Are AI Agents Enhancing Or Replacing Workers?

This was a short post in Forbes that we have built on and expanded on the points raised in the post. That post looked at the potential of AI Agents and the impact on the workforce, but then it discussed what it takes to build an Agent.

AI’s false promise

The advances in AI models, and an understanding of prompting, have dramatically improved what AI can do. At Elements.cloud we’ve been using AI internally since ChatGPT hit the scenes. We’ve been using it internally in every area of the business – product management, marketing, business excellence, IT – and even finance. We’ve also been embedding inside our application to do everything from drawing process diagrams, to writing user stories, and having conversations with metadata.

What is interesting is it has changed the culture and acceptance of AI in the company. Teams are not worried about their jobs. Nor are they dismissive about AI’s capabilities. The mindset has changed to “why can’t it” or “what if we tried” for every activity to see if AI can be a valuable assistant.

This doesn’t happen overnight and it also requires a more forgiving, patient approach that has reaped longer-term rewards. We think of AI as the super excited, unflagging intern. Firstly, if AI doesn’t give you the answers you expect, then look at how you are asking the question and the resources you are giving it. Secondly, we don’t expect AI to perform every task flawlessly. We have become a lot better at working out where it can do a great job, and where a human is better. And finally, as managers, we need to give teams space to experiment and fail, so that they have time to get to the right answers. 

So we can see why there is still skepticism about the value of AI by those who haven’t engaged in pilots and roll-outs and seen the benefits. Security concerns, a lack of clean data, and hallucinations are normally cited as the reasons for not getting started. A recent Forbes article has shown that data quality shouldn’t hold you back. 

But, the benefits of well-architected, planned, and executed AI Agent implementation are huge. The recent McKinsey & Company report The economic potential of generative AI points to the revolution that we are about to see. 

The greatest issue is that expectations from AI are too high and there is a lack of open-minded experimentation. The claims that every job in every industry would be upended have proven to be false. AI is a very valuable assistant. It will not replace jobs.  

But is that about to change with agents?

Are AI agents really that different?

Many websites have bots that answer the simplest of queries. These have been developed by thinking of the potential decision paths and providing access to resources at the end of each path. Depending on how well they have been designed and how complex a task they are trying to support, they range from helpful to very annoying.  We’ve all had the “Press 1 for confusion, Press 2 to get stuck in a loop”.

AI has improved bots conversational tone. AI is able to interpret the discussion so the input can be natural language that is typed or spoken. But the bots are still constrained by their decision trees, and customer experience is based on the quality of their design. 

We use Intercom.com for the Elements support site – support.elements.cloud. We have a huge number of articles ranging from solution guides to deep-dive technical resources.

 Before AI, the chat was pretty good at suggesting articles. But it relied on the end users using the Elements terminology to get decent results. OK, but not ideal. But Intercom launched Fin which is AI powered. The end user can ask natural language questions and Fin interprets them and finds the right resources. Fin is still not an Agent in the true, academic sense of AI. Intercom has fine-tuned it so that it can refine every query, optimize every response, and validate the quality of each answer. However, it is a grey area. If AI is operating like a customer agent – dealing with a customer in a conversational way and answering their questions effectively – then it is getting the job done.

A true AI Agent is able to reason, plan a set of actions, and then apply resources to deliver an answer. The key difference is the reasoning. Now the AI agent is provided with a set of resources that include data, automated actions, and guardrails so that it knows when it needs to be handed back to a human. AI agents can also take and validate input like customer details, order references, or invoice numbers, and provide a personalized response. They can even take action such as sending an email, rescheduling a delivery, or updating an order.

The benefits are that customers can get a more personalized service for simpler tasks without talking to a human. This reduces the CES (Customer Effort Score) and increases CSAT (Customer Satisfaction).

AI Agents deliver results more accurately, at a lower cost, and more consistently than humans. They can also handle seasonal peaks reducing the need to scale the human workforce with part-time staff. For example the product returns process after Black Friday and Christmas purchasing binges.

We need to think of an AI Agent as the newest, greenest, most excited employee. We can teach them about the organization, the culture, the policies, and the resources that they have access to. But, we need to realize that they are a brand ambassador for the organization with the ability to touch thousands of customers. We need to make sure that they are “on-brand”. If not, the pursuit of the lowest-cost customer service transaction is counterproductive.

Potential use cases

Every industry can use AI Agents for both internal and customer-facing tasks:

• ​​Banking: Proactive financing to improve loan terms
• Retail: Grow sales with an always-on concierge
• Healthcare: Match patients to doctors for better care
• Technology: Update old code to reduce tech debt
• Manufacturing: Improve uptime with proactive maintenance
• Telecom: Reduce wait times with 24/7 support
• Nonprofit: Personalize fundraising to increase donations
• Insurance: Payout faster with automated handling

Salesforce has produced a library of Agents that can be copied or used as the starting point for your own Agents. These can be filtered by industry or by product: Sales, Service, Commerce. Access list here.

The end of humans?

Yes, in some cases. In a large-scale assessment of 22,000 tasks in the UK economy by IPPR, the Institute for Public Policy Research, found that about 11% of tasks are exposed to generative AI right now, and this could increase 5x if AI systems became more deeply integrated into organizational processes.  

Nearly three million people were employed as customer service representatives in 2022, with the majority (66%) being women, according to Data USA.  So these are the people at risk when we deploy AI Agents. But it is not all bad news. 

No customer wants to sit on hold for 30-40 minutes to deal with a simple transaction that could have been dealt with by an AI Agent instantly. And the customer isn’t probably in the right frame of mind when they finally get through to a human. AI Agents can deliver lower-level tasks, freeing up humans to do higher-level, more rewarding work. 

Sitting in a cubicle in a call center rescheduling appointments, or answering basic queries, is a job that helps pay the bills but is soul-destroying. The staff turnover in call centers is high because people move if they can find a job with an hourly rate that is slightly better. This is not good for the individual, not good for the organization, and therefore is not good for the customer experience. 

To start with, the most basic transactional tasks will be handled by AI Agents. A critical part of their design will be understanding when and how they hand off to a human seamlessly. We’re going to see some very good and very poor implementations, just like we did when the first touch-tone systems came out. The quality of the implementation will determine the customer experience.

Over time, the best implementations will set an expectation with customers about what a good customer experience should feel like. What is different with AI Agents, is it is no longer determined by the quality or salary level of the person on the end of the phone. It is about the effectiveness of the implementation: the process and the data.

AI Agents will be a huge competitive differentiator. A poor experience could drive customers to a competitor who has mastered the design and implementation of AI Agents. The average cost of a human answering the phone is over $13 according to Brett Taylor, CEO of Sierra.ai – and AI Agent startup. If you recognize his name it is because he was CTO at Facebook, invented GoogleMaps, and went on to be co-CEO at Salesforce after they acquired his company Quip. He left to found Sierra.ai.

When products or services are price competitive, or almost commodities, the difference is the level of service. An AI Agent can deliver a better level of service for 10% of that cost. At that price exceptional customer service delivered by an AI Agent is within reach of any organization It is not just for luxury products, but for every organization, whatever the price point of the product or service that the customer is buying. 

Putting AI Agents to work

There are a number of technology companies offering a platform to develop and deploy AI Agents. We’ve already talked about Sierra.ai. But the biggest player is Salesforce. They are the #1 application for customer data. They also have a strong service offering, in competition with ServiceNow. 

Salesforce made AI Agents a central theme of the recent Dreamforce conference branding it Agentforce.  In a recent interview the CEO, Marc Benioff talked about the huge pivot they have made to AI Agents. In this recent interview, he said they weren’t bold enough when AI first came out and missed an opportunity. This is why the whole company is now laser-focused on AI Agents –  Agentforce.

He even teased the idea of rebranding Salesforce as Agentforce.  MK [Muralidhar Krishnaprasad], Salesforce President and CTO said to him, ‘Marc, you are making a huge mistake. I want you to rename the whole company Agentforce. Right now, you do not understand the fundamental implication of what has happened.’

But is the claim that “anyone can build an agent in minutes” realistic? At Dreamforce over 10,000 Agents were built. This shows that the technology is ready for prime-time and can be implemented by Trailblazer. 

As with every technology, the two key questions are “Could you?” and “Should you?”.  These question the readiness and maturity of your organization’s processes and data to be able to deploy AI Agents. So to answer the questions we need to look at the 5 attributes of an AI Agent. The customer outcomes are only as good as the data the AI Agent can access, the quality of the instructions, and the power of the actions it can take.

Role: What is the scope, topics & instructions?

Picking the tasks that the AI Agent can deliver flawlessly is the starting point. That requires a clear view, at a detailed level, of the operational processes. These processes can then be enhanced taking into account the power of the AI Agent to craft an amazing customer experience. An AI Agent’s ability to reason is its power. So you need to be very clear on the scope of what it can, and cannot do. “The devil is in the detail”. Fortunately, the UPN process mapping notation, with its hierarchy, resources, and attachments is perfect for this, which is why it is the Salesforce standard.

Data: What knowledge can they access?

Typically customer data is siloed in applications. A human is able to pull this 360-degree view of a customer’s interactions together to deliver a seamless service. Access to and quality of accessible data may determine the scope of what the Agent can do. You may need to build an aggregated, harmonized view of the customer that the AI Agent can access for it to work. This is why Agentforce is tightly connected to Data Cloud.

Actions:  What capabilities do they have?

The actions are automated workflows, prompts, or access to external services that the AI Agent will decide to use. These can be existing or new metadata. Remember, the power of an AI Agent is that it can reason and build a plan of action based on the request, the data, and the actions it can take, bounded by the scope of what it can do.

With high levels of tech debt in your applications, it may take some time to identify the existing metadata that you want to use in an action. You need to be certain that the AI Agent cannot inadvertently exceed its remit because you didn’t understand the nuances of the metadata in an action. Also, the AI Agent is reading the documentation about an action to decide whether to use it. When it reads the documentation will it be confused or just disappointed!!! So for actions that are using existing metadata that you have made available to the Agent, there may be some remedial work needed to improve the documentation, detailed testing, and lock down some of the edge cases.

Guardrails: What shouldn’t they do?

The guardrails are critically important. You are allowing the AI Agent to make decisions and build a plan. Thinking through what it cannot do requires more detail than with a new employee, who will probably err on the side of safety. The other aspect of the guardrails, is you need to design the process so that the handoff to a human is elegant and seamless.

We’ve all experienced when a human agent hands us to another human agent and you have to go through security questions and explain the background and the problem all over again. This is an opportunity to build an exceptional customer service experience. So how is that going to work – in detail? How does the human see all the discussions, decisions, and data that the AI Agent has gone through with the customer before it is passed over? Maybe AI can summarize that for them?

Channel: Where do they work?

Again, this requires a detailed understanding of the business processes so that the hand-offs can be determined. However, these processes also need to be updated to reflect that an AI agent is performing some of the work.

Too soon? Not ready?

It’s easy to look at these 5 attributes and decide that you are not ready to deploy AI Agents. This may be true about deploying AI Agents at scale on customer-facing websites and portals. But that shouldn’t stop organizations from launching pilots to establish their readiness, test the technology, and establish profitable use cases. 

Salesforce has made it easy to get started with Agentforce, with an ambitious series of city events, one-day workshops, and even waiving the 1st time fee for AI Certifications. Marc said that he has done a hard pivot. Organizations need to think about what hard pivot they need to make to exploit the power of AI Agents. 

Final word

Not every organization is ready to deploy AI Agents at scale. But, the benefits of well-architected, planned, and executed AI Agent implementation are huge. We are about to witness a revolution. Those organizations that build a foundation of documented processes, data governance, and well-documented orgs will be the winners. Every organization can, and should, get started right now so that they stay relevant