Agentforce
Agentforce vs Zendesk AI agents: where the agent should sit
Both now deflect tier-1 tickets, both ground answers in a knowledge base, and both bill by the resolution. The choice is not which model is smarter. It is where your system of record and the actions you want the agent to take live, and the pricing rewards very different volumes. A practitioner comparison for teams running one, the other, or both.
You already run Zendesk for support. A Salesforce rep is telling you Agentforce will deflect the same tickets and keep the customer record in one place. Both statements can be true. Neither one answers the question you have, which is where the agent should sit.
My rule, before any feature list: put the agent next to the system of record it grounds on and the systems it has to act in. If your tickets, your macros, and your help centre live in Zendesk and the agent mostly needs to answer from articles and take a few actions inside Zendesk, Zendesk’s own AI agents are the lower-lift path. If the refund, the order, the subscription, and the customer’s real history live in Salesforce, Agentforce grounds and acts on that natively and a Zendesk agent has to reach across an integration to touch any of it.
Everything else in this comparison is detail on that one sentence.
What each one is
Zendesk AI agents are the product line built on Zendesk’s acquisition of Ultimate. They sit inside Zendesk’s support suite, ground their answers in Zendesk Guide (the help-centre knowledge base), and resolve conversations in the channels Zendesk already runs: web, email, and messaging. The newer tier reasons through a conversation rather than following a fixed bot script, and can take authorised actions and call third-party APIs mid-conversation through an integration builder.
Agentforce is Salesforce’s agent platform. It grounds in Data 360 and retrievers or in Unified Knowledge, takes actions through Apex, Flow, MuleSoft, and MCP servers, and runs inside the Salesforce platform under the Einstein Trust Layer. A first service agent is scoped to topics and governed actions rather than a single help-centre.
The headline difference is not intelligence. Both reason over a conversation and both ground in a knowledge base. The difference is what the agent can see without an integration, and what it can do without one.
Grounding: the knowledge base is the ceiling on both
On both platforms, retrieval quality decides resolution rate, and the knowledge base is the ceiling. An agent cannot resolve a question whose answer is not written down somewhere it can retrieve.
Zendesk grounds in Guide. If your help centre is well-maintained, Zendesk’s agents answer well, and the setup is close to turnkey because the content and the agent live in the same product. The limit is the one every retrieval system has: the agent resolves the share of tickets that a good article can answer, and stalls on the ones that need the customer’s own data.
Agentforce grounds in whatever you put in Data 360, which is the same place your CRM records, your ingested external data, and your knowledge already sit. That is the real advantage, and it is specific: an Agentforce agent can ground an answer in this customer’s last three orders and the refund policy in the same retrieval, because both are in the data foundation. A Zendesk agent gets the policy from Guide and has to call an API to learn anything about the order.
If your tier-1 volume is mostly “how do I” questions, that gap barely matters and Zendesk’s turnkey grounding wins. If it is mostly “where is my” and “change my”, the account data is the answer, and the agent that sits on top of it has less to reach for.
Actions: inside the suite versus across the platform
A deflected question is worth something. A resolved request is worth more, and resolution means the agent did something rather than only answered.
Zendesk’s newer agents take authorised actions and call external APIs during a conversation: update a ticket field, look up an order through a connected system, trigger a refund via an integration. That covers a lot of tier-1 if the action is a single call to a system Zendesk can reach.
Agentforce’s action surface is the Salesforce platform. An action is an Apex method, a Flow, a MuleSoft API, or an MCP server call, governed by the running user’s permissions and sharing. An action looks like configuration:
Action: Issue refund
Input: Case.OrderId, refundAmount
Runs: RefundService.issueRefund (Apex, WITH USER_MODE)
Guard: refundAmount <= Order.RefundableBalance
Escalate if: refundAmount > tierLimit for the running agentThe depth is the point. If the refund needs to check an entitlement, write to a custom object, respect a sharing rule, and post an event to another system, that is native Salesforce work and the agent inherits all of it. On Zendesk the same logic lives in the integration you build and maintain outside the agent.
This cuts the other way for simple cases. If “resolve” means close the ticket and send a templated reply, Agentforce’s platform depth is weight you are not using, and Zendesk’s in-suite actions are enough.
Pricing: both bill by the resolution, and reward different volumes
Both have moved toward outcome pricing, which makes a direct comparison possible and a little treacherous.
Zendesk bills per automated resolution. Third-party reporting puts the rate near $1.50 per resolution on a committed-volume plan and around $2.00 pay-as-you-go, with a monthly pool of resolutions included per agent seat before billing starts. Reporting also says that since around May 2026 only verified resolutions are billed, so a conversation the agent escalates or one where the customer goes silent is not counted. Confirm all of this against Zendesk’s current pricing page, because the packaging changed during 2026 and the public figures are not from Zendesk’s own documentation.
Agentforce gives you more than one meter, which is a feature and a trap. The Flex Credits model runs at $500 per 100,000 credits, roughly $0.005 a credit, and a standard action consumes about 20 credits, so near $0.10 an action. The Conversations model is listed at $2 per conversation. And the pay-per-resolution help agent bills $2 per resolution, the closest apples-to-apples line against Zendesk’s number.
| Zendesk AI agents | Agentforce | |
|---|---|---|
| Grounds in | Zendesk Guide (help centre) | Data 360, retrievers, Unified Knowledge |
| Actions | Authorised actions + API integration builder, inside the suite | Apex, Flow, MuleSoft, MCP, governed by platform permissions |
| Runs in | Zendesk support suite | Salesforce platform, under the Einstein Trust Layer |
| Resolution price | ~$1.50 committed / ~$2.00 PAYG per automated resolution (reported) | $2 per resolution (help agent), or Flex Credits / $2 per conversation |
| Lowest lift when | Support already lives entirely in Zendesk | System of record and actions live in Salesforce |
The volume logic is what matters. Per-resolution pricing is clean and predictable when resolutions are the unit you care about and most conversations resolve in one pass. Flex Credits reward an org that can model action counts and wants one meter across service, sales, and custom agents rather than a per-resolution line for each. Model both against your real interaction mix before a pricing page decides your architecture, the same caution I gave in the OpenAI comparison.
The hybrid that looks clever and usually is not
There is a tempting middle path: keep Zendesk for ticketing, and ground an agent in Salesforce data through an API so you get the account context without moving platforms.
It works, and it is usually the worst of the three options. You now run two platforms, you pay the integration tax to keep them in sync, and the agent grounds on data it reaches across a network boundary with the latency and failure modes that adds. The latency of a grounded agent is sensitive to exactly this kind of extra hop.
Choose the hybrid only when an organisational reality forces it, a support team contractually on Zendesk and a sales org on Salesforce that will not merge this year. If the two can converge, converge them, and put the agent on the side that owns the customer record.
How to decide
Three checks, in order.
Start with where the customer record and the action live. If both are in Salesforce, the grounding and action advantages are Agentforce’s to lose. If support is self-contained in Zendesk and the agent mostly answers from articles, Zendesk is the lower-lift win.
Then define what “resolved” means for your top tickets. An answer from a help article, both do, and Zendesk is faster to stand up. A refund, an order change, or an entitlement check against your own data favours the platform that owns that data.
Last, price it against your real volume and mix. Put the per-resolution line next to the flex-credit line using your own numbers, and treat every third-party price as provisional until you see it on the vendor’s own contract.
If you are coming off legacy chatbots rather than choosing fresh, the Einstein Bots migration path is a separate decision with its own write-up, and the human handoff design matters more than the platform choice either way.
Understanding the basics
Is Agentforce better than Zendesk AI agents?
Neither is better in the abstract. Agentforce wins when the customer record and the actions live in Salesforce, because it grounds and acts on that data natively. Zendesk’s agents win when support lives entirely in Zendesk and the job is mostly knowledge-base deflection, because setup is close to turnkey and there is no new platform to run.
Can Zendesk AI agents use Salesforce data?
Only through an integration. Zendesk grounds natively in Zendesk Guide, so to use Salesforce records an agent has to call an API during the conversation, which adds latency and a pipeline to maintain. If most of your tier-1 answers depend on CRM data, the agent that sits on that data has less to reach for.
How does Agentforce pricing compare to Zendesk’s per-resolution model?
Zendesk bills per automated resolution, reported near $1.50 committed and $2.00 pay-as-you-go. Agentforce offers a $2 per-resolution help agent that lines up directly, plus Flex Credits at about $0.10 per action and a $2-per-conversation option. Per-resolution pricing is predictable; Flex Credits suit an org that wants one meter across many agent types. Confirm both against current vendor pricing.
Do I need Data Cloud for an Agentforce service agent?
Not always, but grounding quality decides the agent’s quality, and Data 360 is where that grounding lives. An agent grounded in thin data deflects poorly regardless of platform, which is why the data foundation is usually the first piece of work, not the agent.
The agent belongs next to the data
Strip away the feature lists and the pricing tables and the decision is physical. An agent is only as good as what it can retrieve and what it can do, and both are cheapest and most reliable when they sit next to the agent rather than across an integration.
So the question is not which vendor has the smarter model this quarter. It is where your system of record lives. Put the agent there. If that is Zendesk, Zendesk’s own agents are the right call, and I would not sell you a platform migration to avoid them. If it is Salesforce, Agentforce is not a bolt-on chatbot, it is the agent that already stands on your data.
If you are weighing Agentforce against an incumbent support tool and want the grounding, action depth, and real per-interaction cost modelled against your own ticket mix, talk to us, or start with the Agentforce ROI calculator to put a number on tier-1 deflection first.
