Let’s start with the basics: what is an “AI agent”?
An “agent” is generative AI applied for a specific practical purpose.
It uses a Large Language Model (“LLM”) of the same kind that underpins your favourite chatbot. But unlike a chatbot, which simply returns text or images based on your prompts, an agent applies a set of instructions to work through a pre-defined task from start to finish, dynamically establishing context, next steps, and supplying outputs autonomously.
Most importantly, they do something an LLM chatbot as a standalone system can’t do: they interpret all of this input and use it to direct other systems, meaning they can instruct other tools to perform actions without any human input.
With the right governance structures, access permissions, and workflow architecture, agents can represent a tremendous source of value for the enterprise. Properly built, AI agents can independently complete multi-step tasks that might take a human team member several hours in a few minutes.
In the domain of accounts receivable, there are many tasks that are ripe for handing over to agents under the right circumstances. In this piece, we’ll explore some of the most promising use cases that we’ve been experimenting with for our customers. We’ll also explain why and how we’ve chosen these use cases and where we won’t be relying on agents—and why some tasks are best automated by more traditional routes.
What are the limitations of agentic AI?
In order to be effective, agents have to interact with and manipulate external tools, APIs, and other software systems as part of their process. As many have discovered, this isn’t without its risks, as agents allowed indiscriminate access to systems can often act in ways they aren’t supposed to. This is a natural part of the probabilistic way that the underlying LLM technology interprets its inputs.
There’s no way to prevent AI hallucinations, because LLMs work on the balance of probabilities: their “reasoning” is based on predicting the likelihood of a specific outcome based on their training data.
For this reason, it’s not wise to simply deploy agents across all your workflows and simply have them replace entire teams and skillsets. The larger the scope of responsibility you give to an agent, the greater the risk to which it can expose your organization.
To minimize this risk and create agentic systems that add real value to the AR process, we instead have to think smaller rather than bigger.
A well-behaved and successful agent is one designed to do one specific task within a well-defined set of rules and governance frameworks that limit its potential to take the wrong action, but that also make its reasoning transparent and legible to the teams responsible for managing the task it executes.
Where can AI agents have the greatest impact in AR and finance?
Agentic AI is best thought of as a high-performance race car. It works great under specific circumstances, but for many others it’s far too expensive and inefficient. You wouldn’t drive an F1 Ferrari to the supermarket, for instance, no matter how much you might like to.
In the same way, while you theoretically could replace your entire tech stack with AI agents, you probably wouldn’t want to, because it would be both more expensive and less accurate than relying on simpler “conventional” automation approaches.
According to some industry experts, deploying agentic AI in a given enterprise workflow can cost around 10x as much in the development and implementation phase, and a mind-blowing 30x as much to run and maintain over the course of its life.
For this reason, before you start building agentic workflows, you need to ask two questions about the task you’re trying to automate:
What’s the current state of play for this process or task? What’s the best practice the industry currently abides by, and can agentic AI meaningfully improve on this?
Can we automate this task in any other way? If said task is a matter of simple data importation or transformation, or translating information from one system or format to another, you probably don’t need an agent to do it.
Does this task require reasoning for success? AI is strongest in situations where you need to mimic human thinking and reasoning.
This means agents are a genuine asset in any arena that involves:
- Large, complex, unstructured data sources.
- Comparing documents in multiple formats according to unclear rules.
- Finding patterns in un- or underexplored data.
How can you bring agentic AI to bear in your AR function?
Our Chief Product Officer has developed a great framework for deploying agentic AI in finance to develop autonomous workflows. It’s very simple:
Start small with limited agent deployments on limited, potentially high-value tasks. Scrutinize outputs carefully and maintain human oversight, and don’t allow agents to work “independently” yet. Instead, have them make suggestions in a way that the human in the loop can validate easily.
Bring agents working on related tasks that’re performing well together to create a connected “crew” that can handle multiple tasks in succession as an end-to-end workflow. This could look like taking an incoming payment in the AR workflow, matching it to an invoice, purchase order, and remittance document, and posting it instantly with a final check by a human at the end of the process. Each task is handled by a separate agent, and the final result validated by the human in the loop.
When accuracy is high enough, allow the crew to operate autonomously with more cursory reviews triggered when outputs look obviously erroneous or exception handling flags are triggered.
This allows your finance function to work incrementally towards fully autonomous workflows without putting the enterprise at undue risk.
It also presents a clear and intuitive pathway for selecting trial tasks for agentic automation. Together with our customers, we’ve used this framework to develop a suite of “starter” agents that can be incorporated into real AR workflows with relatively low risk that’re currently being rolled out into real production environments.
Let’s have a look at the tasks they can manage, and what makes these tasks good fits for agentic automation with high potential for value creation.
Unmatched payments resolution
Our unmatched payments agent does exactly what you’d expect it to: it cross-references unmatched payments with unstructured data sources like emails, remittance advices, and customer information, identifies the most relevant matches, and generates a list of recommendations for those payments to be matched.
Your teams can accept or reject its recommendations based on audit logs that explain the decision, confidence intervals provided by the platform, and their own judgement.
This eliminates most of the manual number crunching and data cross-referencing we still see in the cash application process: and takes us closer to true 100% automation of all cash application tasks.
Collections email handler
This agent scans inbound customer emails, classifies them in terms of their content, context, and customer relationship status and then flags them to your collections operators with proposals for appropriate responses and follow up actions.
As with all our agents in AR, you can choose exactly how much or how little autonomy you give it in terms of completing tasks end to end.
At the most autonomous end, when you’re satisfied with its performance, it can handle simple actions like composing and sending acknowledgement emails to ensure prompt responses, and even suggest payment plans and terms based on pre-set policies from your governance stack.
Automation improvement identification
We’ve developed the automation improvement agent to help with one of the biggest outstanding problems of enterprise software: many users and organizations simply don’t have the time or resources to optimize how they use their automation solutions.
This agent handles this by taking a look at what settings your teams are using in their Serrala AR tools, compares them to organizational priorities and needs, and identifies where you could improve effectiveness without any need for bringing in additional tools simply by changing how you use what you have.
Of course, the agent doesn’t make any changes itself. It simply tells you what you can do, how to do it, and who to talk to within your Serrala team to ensure it all rolls out smoothly.
Dunning strategy recommendation
Analyzes payment behaviors and outstanding invoice deadlines and presents collections teams with a step by step recommendation for how each individual customer and invoice should be followed up.
A great solution to one of the more manual ongoing challenges for collections departments: figuring out the best way to approach customers in different industries, markets, and credit statuses to ensure they’re most likely to pay promptly and to keep DSO as low as possible.
Customer insight
Analyses digital customer interactions from email conversations to phone logs and payments behaviors in order to build an effective profile of what matters most to each customer, the problems they’re dealing with (both with your AR team and beyond), and how your organization can act to best improve their experience, their relationship with you, and their ongoing value to your business.
This can include everything from simple sentiment analysis of recent communications to stop disputes before they start to predicting likelihood of churn within a given period. A great ally in helping your teams prioritize trickier cases that often go unaddressed until it’s too late.
Dispute outcome prediction
Uses your own customer portal data and machine learning principle to classify ongoing disputes by type and predict their likely outcome based on individual customer history and general trends. Also provides teams with advice on the best course of action to create a mutually beneficial outcome for both parties.
Much like the customer insight agent, this is a brilliant tool when it comes to identifying and providing a plan of action for intractable problems that all collections teams have to deal with, but which are very difficult to prioritize or handle proactively thanks to the complexity of day to day workflows.
Credit risk assessment
Cross-references customer payment behavior data within your own ERP with 3rd party market and publicly available financial information from trusted sources in order to proactively monitor customer credit risk—including when businesses are more likely to start presenting a bigger risk of non-payment.
This allows collections teams to avoid nasty surprises (like formerly reliable customers suddenly becoming a source of potential bad debt) and work to both protect the business and support the customer as best as they’re able, including by flagging existing pre-approved policies and payment plan options and new terms and conditions that your teams can offer to such customer.
Want to learn more about how Serrala can help you adopt agentic workflows in your operations to create a more autonomous finance function?
At Serrala, we’ve been helping CFOs and their teams automate finance operations for over 40 years. It’s our job to help you integrate new technologies within your organization in a way that makes sense both for your challenges and their capabilities.
We’ve developed our newly released Serrala Finance Platform with AI-readiness and autonomous finance principles in mind.
A single working capital intelligence hub that provides the foundation to bring AR, AP, payments, and treasury workflows together in a flexible system that adapts to individual organizations’ digital transformation roadmaps while enabling AI integration in every single process.
This allows us to provide the integrated architecture to deploy agentic AI use cases in a way that ensures total and seamless operational efficiency while still providing complete choice and flexibility as to which parts of the offering your organization uses (and pays for) at every stage of your transformation journey.
This means our solutions empower your organization to start with the workflows you need today, then expand as your needs grow and your transformation plans mature.
To learn more about how Serrala can help your teams bring agentic AI workflows to bear in your finance function, get in touch today to book a demo.
