What are the benefits of implementing AI to replace manual processes in your organization?

Large language models (LLMs) can help the bottom line and enable your organization to pursue more strategic opportunities – the breadth of applications of LLMs can flow through to the bottom line by unlocking revenue growth, creating cost savings, and making it easier to capitalize on new opportunities by allowing staff to focus on more lucrative activities. For example, AI can unlock revenue growth by making it easier to hire key personnel, create cost savings by making processes more efficient, or allow technical staff to focus more on core product development instead of time-intensive technical implementations. 

LLMs can improve end-user experience by expediting and streamlining repetitive processes – AI doesn’t just create internal benefits for a company. If well-implemented, customers also benefit from a more seamless experience that uses automation or self-service options to replace more traditional solutions that are time-consuming and/or more likely to be error-ridden. 

See  more on the benefits of AI for PE-backed companies at the Fractional blog post here.

Laying the Groundwork for AI Implementation

Who should own the implementation of AI for automation within your organization? 

The CEO should own the strategy behind the organization’s investment in AI – the CEO is responsible for taking a proactive approach to using AI to create value for the team and encouraging others to help identify opportunities to introduce automation.

The CTO, VP of Engineering, or VP of Product should own the technical aspects of implementation – this includes adoption of the use case and making tools available across the organization. 

How should companies evaluate and select AI use cases within their organizations?

Identify manual, repetitive processes that are executed at scale – focus on areas where improving pieces of a process or a task can help make an employee’s life easier or streamline a process. Any process that has a large playbook of steps or is offshored (or a candidate for being offshored) is a good place to start. Any use case should have a practical justification that positively impacts the business. 

Test run use cases by using a public LLM to identify promising use cases – don’t overcomplicate the process of identifying AI use cases. Try leaning on Chat GPT or another public AI tool to do tasks that eat up your time. If you find that the tool can be helpful—but that the task is too complicated, so the tool isn’t reliable enough—that’s a potential indicator that this task is a good project to tackle with AI.

How should organizations approach evaluating which LLM best suits their proposed use case? 

Seriously evaluate third party solutions before self-hosting and training your own models – unless you’re one of the cutting edge companies in the AI space, you will never be able to create a better, more specific holistic solution internally. You will spend significant time and money creating a system that becomes irrelevant when the next iteration of AI tools rolls out. It’s better to use existing solutions as building blocks and overlay customizations as needed.

Don’t be overly opinionated about infrastructure before starting a project – don’t commit to a very specific solution before fully understanding the nature of the problem you’re trying to solve and the options for how to solve it. Leaders often want to control an AI project from start to finish, but the most effective applications focus on the desired outcomes, and work backwards. 

Along with a model, consider the technical building blocks if you are going to build an application – infrastructure frameworks and databases (e.g. Langchain and ChromaDB) are crucial to the project of building your own application, but are in a state of flux as tooling to support building AI use cases improves. 

What are the different levels of human involvement in an AI implementation? How do you decide how much human involvement your implementation of AI should have?

There are three different types of use cases:

• Automation first – an automation first approach creates a system that automates a process from end-to-end so that employee involvement becomes unnecessary. If a customer types in a question, an automation first solution would handle the conversation/output without involving a customer service rep. 
• Co-Pilot – A co-pilot approach inserts bits of AI into a process with the goal of having AI accelerate the process or help an employee go through the process more quickly. If a customer service rep is chatting with a customer, a co-pilot solution would offer an autocomplete option every time they go to type in a reply.
• Human first – these tasks are sensitive because they’re often customer facing and new, and they can’t be handled programmatically, so humans have to handle them manually. 

Every task you consider automating falls in one of these four quadrants: 

Co-pilot and automation as existing on a spectrum, not two distinct all-or-nothing types of solutions – it is often more helpful to ask how much you need AI to do to help make a process more efficient. If AI is doing more of the process, you’re closer to the automation end of the spectrum. If AI is providing support and enabling people to do more, you’re closer to the co-pilot end. 

Beginner Use Cases

What are common types of AI use-cases for companies just embarking on their AI journey? 

Tasks that are repetitive, widescale, and medium-to-high impact are ideal opportunities to implement AI – these tasks offer a chance to implement AI in a discrete portion of a process instead of trying to replace it entirely. 

Use Case		Example AI implementation
Reading documentation and extracting relevant information		Combining large quantities of RFP documentation and using judgment to pull out pertinent information to inform decisions. This is an example of “Needle in a haystack” natural language tasks like identifying relevant terms in a contract.
Generating code based on extensive documentation		Writing code to build integrations with third party services based on API documentation
Offloading portions of manual playbooks that are currently offshored		Uploading information such as order numbers from large quantities of documents into a database
Customer support workflows		Scheduling calls with customers or directing them to solutions to common issues.
Lead qualification		BDRs have to research or call a high volume of leads to qualify them. AI can conduct the lead qualification process on a larger scale and at a faster rate than the current employees, removing a bottleneck on sales team capacity.

Customer Support Use Cases

How can AI improve efficiency in customer support call centers?

All companies with call centers should be thinking about AI phone agents – Historically, the technology that allowed you to have an AI on the phone was expensive and not quite good enough. Today, we’re at the point on the adoption curve where both of those factors are shifting. The current variable cost of having an AI on the phone is zero—and it will soon become cheaper and possibly a more reliable experience than speaking to a real person for certain things. AI phone agents can help call centers operate more efficiently by taking on repetitive, high-volume tasks that free up live agents to discuss more urgent, complicated, or important topics with customers. 

There is a wide variety of tool options available – there are also many vendors (e.g. Cisco, Genesys, etc.) of AI-powered call center tools, alongside many vendors who resell the core models and give you the technical primitives that let you answer phone calls, send and receive texts, design flow charts, etc. 

What are the baseline steps to implementing AI in a call center?

Step 1: Identify the call center interaction type where AI can begin to assist first – if the most common reason customers call is to schedule an appointment, and this is a repetitive, low-risk task, it’s a prime target for applying AI. Try to identify a use case with a relatively fixed set of potential outcomes. 

Step 2: Create a solution for that use case and fine tune it – using a call center AI tool, create an option tree that offers new customers who call (and who used to be directed straight to voicemail) the option to either schedule an appointment with an AI phone agent or leave a message. 

Step 3: Continue to extend your solution to comparable use cases –  eventually, you can expand your solution to help alleviate live-agent time constraints by introducing your AI phone agent to inbound calls seeking to make appointments. 

Tips For Implementing AI Phone Agents

• The biggest challenge is building a reliable phone experience for customers – set guardrails to avoid situations where AI “hallucinates” bad policies or gets you into trouble. Text-based support (email, chat, etc.) is often easier for AI to handle correctly. • Use decision trees for topics and ideas – do not use them for keywords. This provides guardrails for your AI while allowing it to use natural language. • Start with a narrow use case that has a clear path toward expansion – select a certain type of phone interaction as your baseline. Don’t start expanding until your initial use case is well-tuned.  • Work with a rockstar engineer – AI phone agent technology is performing better than ever because so many models work together to create each solution. Those solutions and models are changing fast, and you need someone who understands them and can adjust your LLMs as needed.  • Embrace the fact that your solution will be imperfect – don’t aim for perfect; aim for useful. Embed the necessary logic for your solution to detect when it needs to escalate to a human.

Business Processes Use Cases

What are some categories of business processes or services that might previously have been offshored that AI can help automate? 

Quality Assurance – AI can drive efficiencies with certain tasks or processes that are too complicated for AI to do itself. For example, AI can’t do QA by itself, but it can review the job an offshore team does and flag QA opportunities to an onshore team, reducing the risk posed by offshoring. 

Code generation for new integrations – traditionally, companies have offshored teams of engineers who read through API documentation to write new integrations. AI can parse through that documentation and create connectors in a format that’s appropriate for the company, saving that team significant time and resources. 

Opportunities within long, multistep manual playbooks –processes like debt collection have each case go through many of the exact same steps (often in a multi-month process) that are nearly identical for each case. For example, debt collection involves drafting emails, drafting letters, organizing this information, and making phone calls. While the phone call itself should probably be done by a human, the rest of the process can be automated, allowing a large team of debt collectors to reach more people than they could before. 

Data-mapping and entry from written documents –  this can help companies gain greater visibility into data that exists but would take lots of manual entry over a wide range of operations. Data often comes into your organization in formats that are unstandardized or unfriendly to your operations. For example, an ad tech company might receive data about an ad in 25 different formats. An LLM can synthesize that information and make it easy to view and analyze that data in one seamless interface that enables the company to make faster, more informed decisions. 

What are some common examples of healthcare and insurance use cases for AI? 

Use cases for healthcare and insurance
What it is		How AI can help
Prior Authorization – each doctor’s office and pharmacy has a process for getting an insurance company to commit to continuing to cover a prescription. These processes are highly manual and frequently offshored.		AI can automate the prior authorization process by analyzing patient data, insurance policies, and prescription details to generate pre-approval recommendations, significantly reducing manual work and processing time.
Denial Management – when insurance denies a claim or withholds payment to a hospital, doctors must go through a document-intensive manual process to try to get the denial overridden.		AI-powered systems can review claim denials, identify patterns, and automatically generate appeal letters with supporting documentation, streamlining the denial management process and increasing the likelihood of successful appeals.
Dental Eligibility and Benefits Management – the outbound calling of insurance companies is a highly manual and repetitive task often performed by offshore teams.		AI chatbots and voice assistants can handle outbound calls to insurance companies, verifying eligibility and benefits information quickly and accurately, eliminating the need for manual offshore teams.

Overall

How should organizations prepare their organization for an AI-enabled future?

Treat AI as an investment in a functionality whose future costs are trending down – the AI landscape is in a Moore’s Law situation with regards to capabilities and cost. Inference costs are going down and will continue to go down until they’re cheaper than using a human. If a company is debating whether to embark on an automation project, they should keep in mind that the future costs of using that technology will continue to decrease while the impact will increase. 

Get something in production to introduce your company to the process and value-add of AI – even if your first project is tiny and doesn’t significantly change the nature of the business, it’s an invaluable opportunity to start teaching your organization how to think about AI and how to see the value it can offer different functions. Having a win that the whole company can understand will create buy-in for more difficult projects in the future. 

Understand that adoption can drive enormous impact, even if technology stops progressing – a significant unlock comes from training your organization to be open to the new ways that AI can help optimize different processes. Creating a mindset that’s open to playing with AI and testing new solutions has the potential to significantly impact your organization, even if you don’t create extensive custom tools. 

What are the most important things to get right?

Approach AI as a tool to make specific tasks or processes more efficient, not to replace them entirely – AI is unreliable in some ways, and expecting it to perfectly solve every issue and entirely remove the need for employee input is unrealistic. Instead, the value of AI can be unlocked by looking for opportunities to leverage it as a tool to significantly improve existing processes and help employees, not replace them. 

Take an open-minded, flexible approach to evaluating new AI tools – the AI world is changing incredibly fast, to the extent that the best AI tools that we will be talking about in 5 years don’t exist right now. Instead of trying to predict the best ways to use AI in the future, focus on making the most of the tools available today and training your organization to become more adaptive to new technology going forward. 

What are common pitfalls?

Viewing AI as a colossal undertaking that can do everything for everyone – companies that are set on training a gigantic core model that knows everything about their business and can do every important activity for that business are set up to fail. AI isn’t a solve-all solution, and the greatest impact comes from identifying specific ways that AI can help with specific things, instead of expecting it to be able to solve every type of problem.

Trying to do everything at once – it’s easy for companies to get excited about AI and set out to test and implement numerous solutions at the same time. However, this lack of focus indicates a leadership team that’s embracing AI for the sake of AI – not for the sake of trying to achieve specific quantifiable goals that actually help the business. Make a clear, prioritized list of the most important things you want to use AI to do instead of getting distracted by the large list of potential problems that need solving. 

Hiring large teams of AI engineers who don’t have the appropriate training and skills – AI is an emerging field, and businesses that go out and hire large teams of “AI engineers” aren’t necessarily getting the people they actually need to implement the kinds of solutions that will support their business. It can often be better to partner with AI experts who can support you and eventually set your team up for success instead of “repurposing” engineers with different backgrounds to work on areas they don’t have enough familiarity with.