Agentforce Implementation: How to Plan, Build and Launch Your First AI Agent
Planning an Agentforce implementation? What to sort out first, how…
An Agentforce implementation is the work of setting up Salesforce’s AI agent platform in your org and launching agents that do real tasks. In practice that means choosing one narrow use case, getting Data 360 and Einstein switched on, building the agent’s subagents and actions, testing it hard, then releasing it to a small group before everyone else.
Key takeaways
Most teams we speak to are curious about Agentforce, but a bit wary. They’ve seen the keynote demos. What they haven’t seen is a plain account of what it takes to get one agent working properly inside their own messy, real-world org. So here’s ours.
Agentforce is Salesforce’s platform for AI agents that can take action, not just answer questions. An agent can look up an order, update a case, draft a reply or hand over to a person, all inside the rules you set. Salesforce launched it in October 2024 and has shipped several big releases since, most recently Agentforce 360, announced at Dreamforce in October 2025 (Salesforce).
If you read our earlier piece on Einstein Copilot, that product is now part of Agentforce. Salesforce renamed it with no change in functionality (Salesforce Help).
Under the bonnet, the Atlas Reasoning Engine reads a request, works out what the person wants, picks an action, checks the result and loops until the job’s done (Salesforce). You control it with three levers:
Everything passes through the Einstein Trust Layer, which handles grounding in your data, zero data retention with the model provider and toxicity checks (Salesforce).
This is where most of the early pain lives. Salesforce’s own readiness guide lists the prerequisites: Lightning Experience, Data 360 provisioned, Einstein Generative AI enabled, the Trust Layer available and Agentforce turned on, in that order. It also warns that provisioning can take 15 to 60 minutes, so a setup that seems broken may just be early (Salesforce Help).
Data 360 is the one that surprises people. It’s required because the Trust Layer depends on it (Salesforce Developers). If you’re not sure what it does, our guide to what Data 360 does for your customer data covers the basics.
Then there’s the stuff no setup screen will tell you. An agent is only as good as the records and permissions behind it. Duplicate contacts, half-filled case fields and a sharing model nobody fully understands will all show up in its answers. That’s why we’d run a Salesforce health check before any serious AI work, and fix the worst of it first.
Choose something frequent, well-documented and low-risk. Order status questions, password resets, appointment changes. Write down what success looks like (cases deflected, handling time, satisfaction) and when the agent must hand over to a human.
Provision Data 360, enable Einstein, check the Trust Layer settings and sort permission sets. Do this in a sandbox first. It sounds obvious. It gets skipped.
Your agent can only do what its actions allow. Most of these will be Flows or Apex your team already understands, and many will need data from other systems, so this is where Salesforce integration work often comes in. Get the actions reliable on their own before wiring them into the agent.
Agentforce Builder, generally available since February 2026, lets you write instructions in plain language and add deterministic logic with Agent Script where you need certainty, such as ‘always verify identity before discussing an order’ (Salesforce Help). Use that logic for anything with compliance or money attached.
Build a question set from real cases, including the awkward ones: rude customers, vague requests, people asking for things the agent shouldn’t do. Run every version through Testing Center and read the traces. One catch worth knowing: Salesforce notes that Simulate mode in the builder consumes credits too (Salesforce Help).
Release to one channel or a small group of users. Review sessions every week for the first month, tighten instructions, then expand. Treat it like onboarding a new starter rather than flicking a switch.
Agentforce is mostly usage-based. Salesforce’s UK list prices at the time of writing are £400 per 100,000 Flex Credits, with each standard agent action using 20 credits, or £1.60 per conversation on the conversation-based model (Salesforce Agentforce pricing). Prices and bundles change often, so check the current page and your contract before you budget.
The honest advice: estimate actions per conversation, multiply by expected volume, then add a buffer for testing. Agents that call several actions per request cost more than simple Q&A.
If you’d like a second pair of eyes on your plan, our Design & Deliver team can help you scope a sensible first agent. Get in touch.
A tightly scoped first agent with clean data and existing Flows can be live in weeks. Timelines stretch when the agent needs new integrations, a data clean-up or complex security rules. Plan for a pilot phase with a small group before a full launch.
Yes. Salesforce requires Data 360 to be provisioned before you can set up the Einstein Trust Layer and Agentforce. Enterprise Edition and above customers can provision Data 360 at no extra cost, with an annual allowance of credits.
Einstein Copilot was renamed Agentforce in 2024 with no change in functionality. Agentforce has grown well beyond it since, with customer-facing agents, voice, a new builder and usage-based pricing.
At the time of writing, Salesforce lists Flex Credits at £400 per 100,000 credits, with a standard action using 20 credits, or £1.60 per conversation on the conversation model. Check Salesforce’s pricing page for current rates.
Teams with a strong admin and some Flow or Apex skills can build a simple agent themselves. A partner helps most with readiness, data quality, integrations and testing, which is where projects tend to stall.
