- Load current brand voice, positioning and guidelines into Brand Center, and write specific guardrails with legal and compliance before anyone opens Agentforce Builder
- Brand Center, Real-Time Offer Management and the Piper and Hunter agents are generally available now; the Content Agent and Marketing Goals Agent are still in pilot, so plan production work and experimentation separately
- Deploy one contained use case first, test it against ambiguous prompts and edge cases, then expand in phases rather than switching everything on at once
Why it matters
An agent built on inconsistent brand data produces confident, fluent, off-brand content, and it produces it at a volume no human reviewer can keep pace with. The content might still go out on time and on the intended channels, however it’s unlikely to convert the customer reading it or to strengthen brand affinity.
We’ve outlined five practical steps for setting up Salesforce agents, designed to help you build solutions that save time, support business growth, and meet your customer experience standards.
What’s currently available
It’s worth knowing exactly what you’re working with before you plan around it. As of Salesforce’s Connections 2026 announcements, here’s where things stand.
Generally available now:
- Brand Center: stores brand voice, positioning and guidelines as a data layer every agent references when creating assets
- Piper: An AI SDR agent (from Salesforce’s acquisition of Qualified) that identifies and qualifies website visitors in real time and routes warm prospects to sales
- Hunter: A prospecting agent that identifies new contacts, initiates outreach and runs nurture sequences
- Campaign management in Slack via MCP: Managing segments, campaigns and journeys conversationally, without leaving Slack
- Real-Time Offer Management: Generates and adapts personalised offers and promotions from a prompt as conditions change
Rolling out in beta:
- Agentforce Coworker: An assistant inside Salesforce that answers strategic questions in natural language, pulling campaign and performance data to explain things like a pipeline decline
- Agentforce for Slack (Slackbot): The same assistant capability, available in Slack, including building executive summaries and reports on demand
Still in pilot:
- Agentforce Content Agent: Generates omnichannel content (email, SMS, RCS, mobile) from a campaign description, grounded in brand and customer context
- Marketing Goals Agent: Takes a goal, budget and guardrails, then builds the audience, channel mix and content, adapting as it goes
The pilot agents are the ones getting most of the attention, and they’re also the ones worth testing rather than relying on for anything client-facing yet.
1. Start with brand as data, not a document
Brand Center is where this starts, and it’s now generally available. Instead of a style guide someone opens once a quarter, brand voice, positioning and guidelines live as a data layer that every agent references on every asset it creates.
This matters because it changes what “on-brand” means in practice. An agent isn’t checking your work against a PDF. It’s pulling structured rules each time it drafts an email, an SMS or a promotion. If those rules are thin, inconsistent or out of date, every agent built on top of them inherits the problem. Before any agent goes live, the brand data it draws from needs the same rigour you’d put into a CRM data model.
2. Write guardrails before you write prompts
Salesforce calls these guardrails, natural-language instructions that define what an agent can and can’t do, and when it hands a conversation or a decision to a person. It’s a business decision that needs to happen before anyone opens Agentforce Builder.
Salesforce’s own security guidance is direct about this. Guardrails should cover data access, data governance, and data minimisation, giving the agent only what it needs rather than everything available. This is a conversation with legal, compliance, and whoever owns customer data, held early enough to shape the build.
Agent actions aren’t free to run either. Salesforce prices Agentforce on a consumption basis, so a guardrail that lets an agent loop through edge cases indefinitely is also a cost decision. Usage thresholds and monitoring belong alongside the behavioural guardrails, set before launch rather than added after a surprise bill.
The practical output covers what topics the agent handles, what it’s not allowed to touch, and the exact conditions under which it escalates to a human. Specific guardrails produce agents you can trust with volume.
3. Know what’s live and what’s still a preview
Salesforce’s Summer ’26 release drew a clear line worth paying attention to. Brand Center and Real-Time Offer Management are generally available. The Marketing Goals Agent and the Content Agent, the two that generate full campaigns and omnichannel content from a prompt, are still piloting.
We recommend planning in two tracks. Build the foundation now on what’s GA: unify your brand data, set your guardrails, get your segmentation and personalisation working well. Treat the piloting agents as things to test in a sandbox, learn from and prepare for, not things to plan a campaign calendar around yet.
4. Start small, test hard, scale deliberately
The consistent advice across Salesforce’s implementation partner ecosystem is to resist rolling out agents everywhere at once. Pick one focused use case. Something contained, measurable and low risk if it doesn’t go perfectly the first time.
Before that use case goes live, test it properly. Run it against ambiguous prompts and edge cases, not just the clean examples you expect. Review how the agent reasons through a request, not only what it outputs. This is where most of the risk gets caught, and it’s cheaper to find a gap in a sandbox than in a customer’s inbox.
Once that first agent is working reliably, expand. A phased rollout, each stage tested and measured before the next begins, is slower to start and considerably faster to trust.
5. Keep a human in the loop on the moments that matter
Some decisions shouldn’t run on autonomy, no matter how well an agent performs elsewhere. Pricing, regulated communications, and anything brand-critical belong on a named list of moments for a person to review before anything goes out. Naming these explicitly in your guardrails, rather than assuming good judgment will catch them, is what makes an agent program defensible when something needs to be explained after the fact.
Where to start
A first Agentforce Marketing deployment holds together when the sequence is right:
Get your brand data into the Brand Center and keep it current. Write specific guardrails with the people who own compliance and customer data, before any building starts. Confirm that what you’re deploying is GA, not in pilot, and plan for the difference. Choose one contained use case, test it against edge cases, and only then let it run. Name the moments that always need a human, and check them.
None of this is about slowing down. It’s about the order that operations happen in, so the speed Agentforce Marketing promises is the speed you can stand behind.
If you’re mapping out where Agentforce Marketing fits your team’s next quarter, we’d be glad to talk through it.
Timings correct at time of publication.
