Salesforce spent over $500 million marketing Agentforce in the second half of 2025. That kind of spend does something interesting: it educates the market on AI agents while simultaneously generating thousands of buyers who Google "agentforce pricing" and wince at what they find.

If you're one of those buyers - a VP of Operations, CTO, or Head of CX at a mid-market company who saw the Agentforce pitch and started doing math on a napkin - this post is for you. We'll break down what Agentforce actually is, what it actually costs, and when building custom AI agents makes more financial and operational sense.

Fair warning: we build custom AI agents for a living, so we have a bias. We'll be transparent about it. We'll also tell you exactly when Agentforce is the better choice.

What Agentforce Actually Is (and Isn't)

Agentforce is Salesforce's AI agent layer, embedded inside Service Cloud, Sales Cloud, Marketing Cloud, and Commerce Cloud. It launched in late 2024 with seven pre-built agent types that handle tasks like customer service responses, lead qualification, sales coaching, and campaign optimization.

Under the hood, Agentforce runs on the Atlas Reasoning Engine, which pulls CRM data to make decisions. This is a genuine advantage. If your entire business runs on Salesforce, having an AI layer that natively accesses customer records, case histories, and pipeline data without separate integrations is powerful.

But here's the part Salesforce's marketing glosses over: Agentforce is not a standalone AI platform. It's an add-on to an ecosystem. You need active Salesforce licenses, the right Cloud products, and usually a certified Salesforce admin team to configure and maintain agents. As A5 Corp's comparison analysis puts it, Agentforce offers customization within Salesforce's boundaries, while custom AI provides tailored solutions built from scratch.

Think of it like this: Agentforce is a really good kitchen appliance. It does specific things well in a specific kitchen. Custom AI agents are hiring a chef who can cook anywhere.

The Pricing Reality

This is where most buyers' enthusiasm cools. Agentforce offers three pricing models, and you can only pick one per org.

Model 1: Per-Conversation ($2/conversation)

Every customer interaction that touches an Agentforce agent costs $2. The math gets uncomfortable fast:

Monthly Conversations Monthly Cost Annual Cost
10,000 $20,000 $240,000
30,000 $60,000 $720,000
50,000 $100,000 $1,200,000
100,000 $200,000 $2,400,000

Model 2: Flex Credits ($500 per 100K credits)

Each agent action burns roughly 20 credits. A typical conversation with 6 responses and 2 actions uses about 160 credits. That works out to roughly 625 conversations per $500 block - or about $0.80 per conversation at low complexity. But credits burn faster on complex workflows with multiple tool calls, and Cirrius Solutions' cost analysis notes that real-world usage often exceeds initial estimates.

Model 3: Per-User ($125/user/month)

This covers unlimited internal agent usage for employees but is separate from customer-facing agents. It's a different product for a different use case.

The hidden layer: None of these prices include the underlying Salesforce licenses you already need. Service Cloud, Sales Cloud, Data Cloud - those are separate line items. Industry analysts peg the total cost of an enterprise Agentforce deployment at over $13,600 per user per year when you stack everything together.

The CFO problem is real. You're choosing between three consumption models with different cost curves, layered on top of existing license fees, with no volume cap. A retail brand handling 30K customer interactions monthly on the conversation model pays $720K per year in Agentforce fees alone - before Salesforce licenses, before implementation.

Compare that to a custom-built agent on a flat monthly retainer where the cost is identical whether you handle 30K or 300K interactions.

The Lock-In Problem

Pricing aside, the deeper issue is architectural. Agentforce only works inside Salesforce. Your agent logic, your workflows, your trained behaviors - all of it lives in Salesforce's environment under Salesforce's terms.

This creates three specific risks:

Risk 1: CRM migration kills your AI investment. If you ever move to HubSpot, Dynamics, or anything else, your Agentforce agents don't come with you. Every dollar spent configuring and optimizing those agents is a sunk cost tied to one vendor. Greg Banks highlighted this concern on LinkedIn, noting that TCO analysis must account for the growing dependency Agentforce creates.

Risk 2: Your AI can't see your whole business. Most mid-market companies don't run everything on Salesforce. They use Shopify for ecommerce, NetSuite for ERP, Zendesk or Intercom for support, industry-specific platforms for their core operations. An AI agent trapped inside the CRM only sees one slice.

Consider a few real scenarios:

  • A health system using Epic or Cerner for electronic health records needs AI agents that work across clinical and administrative systems, not just the CRM layer
  • A retail chain running Shopify, NetSuite, and Zendesk needs AI that connects all three, not an agent that can only access Salesforce customer records
  • A wealth management firm using Orion or Black Diamond for portfolio management needs AI integrating with those platforms, not just the CRM

Risk 3: Salesforce controls the roadmap. If they reprice, deprecate features, or change the Atlas engine, you adapt on their timeline. Gartner's 2025 analysis of vendor lock-in risks in AI platforms flags this as one of the top concerns for enterprises adopting embedded AI - the vendor's strategic priorities may diverge from yours.

With custom-built agents, you own the code. You can run it on your infrastructure, connect it to any system, and modify it without asking permission.

When Custom AI Agents Make More Sense

Custom agents aren't always the answer. But for specific situations, they're clearly the better investment.

Your operations span multiple systems. If the AI needs to pull from your CRM, your ERP, your support platform, and your industry-specific tools in a single workflow, a custom agent built on your infrastructure handles that natively. Agentforce requires bolting on integrations for anything outside Salesforce.

Your workflows don't fit templates. Agentforce's seven pre-built agent types cover common patterns. If your process is a standard lead qualification flow, those templates work. But if your business has specific routing logic, compliance requirements, or multi-step approval chains that don't map to Salesforce's templates, you're fighting the platform instead of using it.

Predictable budgeting matters. Consumption pricing means your AI costs fluctuate with volume. That's fine if your interaction volume is steady and low. It's a budgeting headache if volume spikes seasonally or you're actively trying to grow usage. A flat retainer means your cost is your cost.

You want to own the software. This is philosophical but practical. At the end of an OpenNash engagement, you own the agent code, the workflow logic, and the deployment. There's no ongoing license fee. No per-conversation charge. No vendor that can change terms. As OpenAI's practical guide to building agents emphasizes, organizations that own their agent architecture retain the flexibility to iterate independently.

You're not deeply embedded in Salesforce. If Salesforce is one of several tools rather than your operating system, Agentforce's main advantage - native CRM data access - doesn't justify the lock-in and cost.

When Agentforce Is the Right Call

Here's where we give credit. Agentforce is a strong product for the right buyer.

You're already all-in on Salesforce. If your sales team lives in Sales Cloud, your support team lives in Service Cloud, and your marketing team lives in Marketing Cloud, Agentforce's native data access is a real advantage. No integration needed. The agent sees everything your teams see.

Your use cases are CRM-adjacent. Lead scoring, case routing, opportunity management, email response drafting - these are exactly what Agentforce's pre-built agents handle well. If your AI needs start and end inside the CRM, the templates save real development time. Zeeg's 2026 analysis of AI sales agents ranks Agentforce highly for teams that need quick deployment within existing Salesforce workflows.

You have the admin team. Agentforce configuration isn't drag-and-drop. You need Salesforce admins who understand Flows, permission sets, and the Data Cloud data model. If you already have that team, the learning curve is manageable.

Your volume is low enough for the pricing to work. At 5K conversations per month, Agentforce costs $10K monthly. That might be cheaper and faster than a custom build for a narrow use case. The break-even point where custom starts winning varies, but it's typically somewhere around 15K-25K monthly conversations.

You're not planning to leave Salesforce. If Salesforce is your platform for the foreseeable future and you're comfortable with that commitment, the lock-in risk is theoretical rather than practical.

The Comparison at a Glance

Factor Agentforce Custom AI Agents
Pricing model $2/conversation or Flex Credits Flat monthly retainer
Software ownership No - you license access Yes - you own the code
CRM dependency Salesforce required Works with any CRM or none
Customization Configure pre-built templates Built from scratch for your business
Cost at 50K convos/mo ~$100K/month + SF licenses Flat fee regardless of volume
When contract ends Lose access to everything Keep the software forever
Implementation Requires SF admin team Handled by the builder
Data residency Salesforce infrastructure Your environment, your rules
Best for SF-native enterprises Businesses wanting ownership and flexibility

Making the Decision

Here's a simple framework. Answer these four questions:

  1. Does more than 80% of the data your AI agent needs live in Salesforce? If yes, Agentforce has a natural advantage. If no, you'll spend as much on integrations as you would on custom development.

  2. Are you comfortable with your AI costs scaling linearly with success? If your goal is to automate more interactions over time, consumption pricing means success gets expensive. Flat pricing means success gets cheaper per interaction.

  3. Would you be okay if Salesforce doubled Agentforce pricing next year? They've done it with other products. If that scenario makes you nervous, ownership matters.

  4. Do you have a Salesforce admin team already? If not, you're adding headcount for Agentforce that you wouldn't need for a managed custom solution.

No single answer disqualifies either option. But if you answered "no" to three or more, custom agents are likely the better path.

The AI agent space is moving fast. Salesforce is spending heavily to own the narrative, but the narrative isn't the whole story. The buyers who come out ahead will be the ones who looked past the marketing, did the math, and chose the architecture that fits their business - not the one with the biggest ad budget.

Want to see what custom AI agents would look like for your business? Talk to us.