Salesforce AI review 2026: Agentforce, Einstein, pricing and is it worth it?

Quick answer first: Salesforce AI is the set of AI tools built into Salesforce's CRM. Today that mostly means Agentforce, its platform for AI agents that answer customers and do work inside your records, plus the older Einstein features and Data 360 (the product that used to be called Data Cloud) underneath. I'd say it's probably the most deeply wired-in AI you can get if Salesforce already runs your business. It's also one of the hardest to price.

I'll be straight with you about how this review was made. Nobody on our team ran Agentforce in a live org, so you won't find made-up stopwatch timings here. Instead, we pulled together Salesforce's own Dreamforce 2026 announcements, independent pricing breakdowns, the published rate card details and other reviewers' findings, all as of late October 2026. Where someone else did the hands-on work, I'll name them.

My short version: brilliant plumbing, confusing meter. If your sales and service teams live in Salesforce, Agentforce can do things no bolt-on chatbot can. If you were hoping for a simple monthly price, pour yourself a coffee before opening the pricing page.

The Salesforce AI stack

"Salesforce AI" isn't one product. It's a stack, and the names have changed so often that even Salesforce admins joke about it. Here's the 2026 map:

• Agentforce is the agent platform. Agents can answer questions, update records, run Flows and call APIs, rather than just writing text. Agentforce 360, unveiled at Dreamforce in October 2025, was described as the fourth edition of the platform, and it's still the current platform name heading into November 2026.
• Einstein is the original Salesforce AI brand. In 2025, Agentforce add-ons and Agentforce 1 Editions replaced the older Einstein add-ons and Einstein 1 Editions entirely. The name lives on in pieces like the Einstein Trust Layer, Einstein Bots (which can now be upgraded to Agentforce agents) and AI Models, which Salesforce Break notes was formerly Einstein Studio.
• Data 360, formerly Data Cloud, is the data layer that pulls CRM records, warehouse data and unstructured content together so agents have context. Salesforce Break describes it as the platform that powers Agentforce with customer context.
• AIforce is the newest layer, announced at Dreamforce 2026. Salesforce says it brings the full power of Salesforce to wherever you already work, whether that's Claude, Slack or Salesforce itself.

So when people search "salesforce einstein ai" or "einstein ai salesforce" in 2026, what they'll actually buy is mostly Agentforce, running on Einstein's trust plumbing and Data 360's data.

The latest Agentforce release

Dreamforce 2026 didn't come with a new version number. Instead, Salesforce shipped a pile of features on top of the Agentforce 360 platform. Its own recap lists 21 announcements. The ones that matter most:

• Agentforce Coworker, an AI teammate inside Salesforce that answers questions across your CRM and Slack and can build and execute a plan. Salesforce lists it as generally available.
• Job-ready agents. Prebuilt agents with names. Cyntexa counts seven: Piper, Hunter, Casey, Paige, Fin, Carter and Marshall. Piper handles inbound sales development, Marshall handles back-office and supply chain work, and eesel reports that Casey, the service agent, has been generally available since September 11, 2026.
• Long-horizon agents that chase a goal over days or weeks, such as "re-engage my at-risk deals", checking in for approval along the way. Salesforce lists them as pilot now, GA November '26.
• Koa, Salesforce's first CRM reasoning model, built on NVIDIA Nemotron and trained on synthetic scenarios only. It's in pilot, with GA expected Winter 2026.
• Claudeforce and Slackforce. Salesforce data and actions inside Claude (beta) and Slack (GA in October). Cyntexa reports Claudeforce uses MCP and ships with 37 prebuilt sales skills.

The theme is obvious once you squint: Salesforce wants its data and permissions to follow you into whatever AI tool you open. I think that's a smart bet, and it's the most interesting thing about the platform right now.

Key features

Agents that take real actions

This is the big one. An Agentforce agent doesn't just chat. As eesel's reviewer explains, actions reuse what you built: if your admins already have a "process return" Flow, the agent can call it. It works within your sharing rules and field-level security, so it can't read what its user profile can't.

That's the difference between Agentforce and most of the tools in our AI chatbots for business roundup. Those can answer questions. Agentforce can change the record.

Agentforce Builder, Agent Script and subagents

You build agents in Agentforce Builder. Agent Script, introduced with Agentforce 360, lets you mix fixed rules with AI reasoning, which CEO Insights India described as handling "if/then" scenarios more predictably. One naming trap: eesel notes that what Salesforce used to call topics have been called subagents since April 2026, with no change in function.

Honestly? Agent Script is the feature I'd care about most as a buyer. Refund limits and identity checks shouldn't be left to a model's mood.

Einstein Trust Layer and model choice

Agentforce 360 runs on reasoning models from Anthropic, OpenAI and Google Gemini. Requests run through the Einstein Trust Layer, which eesel says adds zero data retention, toxicity detection, and an audit trail. You can also bring your own model: Salesforce Break lists OpenAI, Microsoft Azure, Google Vertex AI, or Amazon Bedrock as supported providers, with requests still routed through the Trust Layer.

Data 360 grounding

Agents are only as good as what they can see. Data 360 unifies records, warehouse data and documents so the agent answers from your data, not the open web. It's powerful, and it's also a separate meter (more on that below).

Testing and observability

Salesforce seems to have heard the "it worked in testing, then went weird in production" complaints. New at Dreamforce 2026: Agent Optimizer, which reads production sessions and ranks failure patterns (GA October), Agent Health Monitoring with 16+ built-in metrics, A/B experiments and voice conversation testing. There's also a one-click path to convert an existing Einstein Bot into an Agentforce Service Agent.

Voice

Agentforce Voice handles phone conversations, and Salesforce added six new languages at Dreamforce: Spanish, Italian, German, Portuguese, Japanese and Canadian French.

Salesforce AI pricing and plans

Deep breath. Salesforce sells Agentforce several ways, and you can't simply pick the cheapest one. Salesforce's own pricing pages blocked our automated checks, so the table below comes from independent breakdowns that quote the official pricing page and rate card (eesel, Salesforce Break and eesel's review). Prices are list, in US dollars, checked October 2026.

Flex Credits in practice

One standard action costs 20 credits, which is $0.10. Voice actions cost 30 credits ($0.15). Jitendra Zaa's guide adds that each action includes up to 10,000 tokens. A Help Agent resolution costs 400 credits, or $2, according to eesel's reading of the rate card. Prompts to Salesforce-managed models draw 2, 4 or 16 credits depending on the model tier, per Salesforce Break.

Salesforce's own examples, as quoted by eesel, run from 1 action ($0.10) for an onboarding question to 6 actions ($0.60) for a field-service scheduling request.

The fine print that bites

• The $2 conversation is fading. eesel reports the live pricing page now labels it "For existing Agentforce Conversations customers". New buyers land on Flex Credits.
• Pick one model per org. You can't mix Flex Credits and Conversations in the same org, and unused Flex Credits don't roll over.
• The credit pool number moved. July write-ups listed 2.5M Flex Credits per org per year with Agentforce 1 Editions; eesel's later review lists 2.75M with the Max Edition. Check your quote.
• Other meters still run. Data 360 work is billed in credits on its own tiered rates, and Salesforce's worked examples, eesel notes, "do not include other costs like Data 360 credits or other consumption services".
• You still need the seat. Agentforce needs Enterprise, Performance, Unlimited or Developer Edition. For service teams, eesel lists Service Cloud from $195 (Core) to $550 (Max) per user per month, with Casey arriving at the $395 Advanced tier.

So what does the free plan really include? Foundations lets you build and test agents at no cost, and eesel says Salesforce includes credits for the first 1,000 Agentforce for Service conversations. It isn't a free way to run agents at scale, and it sits on top of a paid Salesforce org.

Performance and independent testing

We didn't benchmark Agentforce ourselves, so here's what the people who've lived with it report.

eesel's reviewer, who went through Salesforce's docs, the rate card and the reviews on G2, gives Agentforce 3.7/5. The upside, in their words: inside a clean Service Cloud org, "nothing else is this native." The downsides: G2 reviewers rate it 4.3/5, eesel reports, while repeatedly naming "setup takes real admin skill" and hard-to-forecast consumption pricing as the two pains.

The production gap is real, too. eesel quotes a Salesforce architect: "The agent can behave perfectly in preproduction and then behave completely differently in production." That's exactly why the new observability and A/B testing tools matter.

On results, Salesforce reports that 50% of Engine's chat inquiries are fully resolved by its help agent, per eesel. That's a vendor-supplied figure, so it's likely a best case, not a promise. And when Agentforce 360 launched, Salesforce said it already served 12,000 customers, including Lennar, Adecco and Pearson.

Cost is the loudest complaint. eesel cites a G2 reviewer: "Due to the pricing per dialogue, any loops or logic failures quickly 'eat up' the budget." An agent that loops still bills. Ouch.

Who Salesforce AI is for

• Salesforce-first companies with a capable admin team. If Salesforce is your system of record, Agentforce's native access to records, Flows and permissions is hard to beat.
• Large service operations handling high volumes across chat, messaging and voice, where even partial automation pays back quickly.
• Regulated or security-conscious teams that want the Trust Layer, permission-aware agents and audit trails. Our enterprise AI solutions roundup covers how other platforms handle the same needs.

Who should skip it? Small teams on Starter or Pro Suite, anyone without an admin to own the build, and anyone who needs a fixed, predictable AI bill this quarter.

Salesforce AI alternatives worth considering

• Microsoft Copilot makes more sense if your company runs on Microsoft 365. See our Microsoft AI review.
• Atlassian Rovo fits teams whose work lives in Jira and Confluence rather than a CRM. Our Rovo AI review has the details.
• A standalone support AI, such as eesel, which pitches one price per resolution, no credits to size. Tools like this sit on top of your helpdesk and are lighter to set up, though they can't act as deeply inside Salesforce. Our best AI chatbots for business list compares more options.

If you're still working out what "agents" means for your business in the first place, start with our explainer on agentic AI.

Verdict: is Salesforce AI worth it?

For the right company, yes. Agentforce in late 2026 is a serious platform: real actions, permission-aware agents, a growing roster of job-ready agents and, finally, decent tools for testing and monitoring them. The AIforce push into Claude and Slack is genuinely clever.

But the bill is a puzzle. You pay for the seat, then the actions, then the data, and the forecast is fuzzy until you've run it for a while. I'd give it 4 out of 5, and I think that's about right: top marks for features, lower marks for ease of use and value.

Use it if Salesforce is your system of record and you have someone to own the build. Skip it if you want a simple chatbot with a simple price.

FAQ

What is Salesforce AI?

Salesforce AI is the AI built into Salesforce's CRM. In 2026 it centres on Agentforce, a platform for AI agents that answer customers and act on CRM records, supported by the Einstein Trust Layer for security and Data 360 for unified customer data. AIforce, announced at Dreamforce 2026, extends it into Claude and Slack.

Is Salesforce Einstein AI still available?

Partly. The Einstein add-ons and Einstein 1 Editions were replaced by Agentforce add-ons and editions in 2025. The Einstein name survives in the Einstein Trust Layer and Einstein Bots, and Salesforce now offers a one-click upgrade from an Einstein Bot to an Agentforce Service Agent.

How much does Agentforce cost?

Building is free with Salesforce Foundations. Running agents costs $500 per 100,000 Flex Credits, roughly $0.10 per standard action, or $125 per user per month for unmetered employee add-ons. Max Edition starts at $550 per user per month. The older $2 per conversation model is now aimed at existing customers.

What is the latest version of Agentforce?

Agentforce 360, launched at Dreamforce in October 2025 as the fourth edition, is still the current platform. Dreamforce 2026 added features rather than a new number: Agentforce Coworker, job-ready agents, long-horizon agents, Agent Optimizer, the Koa reasoning model and AIforce.

Does Salesforce AI work with other AI models?

Yes. Agentforce uses models from Anthropic, OpenAI and Google, and you can bring your own model through OpenAI, Azure, Vertex AI or Bedrock. Requests still pass through the Einstein Trust Layer. Salesforce's own Koa reasoning model is in pilot.