Learnings from early adaptor Agentforce projects

Everyone’s talking about AI agents, but what’s it actually like to work with one in a real business setting? Lets look beyond the hype curve and into the reality of AI agent projects. 

In the Salesforce ecosystem, the focus has since Dreamforce 2024 been heavily focused on Agentforce and understanding the capabilities, roadmap and, most importantly, the impact it can bring into transforming customer interaction.

Agentforce use cases in action

At Columbia Road, we have been working on early adaptor cases around enhancing customer interaction with AI in sales, marketing, and service. Typical use cases in those areas include:

• Creating target groups and segments based on AI-modelled propensity to churn for better marketing impact on loyalty
• Using speech data entry and AI to enhance sales activity and deal logging for better UX, data quality and efficiency
• Detecting and describing underlying data trends and bringing AI-assisted insights to sales users for better sales impact
• Tapping into existing knowledge bases, product materials and technical sales support materials for service agent support for service efficiency and faster human agent onboarding 
• Automatically handling service agent tasks and back office actions with an autonomous AI agent for faster case resolution in simple cases
• Automating product search, selection and quoting to increase sales hit rate 

The pace at which Agentforce as a technology is developing is impressive. What has maybe been even more impressive is how fast customers have been ready to identify opportunities for impact in applying AI agents. 

In this blog, we’ve summarised our main learnings so far to help understand what to expect in a Salesforce Agentforce project, best practices, maturity and how to scale for impact.

1. How does starting an Agentforce implementation look like?

• Start simple – New users should focus on the basics first: standard prompts, actions, and topics. Solve small use cases before expanding or customizing the agent while collecting other impactful ideas into backlog. 
• Agentforce is quick to set up, but fine-tuning takes time – You can get an agent up and running fast, but most of the effort goes into tweaking and testing to get the best results.
• You need Data Cloud for most of the use cases – It’s a must-have for prompt building, proper grounding, and reporting. Without it, you only get a small part of the functionality.
• Ideate, Experiment, tweak on feedback & repeat – The real magic happens in adjusting instructions, actions and prompts based on user and customer feedback. Instead of overthinking every phrase, experiment with different options to see what works best.

2. Agentforce is evolving fast, so be prepared for some gotchas and challenges for early adopters

• There are some bugs or irregularities in the base setup. A few issues, like better user context handling, are still being ironed out. But the product is evolving fast, so expect regular updates, fixes and enhancements. 
• Agentforce evolves through monthly release windows so keep an eye all the time on release notes and beta releases.
• There were some missing features at the time of implementing these projects. If you need more advanced customizations (like better user context handling or mobile record context), you might need workarounds—but Salesforce is working on improvements.
• Some deployment steps aren’t fully automated yet, but new features are rolling out quickly. Automated deployments are in the very close roadmap.

3. Agent performance and resources

• Agents will most likely impress you on their ability. Together with our customers we have been pleasantly surprised by how well Agentforce can answer very industry specific questions, that used to require human thinking, investigation and collaboration between experts.
• Reusing existing functionality you have built over the years is straightforward. Even in heavily customised environments with lots of customisation (and dare I say, technical debt), we had no issues in triggering existing processes and logic
• Keeping context is sometimes tricky. The agent does a great job on initial responses but can sometimes get confused when carrying over context from previous messages.
• There is a great knowledge base. Even though it’s a new product, there are plenty of articles and videos available to help with setup and troubleshooting.

4. Scaling & adoption strategy for impact

• Think big, start small – A good approach is to launch with one agent, refine it, then gradually scale up (e.g., one agent turns into two, then four, etc.). Remember to set a well-defined definition of done for first agent iterations!
• Involve real users early – Understanding the actual questions and needs of customers (or internal teams) makes configuration much easier and more effective.
• Fine-tuning is everything – Small tweaks to prompts and instructions can make a huge difference. It’s best to adjust based on real customer feedback.
• AI-powered setup is different – Unlike traditional software implementations, where 70% of the effort goes into building and 30% into testing, Agentforce flips that—30% building, 70% testing and refining.
• Pricing is usage-based – Works like a pay-as-you-go model, similar to cloud platforms like AWS, Azure, and GCP. Assess cost early on and build a solid business case.

Final thoughts: act now to turn Agentforce potential into real business value

Looking back at these pilots and PoCs, it seems that the time to act is now. Waiting for perfection may mean falling behind. Platform AIs have matured enough to offer a jump-start in scaling AI-led interaction.

Ready to explore further?

Download our canvas and start shaping your first agent use case today: link

Read more on how to use AI for sales impact here: https://www.columbiaroad.com/services/ai