The Human Side of an Agentforce Implementation 

Implementing artificial intelligence solutions can be a painful task for the uninitiated, and not just from a technical aspect either. While legacy issues like messy data can cause headaches for those trying to introduce AI to their companies, it’s also worth considering how human-related problems – like misconceptions around these new tools – can also be a problem. 

We spoke to Salesforce Application Manager & AI lead at Arjo UK, Will Turner, about what to keep an eye out for when implementing AI solutions – including Salesforce’s own agentic AI suite, Agentforce.

‘OK, You’ve Implemented AI… Tell Us How to Use It’

Will told Salesforce Ben that his company implemented Agentforce – or Copilot as it was then – one year ago, and is also currently involved in a project to implement Microsoft’s Copilot sales product, examining how people can use AI in Microsoft applications to draw data out of Salesforce. 

Speaking about implementing AI generally, he said: “I think there is an anticipation from the user’s perspective that when you implement a feature for Salesforce, maybe you’re implementing a new business process or you’re building something, that’s always usually followed by some training and some guidance: ‘Right, this is how you use it we’ve implemented this process, this is what you do to take advantage of it.’ 

“So the users still have that perspective of, ‘OK, you’ve implemented AI, tell us how to use it. ’ AI doesn’t really follow that rule.”

This is because use cases are so varied and so generic that, while certain requirements can be anticipated, generally, it’s best to understand from the users themselves how they want to use it. 

“The only reference point that users will tend to have is if they’ve used something like ChatGPT or something like that. Usually, when they experience AI for the first time, particularly if it’s outside of work, they have zero expectations,” said Will. 

Whatever results they get from using AI in that way, they’re usually pretty impressed from the start. However, when it comes to using it when they have much more specific requirements around work they usually have that expectation, but AI doesn’t tend to always match those expectations from the off.”

So, managing a user’s expectations about what exactly they can get out of AI is a critical aspect of any successful implementation.

Get Good At Asking The Right Questions

Will added that users don’t always know precisely what the right question to ask an AI is, leading to less than optimal results. 

He said: “When they use chat GPT, they just ask anything. They’re kind of unrestricted. But when they’re at work, even though they probably don’t realize it, they actually have more specific ideas or expectations in mind. 

“If they don’t craft their question [specifically], they don’t always get the answer they want the first time round.”

The answer the user receives can sometimes show them an opportunity to rewrite their instructions, which can lead to a more specific response.

“Those initial expectations from a user’s perspective can be a bit hit and miss. My advice to the Salesforce admins is that clear prompt writing is a skill in itself, and I mean that in both the sense of the admin technically writing prompts or all the users themselves writing the prompts.”

AI Use is Personal

Will stressed that one important point is that the use of AI is “very personal”, meaning how the tools are best utilized depends on a wide number of variables like internal processes, customers, KPIs, and objectives, etc. 

He said: “Let’s take sales as an example. Yes, a company wants to sell more, right? Everyone’s trying to sell. But the way each salesperson goes about selling the products for a company can be vastly different. 

“Some prefer more text communication. Some people want to spend more time face-to-face with the customer. Some salespeople like to go to the nth degree, following up and checking in with the customer throughout the buying process. Some are a bit more hands-off. Some prefer a data-driven approach. 

“You’ve got all these different factors. So whilst they’re all trying to achieve the same ultimate objective… there’s always going to be different methods.” 

Will said that AI is not necessarily going to change the way people work; it’s more that people are going to fit these tools into the way they currently work. 

“Often, early in our Copilot or Agentforce implementation, we would find that taking requirements from one user doesn’t make sense for [another].

“For example, we started with a little pilot group of the salespeople and each of them had their own individual requirements for AI that we then had to configure Copilot prompts for to become more productive and more efficient.”

All salespeople “really love” the fact that AI can generate content like emails for them, Will says. 

However, every person has their own style, their own language, and their own preferences when it comes to written communication, and in the early days of AI, these tools would produce “fairly standard” written content, he added. 

“At the start, people are kind of quite happy. It’s like, ‘I can draft an email, I can send that out to one person, and I’ll go to the next lead for example, I bring up my lead list view and I can generate an email for them all and I can… type up one instruction, ‘Please can you draft an email, suggest a meeting, let’s just talk about something’.

“They’ll copy that instruction, and then they’ll cycle through the lead records with each one sending off the emails, and it’s kind of like, okay, bang. I’ve contacted 10, 15, 20 leads, all in a much faster space of time, rather than writing out one email template and then just changing the first name.”

Initially, this is found to be quite useful, but then with repeated use, the AI is seemingly taking the same approach every time, so they might then want to personalize the emails to make them fit the salesperson’s unique writing voice, Will said. 

The Sky is The Limit… Tell Me What to Write 

AI can sometimes be viewed as a tool with limitless potential, and while this might sound terrific on paper, when it comes to actually implementing it for practical purposes, some users might be left wondering where to even begin. 

Will said: “With the amount of explaining you do to your audience to say, ‘Look, you guys are going to drive the requirements, you guys are going to tell us how you want to use it… but still certainly initially a lot of them were like, ‘Well look, I’ve enabled you, I’ve enabled AI for you, you can now do this, what are your first thoughts, are you confident enough to start?’

“One or two were kind of like, ‘Let’s take it from here and I’ll run with it and what hurdles I come across.’ And a lot of them just sat there going, ‘Tell me what to write.’”

It’s not going to be the same education for every single person, Will said. 

“Some people are going to need some real handholding. Some people are going to want to literally run before they can walk.”

Understand Your Company’s Attitude to AI… And The Terminator Movies

From fairly realistic and immediate concerns about AI like its ability to perform the jobs of entry-level people, ranging all the way to apocalyptic and sci-fi dystopias where Skynet takes over the world, a lot of people have developed notions about the place of artificial intelligence in human society. 

Will told Salesforce Ben: “Probably the most important topic to tackle before you go anywhere for any user is… understand your company’s attitude towards AI.

“There was lots of media influence, especially this time last year… you couldn’t surf any news website or anything technology-related [without encountering] stuff about AI – positive, negative… That influences people at all levels in your company.”

Obtaining buy-in from the organization’s senior level does take some time because they have their concerns, but you also need to get the buy-in at the user level as well, he added. 

This includes understanding what their thoughts are around security and how they’ve been influenced by the media.

Will said that he remembers some initial training sessions where he mentioned Skynet – the dystopian AI that brings about the end of humanity in the Terminator films.

He added that even though people might dismiss this kind of fantastical sci-fi scenario, he still found value in explaining the boundaries that would keep the AI in check. 

“When I first talked about it, ‘Is anyone concerned a Skynet situation could come out of this sort of thing’, no one wanted to admit it, but going through and explaining the boundaries of which AI is going to be used to say that this isn’t going out of our organization, this isn’t going out beyond anywhere outside of our control, I could tell that certainly with a couple of users, there were looks of relief on their face.

“So, even though it sounds quite farfetched, I think it is worth reassuring your users that the security is the main concern.”

Being able to run through the Einstein trust layer with users, showing them the security, the data masking, the zero data retention, and looking at the toxicity detection was “definitely” an “essential step” when talking to users, Will said. 

“Obviously prior to that, we do have a cyber security team at work; we do have compliance and legal, so approaching all those internal departments was obviously a big part of the implementation as well,” he added. 

Will said that his overall advice was to leave “no stone unturned”, making sure you take in all the feedback from internal departments and stress whether something has been cleared by legal and cyber security teams.

“Engage early with all your relevant stakeholders, [listen to] everybody’s concerns and make sure you provide a clear answer to them all, because transparency is trust.”

Final Thoughts

There can be little doubt at this point that artificial intelligence has the potential to streamline a great deal of business processes, and it seems that, as time goes on and these tools become more sophisticated and effective, more businesses will want to take this opportunity. 

But how exactly each company will use AI will differ, sometimes drastically, so having a clear goal, as a user, of what precisely you want the AI to do, seems a crucial part of any successful implementation. 

Learning proper prompting, staying transparent, securing the backing of legal and cybersecurity teams, and dispelling misconceptions – even far-fetched and whimsical ones – are also good ideas for anyone looking to implement AI solutions.