An Agentforce Reality Check ✅: What's Actually Working in Production (And What Still Needs Time to Cook)

How can you use Agentforce to its fullest? Planning with a strategic roadmap.

A year into Salesforce's big bet on autonomous AI agents, we're starting to see what separates hype from reality. If you've been watching the Agentforce announcements, or sitting in demos promising to revolutionize your business, you've probably wondered: Is anyone actually using this thing?

The answer is yes. But it's complicated.

As consultants who live in the trenches of Salesforce implementations, we think it's time for an honest conversation about what Agentforce is delivering right now, what it promises for tomorrow, and most importantly: what you need to have in place before you even think about building your first agent.

The Numbers Look Good (With Context)

Let's start with what's working. Salesforce closed over 5,000 Agentforce deals by early 2025, with 3,000 paid deployments. Nearly half of the Fortune 100 are using Data Cloud and AI together. That's impressive traction for a product barely a year old.

And the customer stories? They're genuinely compelling.

Engine, a travel platform serving over a million travelers, deployed an AI agent named Eva that now handles 30% of customer support cases. The result: a 15% reduction in average handle time and an estimated $2 million in annual savings.

During tax season 2025, 1-800Accountant's Agentforce agent autonomously resolved 70% of administrative chat engagements, freeing up CPAs to handle complex cases that actually require human expertise.

Those aren't vanity metrics. That's real operational impact.

But Adoption Isn't Quite What the Keynotes Suggest

Here's where it gets interesting. Despite the success stories, Salesforce CFO Amy Weaver described Agentforce's near-term financial impact as "modest" during the FY25 Q4 earnings call.

Translation: lots of companies are buying Agentforce, but fewer are fully deploying it at scale.

Some of those 5,000 deals? They're bundled into broader Salesforce platform contracts. Companies are technically paying for Agentforce but haven't turned it on yet. And the ones who have? Many are still in pilot mode, testing a single use case before expanding.

This isn't a criticism. It's a clever implementation strategy. But it's a far cry from the "digital labor revolution" that gets promised on stage.

The Real Blockers: It's Not What You Think

When we talk to clients hesitant about Agentforce, pricing is usually the first thing that comes up.

The original $2-per-conversation model drew immediate pushback. It was unpredictable, expensive at scale, and confusing. (What exactly counts as a "conversation" when a customer follows up 25 hours later?)

Salesforce listened and pivoted in May 2025 to a Flex Credits model: $0.10 per action, down from $2 per conversation. Better? Yes. Perfect? Not quite. Organizations still worry about usage creeping up unpredictably, especially when they're not sure how many "actions" their agents will trigger in production.

But here's the thing: pricing isn't the biggest problem. The biggest problem is data.

Every successful Agentforce implementation we've seen has one thing in common: they did the boring or tedious work first. They cleaned up their Salesforce data. They implemented Data Cloud properly. They documented their business processes and actually knew what they wanted the agent to do.

Are the organizations struggling? They're trying to deploy AI on top of messy CRMs with incomplete records, inconsistent naming conventions, and no unified view of their customer data.

Agentforce can't fix your data problems. What it will do is expose them faster and at scale. Not quite the goal when we talk about scaling AI.

What Actually Works Right Now

Based on what we're seeing in production, here's where Agentforce shines:

Customer service automation for well-defined, repetitive tasks. Think cancellations, basic inquiries, status updates—the stuff your team handles dozens of times a day. If you can document a clear process and your data supports it, agents can handle it autonomously.

Employee-facing use cases with Slack integration. Internal agents helping with HR questions, IT support, and onboarding are showing strong results. Why? The interactions are more forgiving, the data is usually cleaner, and employees are more patient with AI assistants than customers are.

Pre-built industry solutions. Salesforce now offers over 100 industry-specific actions from patient scheduling to vehicle servicing. These templates dramatically reduce time-to-value because you don't have to start from scratch. We’ve built ingestion agents that push against monotonous data entry that can be used across industries. Give us a shout if you’re interested in this kind of use case.

What With Agentforce Still Needs Work

Complex, multi-step workflows. Agentforce can handle sequential tasks, but when business logic gets intricate, like deal approvals with multiple stakeholders or escalation paths that depend on nuanced context, you'll hit limits fast.

Anything requiring deep judgment calls. AI agents are great at following rules. They're terrible at reading between the lines. If your process requires the intangibles: understanding customer frustration, catching subtle red flags, or making calls that balance competing priorities, you still need human oversight.

Sales use cases (for now). This is where the gap between promise and reality is widest. Despite the hype around AI SDRs, most sales teams aren't seeing the ROI yet. Agentforce was designed first for customer service. Sales functionality is improving, but it's not where service capabilities are today with a non-artificial intelligence behind them. (Read: Humans.)

The Honest Prerequisites Checklist

Before you build an Agentforce agent, ask yourself:

Is your Salesforce data actually clean? Not "clean enough". Actually clean. Duplicate records, missing fields, and inconsistent formatting will torpedo your agent’s accuracy.
Have you implemented Salesforce Data Cloud? If you're pulling data from multiple systems, you need a unified view. Your agent can't reason across disconnected silos.
Can you document the process in plain English? If you can't explain exactly what the agent should do in every scenario, you're not ready to build it.
Do you have realistic expectations? Agents work best when they handle 70-80% of a specific task type and escalate the rest. If you're expecting 100% automation, you'll be disappointed.
Can you dedicate resources to iteration? The best Agentforce implementations we've seen treat it like a product, not a project. They test, refine, monitor, and continuously improve.
Are your automation capabilities sized appropriately to your expectations? Remember, Agentforce won't magically generate a document package for you. Agentforce won't intuit how to connect to your ERP to create an invoice or your OMS to process a refund? For all of its wonderful context-fueled planning capabilities, you still need to invest in developing capabilities that the agent can latch onto. Make sure your eyes aren't larger than your stomach in this space, as the agent can only take actions you explicitly expose to it.
What are your UX expectations? Do you want the agent to be a companion embedded in your website, running alongside your users in Salesforce, or are you looking for something more nuanced? We've heard mixed feedback on the standard interaction patterns Salesforce uses to surface agents. That is improving with each release; however, it is essential to consider how agent interactions and their drivers fit into your existing workflows, or otherwise. Remember, Agents can be invoked by Apex and Flows under the hood to help grease squeaky hinges on your more involved processes.

The Atlas Reasoning Engine: What Makes Agentforce Different

What sets Agentforce apart from traditional chatbots isn't just the AI. It's Salesforce’s Atlas Reasoning Engine, the "brain" that powers autonomous decision-making.

Unlike linear chatbots that follow predefined conversation trees, Atlas uses a sophisticated reasoning loop that retrieves data, evaluates options, generates a plan, and refines its approach based on outcomes. It's designed with "System 2" reasoning. That’s the kind of deliberate, multi-step thinking humans use to solve complex problems.

Early results are promising. Atlas has shown a 2x increase in response relevance and 33% improvement in end-to-end accuracy compared to competitor solutions and DIY systems. This isn't just faster automation. It's smarter automation that adapts to context and learns from feedback.

The catch? Atlas is only as good as the data and guardrails you give it. Feed it messy data or vague instructions, and even the most sophisticated reasoning engine will struggle.

Understanding the Data Foundation

Here's a reality that surprises many organizations: Agentforce requires Data Cloud to reach its full potential.

Data Cloud isn't just a nice-to-have. It's the foundation that gives agents context and grounding with near real-time data from multiple systems, and ground their responses in an accurate business context. Without it, agents are limited to whatever data lives directly in your Salesforce org.

For use cases requiring truly instant metrics, agents can go outside of Data Cloud to retrieve the latest and greatest for your users and customers.

Data Cloud unifies structured data (like CRM records) with unstructured data (like emails, PDFs, and chat transcripts), creating a 360-degree view of each customer. It connects to data warehouses such as Snowflake and Databricks via zero-copy integrations, enabling you to leverage existing data investments without duplicating data in Salesforce.

The reality? Most organizations underestimate the data preparation work required. Identity resolution, data mapping, or governance policies aren't afterthoughts. They're prerequisites for agents that deliver accurate, trustworthy results.

Where This Is Headed

Despite the growing pains, the trajectory is clear. Agent creation among early adopters surged 119% in the first half of 2025. Customer service conversations led by agents grew 22x in the same period. Companies that started with one use case are confidently expanding to two, three, and more.

The technology is improving fast. Agentforce 3.0 (released in June 2025) introduced better observability via the new Command Center, lower latency, improved accuracy, and native support for connecting to external tools via the Model Context Protocol. The platform is maturing quickly.

Salesforce recently acquired Informatica for enterprise-grade data management capabilities, signaling its commitment to solving the data quality challenges that hold back AI adoption. This acquisition could be a game-changer for organizations struggling with data governance and integration.

But here's our take: Agentforce is a tool for businesses that have their fundamentals in order. If your Salesforce org is a mess, if your data quality is questionable, if you're still figuring out your processes—fix those things first. Agentforce won't solve those problems. It will just make them more obvious.

For organizations that have done the groundwork? The potential is real. You can genuinely offload repetitive work, scale support without hiring proportionally, and free your team to focus on high-value interactions.

Just don't expect magic. 🧙🏻‍♂️

Expect work. 📊

Innovative, strategic work, with AI as a powerful accelerant. 🔥

The bottom line: Agentforce is delivering real results for companies that approach it thoughtfully. But if you're chasing the hype without doing the tedious prep work, you're setting yourself up for disappointment. The wins are there—they just require more honesty, less buzzword-chasing, and a willingness to solve the data problems everyone wants to ignore.

Ready to explore whether Agentforce makes sense for your organization? Let's talk about what you actually need, not just what sounds cool on a demo. Start with a SprintZero, our process for complex planning and roadmapping, and Contact Digital Mass to start the conversation.