You’ve probably already run the pilot. The summaries looked good. The recommendations made sense. Everyone in the room nodded. Then it didn’t move to production. Or it did, and three months later, you stopped using it. Is that because they don’t have conversation catch-up or because it’s hallucinating to fill in the data gap of disconnected systems?
You’re not alone. A year ago, most of the conversation around Agentforce centered on what it could do. Today, more of it centers on why deployments stall between pilot and production. Salesforce Ben has covered the governance questions. RevOps teams have written about the operational gaps. Solutions4SF has documented the cost overruns. None of these sources argues that Agentforce doesn’t work. They’re arguing that deploying it and operationalizing it are two different jobs, and most teams only budget for the first one.
Quick answer: Agentforce pilots in contact centers usually fail for the same handful of reasons: incomplete customer data, technical debt in the surrounding systems, unclear governance, an underestimated integration bill, and an assumption that Agentforce can replace infrastructure it was never built to replace. Voice-heavy contact centers hit every one of these issues first, and usually hardest.
What’s Actually Standing Between You and a Working Agentforce Deployment
None of what follows is a knock on the AI. Agentforce’s reasoning engine does what Salesforce says it does. The friction shows up in the layer underneath it, the data, systems, and decisions your organization made long before anyone mentioned AI.
Data quality and context.
Agentforce reasons over whatever reaches Salesforce, and for most organizations, that’s a fraction of the full picture. Customer history lives in Salesforce. Call transcripts, recordings, and interaction details often live somewhere else entirely. Agentforce sees the part of the story that made it into the CRM and treats that as the whole story.
Technical debt in your existing environment.
You didn’t build your stack with Agentforce in mind, because Agentforce didn’t exist when you built it. Avaya, Cisco, Genesys, Teams, SBCs, carrier contracts, CCaaS platforms: every one of those represents a real investment and a real dependency. Salesforce Ben and the broader RevOps community have both flagged this as one of the most underestimated barriers to AI adoption, not because the technology is hard, but because untangling years of infrastructure decisions is.
Governance and trust.
As Agentforce takes on more autonomous actions, someone has to decide what it’s allowed to do without a human checking first. Salesforce Ben has covered this extensively: the conversation shifts fast from “what can the AI do” to “who’s accountable when it’s wrong.” Most organizations haven’t had that conversation yet, and Agentforce surfaces the need for it the moment it goes live.
The gap between demos and production.
Every source covering Agentforce, vendor-side, and independent, converges on this point. Demos run on clean data and a handful of well-defined workflows. Production runs on duplicate records, inconsistent processes across teams, and years of edge cases nobody documented. A pilot can look great and still not survive contact with the real environment.
Underestimated cost and complexity.
Solutions4SF and ongoing community discussion both point to the same pattern: the license is the smallest line item. Data preparation, integration work, governance design, and change management routinely cost more than the platform itself, and most budgets don’t account for that until they’re already over.
Agentforce is an intelligence layer, not a replacement.
This is the thread connecting everything above. Salesforce Ben, the RevOps research, and AMC’s own work with DaVinci all land on the same conclusion: Agentforce coordinates actions, surfaces insights, and automates workflows, but it doesn’t replace the systems that already run your business. It depends on them. If those systems aren’t feeding it the right information, Agentforce can only work with what arrives.
Why Contact Centers Feel All of This First
Every department deals with some version of these six challenges. Contact centers feel them earlier and more visibly, because customer conversations generate an enormous share of the context your business runs on. And most of it never reaches Salesforce in a form that Agentforce can use.
So when Agentforce recommends a next step or summarizes “the situation,” it works from a file missing most of the conversation. And your agents notice. An agent picks up a call, and Agentforce surfaces a recommendation that doesn’t reflect the conversation they had with this same customer ten minutes ago, because that call hasn’t synced yet. They shrug it off once. By the third time, they’ve mentally filed Agentforce under “sometimes wrong, don’t bother checking.” At that point, you don’t have an adoption problem you can fix with training. You have an agent population that learned, correctly, that the tool lacks the full picture, and stopped relying on it.
This is also why two contact centers can run what looks like the same pilot and get opposite results. The difference usually isn’t the use case or the Salesforce configuration. It’s whether the contact center and Salesforce were already exchanging information in real time, or close to it, before Agentforce arrived.
One Fix to Address Your Agentforce Implementation Challenges: Your Voice Architecture
Here’s the part that should be encouraging: most of what’s described above traces back to one thing, and that one thing is fixable without ripping out anything you’ve already built.
Voice is where most of your richest, least-captured customer context originates: recordings, transcripts, routing history, sentiment, resolution paths. If you plan for how that data moves, who owns it, and where it lands, you address the data quality problem, reduce the technical debt friction, and give governance something concrete to govern. You also stop the trust breakdown before it starts, because agents start seeing context that actually matches the call they’re on.
This is architecture planning, not a rebuild. It means mapping where voice data lives today, how it moves between systems, who owns each piece, and which platform serves as the source of truth going forward. AMC’s DaVinci Voice Toolkit was built around exactly this kind of planning, helping organizations work out data ownership and voice orchestration before they scale AI on top of it.
Get this right, and agents see context that lines up their call, customer history feels complete, and post-call work starts happening on its own, because the data finally lands where it needs to.
Steps to Take Before You Implement & Scale Agentforce
You don’t need to solve all six challenges before you start. But you do need answers to a few questions, and most teams can get them faster than they expect.
- Start by finding out where your voice data actually lives right now. Recordings, transcripts, dispositions: pick three recent calls and trace where each piece of information ended up. If you can’t find all of it, you’ve just found your first integration project.
- Next, check how long it takes for a disposition to reach Salesforce after an agent sets it. Minutes are good, but hours mean Agentforce is working from stale information during exactly the interactions where real-time matters most.
- Then look at what an agent’s screen actually shows when a call connects. Does it reflect anything from the customer’s last interaction, or does it start from zero every time?
- Finally, before you expand use cases, decide who owns voice data governance: who decides what gets recorded, where it’s stored, how long it’s retained, and who can access it. This conversation gets easier before AI is in the room, not after.
None of this requires replacing your telephony platform, your CCaaS provider, or your carrier relationships. It requires knowing where your voice data lives and building the connections that let it reach Salesforce in a form Agentforce can actually use.
Connecting Your Existing Voice Stack to Agentforce, Without Ripping It Out
Most contact centers we talk to run a voice stack that works: carriers, numbers, routing logic, and compliance setups that took years to get right. DaVinci connects what you already run to Salesforce and Agentforce Voice through a bring-your-own-carrier approach. It sits underneath as the voice control layer, so transcripts, call events, and routing data reach Salesforce in a form Agentforce can use. You keep the telephony environment your team built. The data finally has somewhere to go.
We’ve run this at real scale: one deployment processes roughly 1.9 billion minutes of transcription a year across more than 10,000 agents, with Salesforce Voice running on top of telephony infrastructure that already existed. At that volume, voice carries most of the customer relationship, and the systems underneath it have to keep pace.
If you’re working through any of the six challenges above, particularly the voice piece, we’re happy to look at your setup with you.
