Agentforce vs custom AI agents: which fits a mid-sized business?

Salesforce now offers several ways to pay for Agentforce. When it introduced Flex Credits in May 2025, one Agentforce action consumed 20 credits, priced at US$0.10, and the current UK pricing page lists £400 per 100,000 credits alongside per-user licences and per-conversation pricing. For a mid-sized business, that raises a practical question: should AI agents be built inside Agentforce, or as custom agents that sit across your systems?

The stakes are higher than a licensing decision. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Those three reasons are a useful checklist for comparing Agentforce with custom agents, and this guide uses them throughout.

Quick answer

Agentforce suits businesses whose customer data, users and processes already live in Salesforce, because agents can act on that data natively. A custom agent suits teams whose work spans several systems, or who do not run Salesforce at all. Decide on where your data lives, how many systems the agent must touch, and how usage is metered.

What Agentforce actually is

Agentforce is Salesforce's platform for building and running AI agents inside its ecosystem. According to Salesforce's overview of how Agentforce works, Agent Builder lets admins and developers use natural language to define an agent's goals, instructions, actions and guardrails. Existing workflows, prompt templates, Apex code and APIs can be turned into actions the agent is allowed to take.

The same page describes the Einstein Trust Layer as a set of guardrails that protect data privacy and improve output safety, with features such as dynamic grounding and toxicity detection. It also describes Agentforce Guardrails as a mix of user-defined safeguards and Salesforce-managed protections that keep an agent on its instructions and block off-topic conversations.

In practice, that means an Agentforce agent is only as useful as the actions you give it and the data it can reach. Salesforce's own Trailhead training on grounding an agent with data lists the building blocks: Apex, Flows, prompt templates and MuleSoft APIs for actions, and records, external sources through Salesforce Connect, Data Cloud and knowledge for grounding. Data Cloud now appears in Salesforce material under the name Data 360, so you will see both labels.

If your team lives in Sales Cloud or Service Cloud, that is a strong starting position. The objects, permissions and audit trails the agent needs already exist, and your admins already know the tooling. That native fit is the real argument for Agentforce, and it deserves to be taken seriously before anyone proposes something custom.

How Agentforce is priced

Pricing is where mid-sized businesses most often misjudge the total cost, so it helps to read the vendor's own page closely. Salesforce's Agentforce pricing page lists several models side by side, and the UK figures below are taken from it at the time of writing. Salesforce states that pricing is subject to change, so check the page before budgeting.

The page says an Agentforce action costs 20 Flex Credits and a voice action costs 30. Salesforce's launch announcement for Flex Credits put a standard action at US$0.10. The same pricing page notes that unused Flex Credits do not roll over into later subscription terms, and that Data 360 credits are not included in its worked cost examples.

Those last two details matter. A pilot that handles a few hundred actions a week looks cheap per action. A production agent that handles thousands of enquiries, with Data Cloud running underneath, has a cost profile you can only estimate by modelling your real volumes. Ask for that model in writing before you commit, and compare it with a fixed-scope build rather than with the per-action headline.

Custom agents have their own cost structure, made up of build effort, model usage and ongoing maintenance. Our guide to what an AI agent costs a small business explains how to think about those components without relying on a single figure.

Where Agentforce is the better choice

Agentforce earns its place in a few clear situations. The first is when Salesforce is the system of record for the work being automated. If an agent needs to update opportunities, triage cases or draft replies from knowledge articles, building it where the data lives avoids a whole layer of integration and permission mapping.

The second is when your admins are the people who will maintain it. Salesforce's developer guidance on securing Agentforce actions describes custom actions as autolaunched flows, invocable Apex classes or prompt templates, and says employee-facing actions run as the user who submitted the prompt. That inheritance of user permissions is a real governance benefit if your sharing model is already well designed.

The third is when your data warehouse is one of the platforms Salesforce can federate. Salesforce's Trailhead material on zero copy data federation says Data 360 can query external data at runtime, with no data persisted, and names Google BigQuery, Snowflake, Amazon Redshift and Databricks as supported sources. If your analytics already sit in one of those, an Agentforce agent can reach them without a bespoke pipeline.

The fourth is simpler: your organisation already buys through Salesforce, procurement is settled, and the security review for the vendor has been done. For a mid-sized firm with limited IT capacity, avoiding a second vendor relationship has genuine value.

Where Agentforce gets harder

The difficulties tend to appear when the work does not sit neatly inside Salesforce. Many mid-sized businesses run quoting in one tool, finance in another, project delivery in a third and customer conversations across email, WhatsApp and a helpdesk. An agent that must read from and write to all of them is still possible in Agentforce, but the connection usually runs through MuleSoft APIs, Apex or external services. Each is a build and maintenance task in its own right.

Cost predictability is the second pressure point. Metered pricing rewards agents that handle a modest number of high-value actions and penalises chatty, high-volume workflows. If your use case is front-line enquiry handling with unpredictable peaks, model the credit consumption carefully, including the Data 360 usage that sits outside the examples on the pricing page.

Security design is the third. The same developer guidance says actions should not accept record IDs, object names or field names as user-controlled input, that private service-agent actions must require a verified customer identity, and that LLM-generated output should be treated as untrusted and validated before use. It also says agents should ask for confirmation before actions that modify data or send emails. None of that is unusual, but it is design work that someone on your side must own, whichever platform you choose. See the guidance on securing Agentforce actions for the full list.

Finally, there is the question Gartner raised about whether the problem needs an agent at all. In the same release, Gartner analyst Anushree Verma is quoted as saying that many use cases positioned as agentic today do not require agentic implementations. A platform that makes agents easy to build can make it easy to build agents you did not need. The remedy is the same on any platform: define the process, the owner and the measurable outcome before building anything.

When a custom AI agent makes more sense

A custom agent is built around your process rather than around a platform. It connects to the systems you already use through their APIs, uses whichever language model suits the task, and puts the controls you need around actions that carry risk. It does not replace Salesforce. It can read from and write to Salesforce while also working across the rest of your stack.

Custom tends to win in four situations. The first is cross-system work, such as an agent that takes an enquiry from email, checks stock in an ERP, drafts a quote in your finance tool and logs the result in the CRM. The second is when you are not on Salesforce, or are on a lighter CRM where a full Agentforce deployment would be disproportionate. Our comparison of HubSpot Breeze and custom AI agents covers that scenario for HubSpot users.

The third is when you want control over the model, the prompts and the hosting, for example because of data residency requirements or a need to swap models as better ones appear. The fourth is when the workflow is specific enough that a general platform would need heavy customisation anyway. In that case the platform's native advantage shrinks, and you may be better served by a purpose-built agent.

Custom is not free of trade-offs. You carry more responsibility for monitoring, testing and maintenance, and you need a clear owner on the business side. Our overview of custom AI agents versus off-the-shelf tools sets out those trade-offs in more detail, and the piece on how AI agents integrate with existing tools explains what integration involves in practice.

A practical way to decide

Work through these questions in order. They are designed to surface the real constraints before a vendor conversation anchors the discussion.

The answer is sometimes both. A common pattern is to keep Salesforce-native tasks, such as case summarisation, inside Agentforce, and use a custom agent for work that crosses systems. Splitting the work this way avoids forcing either tool into a role it does not suit. If you want the Salesforce side of that explained, our page on Salesforce AI integration describes how custom agents work alongside the platform.

How Klevere approaches the choice

Klevere builds custom AI agents, with 500+ agents deployed across 50+ projects and 12 industries, so we have an obvious interest in custom work. We try to be honest about where it is not the right answer. If your workflow lives almost entirely inside Salesforce and your admins are comfortable with the tooling, Agentforce may be the shorter route, and we will say so.

Our process starts with the work rather than the technology. We map the process, identify the systems involved, agree the one outcome that matters and note the actions that need human sign-off. Only then do we compare options, which may include a platform agent, a custom build or a combination. That sequence is what protects against the failure modes Gartner describes: costs that grow unnoticed, value that was never defined, and controls added after the fact.

When a custom agent is the right fit, our AI agent development service covers design, build, testing and handover, with approval steps placed wherever an action is hard to reverse. We keep the agent connected to the tools your team already uses rather than asking them to move, and we document what it can and cannot do so that ownership stays with you.

Frequently asked questions

What is the difference between Agentforce and a custom AI agent?

Agentforce is Salesforce's platform for building agents that work on Salesforce data and use Salesforce tooling, with metered pricing. A custom agent is built for your specific process and connects to whichever systems you use, including Salesforce. The practical difference is where the agent lives, who maintains it and how you pay for usage.

How much does Agentforce cost?

Salesforce's UK pricing page lists Flex Credits at £400 per 100,000 credits, conversations at £1.60 each and a user licence at £4 per user per month, with add-ons and editions in US dollars. Pricing is subject to change, and Data 360 credits are not in its examples, so request a written estimate against your own volumes.

Can Agentforce work with systems outside Salesforce?

Yes. Salesforce documents routes through MuleSoft APIs, Apex, Flows, Salesforce Connect and zero copy federation to platforms such as Snowflake, BigQuery, Redshift and Databricks. The question is effort and cost: each connection is something you build, secure and maintain, so count that work when comparing against a custom agent.

Do I need Data Cloud to use Agentforce?

Not necessarily for every use case. Salesforce's training on grounding an agent lists records, external sources and knowledge as options alongside Data Cloud. Data Cloud becomes more important when you need to unify data from several sources. Check licensing and usage costs for your design, because Data 360 credits are separate.

Is a custom AI agent riskier than Agentforce?

Not inherently. Risk depends on how permissions, approvals and monitoring are designed. Agentforce gives you built-in guardrails and permission inheritance, while a custom agent gives you more control but more responsibility. In both cases, restrict what the agent can do, validate its outputs and require human approval for actions that are hard to reverse.

Can we start with Agentforce and move to custom later?

Yes, and it is a sensible route. Keep business rules, prompts and process documentation outside any single platform so they can be reused. That makes it far easier to move one workflow to a custom agent later if costs or cross-system needs outgrow the platform, without rebuilding everything from scratch.

If you are weighing Agentforce against a custom build, a free AI audit is a practical first step. We will look at your systems and processes, and tell you plainly which approach fits, even when that is the platform you already pay for.