An agent is only as good as the org underneath it.
Agentforce is genuinely useful when it's grounded in clean data, given tightly-scoped actions, and fenced in by real permissions. It's a liability when it's switched on over a messy org and pointed at customers. Most of the work is the part nobody demos — and that's the part we do.
What Agentforce actually is, without the pitch
Agentforce is Salesforce's platform for building AI agents that live inside your org — they can read your records, take actions through Flow and Apex, and hold a conversation with a customer or an employee. Unlike a bolt-on chatbot, it runs under Salesforce's own permission and sharing model, which is exactly why it's worth doing and exactly why it has to be done carefully.
The demos make it look like a toggle. In practice it's a build: deciding what the agent is allowed to handle, writing the actions it can take, connecting it to data it can trust, locking down what it must never do, and testing it against real cases before a customer ever sees it. Skip any of those and you get a chatbot the company quietly switches off three months later.
What actually gets built
"Deploy an agent" sounds like one task. It's six, and every one of them is where projects go wrong.
Topics — what it's allowed to handle
Each topic is one job the agent can do, with instructions written tightly enough that it doesn't wander into territory nobody planned for. Scoping topics narrowly is the single biggest lever on whether the thing behaves in production.
Actions — what it can actually do
Standard actions, Flows, and Apex where the logic genuinely needs code. Every action is a door you're opening into your org, so each one gets defined inputs, defined outputs, error handling, and a reason to exist.
Grounding — where its answers come from
Records, knowledge articles, and whatever else it needs so it answers from your reality instead of improvising. This is the stage where data quality stops being an abstract complaint and becomes the whole project.
Guardrails — what it must never do
Permissions, sharing rules, field-level security, and an explicit escalation path to a human. The agent runs as a user; that user gets scoped as carefully as any employee you'd hire.
Testing before anyone real touches it
Run it against actual cases and actual edge cases, watch where it fails, tighten the instructions, repeat. Then a limited rollout to a small internal group before it ever faces a customer.
Documentation your admin can own
Every topic, action and guardrail written down, plus a walkthrough for whoever maintains it after we're done. Built so the next person can change it without calling us — that's the point.
When it's worth it, and when it isn't
We'd rather lose the project than build you something you'll switch off in a quarter.
Worth building
- High volume of repetitive, answerable questions
- The answers already live in Salesforce or your knowledge base
- Your data is reasonably clean — or you're willing to clean it
- There's a clear escalation path to a human
- Someone internally will own it after handover
We'll tell you not to
- Duplicate records and stale data nobody trusts
- The real problem is a broken process, not a slow one
- Low volume — a Flow or a better page layout is cheaper
- Nobody internally has time to own it
- The goal is "we should have AI" with no use case behind it
How Agentforce work is packaged
Every engagement is scoped on a free call and quoted at a fixed price before anything is built. No hourly meter.
Agentforce Readiness
The honest starting point. We review your org, data, permissions and candidate use cases, then hand you a ranked shortlist of what's worth automating, what has to be fixed first, and a written build plan you own — whether or not we build it.
Agentforce Deployment
The full build: topics, actions, grounding, guardrails and testing, done in your sandbox and deployed to production with documentation and a walkthrough for your admin. Credit applied if you started with a Readiness review.
Agent Care
Agents drift as your org changes. Ongoing tuning, new topics as you find them, monitoring of what it's getting wrong, and a person who already knows how it was built. Cancel any time — nothing is held hostage.
An agent is a user. We scope it like one.
Agentforce runs under Salesforce's own permission model, which means the agent sees exactly what its profile, permission sets, sharing rules and field-level security allow — and nothing else. That's only a safeguard if someone deliberately scopes it. Every build gets least-privilege permissions, sandbox validation before anything touches production, and a human escalation path that always exists.
Agentforce questions, answered.
Do we need Data Cloud to use Agentforce?
Not always. Plenty of useful agents ground entirely on standard Salesforce records and Knowledge articles. Data Cloud matters when you need to unify data that lives outside Salesforce or do large-scale retrieval across it. Part of a Readiness review is telling you honestly which side of that line you're on before you buy anything.
How long does an Agentforce deployment take?
It depends almost entirely on the state of your data and how many topics you want live. A single tightly-scoped topic grounded on clean records is a matter of weeks, not months. A broad agent across a messy org takes longer — and the extra time is data cleanup, not agent building. Scope and price are locked before work starts.
Is it safe to point an AI agent at our customer data?
It can be, if it's built properly. The agent runs as a Salesforce user, so it sees exactly what that user's profile, permission sets, sharing rules and field-level security allow — no more. We scope that user deliberately, test what the agent can reach, and build an escalation path to a human for anything outside its lane. See the Approval Gate section above.
What if our Salesforce data is a mess?
Then that's the project, and we'll say so. An agent grounded on duplicate and stale records will confidently give wrong answers, which is worse than no agent at all. A Readiness review tells you exactly what has to be cleaned and roughly what it costs, so you can decide before committing to a build.
Can you work alongside our existing admin or partner?
Yes, and it usually goes better that way. Your admin knows the org and the business; we bring the development and agent work. We're also available to consultancies and partners as white-labeled overflow capacity, working inside your process and under your NDA.
Are you a certified Salesforce partner?
No. Unlock Integrations is an independent studio, not a Registered Salesforce Partner, and doesn't lead with badges. What you get is a senior developer with 15+ years in software and deep hands-on Salesforce experience doing the work directly, start to finish.
Tell us what you'd want an agent to handle.
Book a free scoping call. Describe the job you'd hand to an agent and we'll tell you honestly whether it's a good fit, whether your org is ready, and what it would take — then send a fixed-price quote. No obligation, and you talk to the senior developer who'd actually build it.
Book a free scoping call →