Most Agentforce pilots stall because nobody wired the actions, the data, or the measurement. We do those bits.
The licences are the easy bit. Getting it to actually do work in your business is where it usually dies.
Conversations are happening, but nobody knows whether any of them resulted in a closed case, a booked call, or a saved hour. No measurement, no business case for renewal.
It cheerfully describes how to log a case, but doesn’t log it. That’s a chatbot, not an agent. Agentforce only earns its name when the actions are wired in.
Duplicate accounts, half-filled fields, inconsistent picklist values — the agent reflects all of it back. Garbage in, fluent-sounding garbage out.
The consultancy left, no one internal owns the prompts or topics, and the next iteration is “parked pending capacity”. Six months later it’s shelfware.
Five components. Skip any one of them and the pilot stalls.
Topics tell the agent what kinds of request it handles. Actions are the things it can actually do — invoke a Flow, update a record, call an API. This is where most builds underinvest.
Salesforce’s reasoning engine that decides which topic to use and which action to invoke. Configured properly it’s reliable; left at defaults it picks the wrong thing.
Data masking, prompt defence, audit logging. The bit that lets you put this in front of customers without your legal team having a quiet breakdown.
Slack, the website, the mobile app, an embedded experience inside Salesforce. Each channel needs its own configuration and its own measurement plan.
Conversation volume, action success rate, deflection, escalation, customer satisfaction. Without these you can’t answer “is this working?” — and renewal becomes a fight.
We’d rather tell you up front than three weeks into a pilot.
Fixed scope, fixed price, working software at the end. No discovery-then-implementation two-stage dance.
Pick the one or two topics worth shipping first. Audit the data those topics depend on. Identify the actions to wire.
Configure the topics, build the actions (Flows, Apex, external callouts), wire the channels.
Guardrails, masking, audit. Set up the conversation, action and outcome dashboards so you can answer “is it working?”
Live pilot with a real audience, iterate on what surfaces, hand the keys to whoever owns it internally. Documented, not mystical.
No, not for v1. You can ship a useful Agentforce build against Salesforce data alone. Data Cloud only becomes necessary when you want grounded retrieval across multiple systems — emails, support tickets, ERP records, knowledge bases. Start without it; add it when there's a specific reason.
Licences are consumption-priced per conversation, on top of your existing Salesforce edition. Implementation cost depends on how many topics and actions you need wired up, and how clean the underlying data is. We cover the real numbers in our Fox Den write-up: see /news/agentforce-cost-uk-licences-implementation.
Four weeks for a focused build — one or two topics, the actions wired in, Trust Layer configured, measurement in place. Anything claiming a fortnight is either trivial or skipping the bits that actually matter (data, guardrails, who owns it after we leave).
No. Einstein generates text. Agentforce reasons about a request, picks an action, executes it, then reports back. The point is that it does the job — updates the record, raises the case, books the meeting — rather than just describing how someone else could do it.
Yes. We build in a sandbox, validate the actions against your real data shape, configure the Trust Layer to mask anything sensitive, and only promote when you are happy. Rollback is a deploy, not a rebuild.
The Trust Layer plus grounded retrieval keeps the agent answering from your actual data rather than guessing. Combined with explicit action definitions and a measurement loop, hallucination risk drops to where it's manageable for production. We won't pretend it's zero.
If you also need Einstein or Data Cloud, see Salesforce AI. Just need a quick check? Free 30-min Health Check →
Book a free 30-min Health Check. We’ll look at your Salesforce, the work you’d want an agent to do, and tell you honestly whether Agentforce is the right answer.
Book a Free 30-min Health CheckNo card required. No pitch. Written summary after.