AI models with our partner Cirrascale.  Be sure to sign up for free tokens and retrieve your API key to try out generative AI in your workflows.  Explore other topics in the cloud AI blog series.