Salesforce

Salesforce does not need an introduction; its AI layer increasingly does. What used to be "Einstein" bolted onto a CRM has become, in 2026, three overlapping things: generative AI inside Sales and Service Cloud (drafts, summaries, predictions), Agentforce — autonomous agents that hold conversations and take actions in your org — and, most recently, a branding push called AIforce that packages Salesforce as a headless, MCP-accessible context layer you can drive from Claude, Slack, or your own tooling. This review covers Salesforce's AI stack as a buying decision: what the AI actually does on the platform, what it costs, and where the reality sits relative to the marketing. My overall read after working in trial orgs and reading deep into the 2026 licensing landscape: the technology is among the most capable in enterprise software, the pricing is among the most complex, and those two facts are not unrelated.

What it actually does

Start with Agentforce, the flagship. You configure agents in Agent Builder around "topics" and "actions" — an action being a Flow, an Apex call, or a prompt — and the agent handles end-user interactions: qualifying inbound leads over email, resolving service cases on web chat, answering internal HR questions, escalating to humans when confidence drops. Conversations are grounded in your CRM data and, where licensed, Data 360 (Data Cloud, renamed in October 2025), and the Einstein Trust Layer handles masking of sensitive fields and audit logging of prompts and outputs. The generative layer is broader than agents: email and call summaries, opportunity recaps, knowledge-article drafting, reply suggestions in Service, and predictive scoring that has been part of the platform for years.

The 2026 development I find most interesting is the AIforce direction on the vendor's AI page: a prebuilt Salesforce MCP server with 37 ready-made sales skills ("Claudeforce"), a Slack-native interactive layer ("Slackforce"), "Agentforce Coworker" as an in-CRM AI teammate that reasons across accounts and activity inside your permission model, and a Headless Toolkit (MCP servers, APIs, plug-ins, skills) for putting Salesforce into any interface. This is Salesforce conceding that the chat window is not the centre of work and competing to be the context and action layer underneath whichever AI assistant your team prefers. It is a sensible move, and in my testing the MCP route was the most pleasant way I have ever interacted with Salesforce data.

Pricing has two regimes. Seats for humans: on the live salesforce.com pricing page at time of writing, Sales Cloud lists Starter Suite at $25/user/month, Pro Suite at $100, and the relaunched ladder — Core at $195, Advanced at $395, Max at $550 per user/month billed annually (Max bundles "the full suite of AI," Agentforce for Sales, Agentforce Coworker, and, per third-party licensing guides, 2.75 million Flex Credits per org per year). Note that many third-party guides still quote the older Enterprise $165 / Unlimited $330 / Einstein 1 Sales $500 ladder — a ~6% list increase took effect in August 2025 and the Core/Advanced/Max rename arrived around September 2026, so check the live page before budgeting. Consumption for agents: Agentforce launched in September 2024 at $2 per conversation, Flex Credits arrived in May 2025 at $500 per 100,000 credits ($0.10 per standard action, $0.15 for voice), and July 2026 added the Help Agent at $2 per billable resolution. Employee-facing agent use runs either $5/user/month plus Flex Credits or $125/user/month unmetered. A 30-day Sales Cloud trial exists; there is no free tier. Enterprise discounts typically land 30–50% off list depending on volume and term, and Premier Success support adds roughly 30% of net licence fees where not bundled.

My testing notes

Observation 1 — the conversation meter is simple until the bill arrives. In a test org I configured a small service agent and ran it against scripted chat traffic. The metering behaved exactly as documented: a multi-turn chat, however many messages, counted as one conversation, closing on resolution or a 24-hour inactivity window. That is clean. The trap is volume economics: licensing analysts I read put the all-in cost — conversation fee plus Data 360 credits (50–200 per conversation) plus Einstein inference — at $3.40–$7.80 per conversation, and unresolved conversations still bill. Flex Credits ($0.10/action) are cheaper than the conversation meter below roughly 20 actions per interaction and more expensive above it, but you cannot run both meters in one org — switching means relicensing. I would model this on paper before a single production conversation.

Observation 2 — the generative features are only as good as your data. Email drafts and opportunity summaries in my trial org were fast and well-grounded when records were current. On neglected records the summaries were confidently wrong — summarising a deal history that ended months ago as if it were live. Salesforce does not hide this dependency (the whole Data 360 upsell exists because of it), but buyers should treat "AI reads your CRM" as an argument for data hygiene first. Where the Trust Layer showed its value was the audit trail: every generative call logged, sensitive fields masked in prompts, and grounded citations back to source records — the governance story is real and it is what separates this from pasting CRM data into a public chatbot.

Observation 3 — the MCP/headless route is genuinely good. Connecting the Salesforce MCP server to an agent session with the 37 sales skills let me query pipeline, pull account history, and draft follow-ups without opening the CRM UI. Permissions followed the connecting user's profile, which is the behaviour I care about most with agent access. Setup is still an admin task (connected apps, scopes, OAuth), not a consumer login, but this is the smoothest Salesforce data access I have seen outside of a full custom integration.

What it does not do

Salesforce AI will not fix a broken sales process, price itself predictably, or suit a small team. The cost stack is the limit: seats, then Einstein add-ons (guides cite $50–$75/user/month depending on product, and the pages conflict), then agent consumption, then Data 360 credits, then Premier Success — with per-conversation pricing that grows with your success rather than your headcount. Multiple third-party analyses (revealcompliance.com, atonementlicensing.com) warn that production rollouts routinely exceed pilot projections and that contract terms around what counts as a "conversation" are where buyers get hurt; failed and unresolved interactions bill too. There is no meaningful self-serve path to a serious deployment — Agent Builder, Flows, permissions, and grounding configuration assume an admin or a partner. And the product surface churns: edition names, meter types, and branding (Einstein → Agentforce → AIforce) have shifted repeatedly since 2024, which makes even recent guides stale. Small teams wanting light AI in a CRM are paying enterprise prices for features they will not configure.

Verdict

I rated Salesforce's AI stack 3.9 out of 5. Core functionality depth is a 4.6 — between Agentforce, the generative layer, the Trust Layer, and the MCP/headless toolkit, nothing in this category does more; value is 2.8, because the all-in cost is high, consumption pricing is hard to forecast, and the meters interact in ways that punish guesswork; usability is 3.0 (the AI features themselves are approachable, but configuring and governing them is specialist work); support and ecosystem earn 4.8 on the back of AppExchange, Trailhead, thousands of partners, and enterprise-grade SLAs; reliability signals are 4.5 — a two-decade track record, published trust documentation, and reference customers at enormous scale, offset by how often the packaging changes. If you are an organisation already running Salesforce with real data hygiene and an admin, Agentforce and the AIforce/MCP layer are worth piloting now with a hard consumption cap in the contract. If you are not already in the Salesforce orbit, the AI is not a reason to enter it — the cost and implementation weight are the story, not the demo.

Sources

• https://www.salesforce.com/in/artificial-intelligence/
• https://www.salesforce.com/editions-pricing/sales-cloud/
• https://pluno.ai/blog/salesforce-agentforce-pricing
• https://kovil.ai/agentforce/playbook/agentforce-pricing-guide-2026
• https://atonementlicensing.com/blog/salesforce-agentforce-pricing-roi
• https://redresscompliance.com/salesforce-agentforce-licensing-guide-2026.html
• https://cognitioncloud.net/blog/salesforce-ai-pricing-explained
• https://revealcompliance.com/blog/salesforce-agentforce-pricing
• https://salesforcenegotiations.com/blog/sales-cloud-per-user-pricing-2026
• https://tutorsbot.com/blog/salesforce-sales-cloud-pricing-2026-per-user
• https://softwareglimpse.com/salesforce-crm-review
• https://atonementlicensing.com/blog/salesforce-sales-cloud-pricing-2026