Probably not yet, and that's normal. Most Salesforce orgs weren't built with an AI agent in mind, and the gaps don't show until one starts reading your data.
An agent doesn't fail loudly. It reads half-empty fields, a stale Knowledge base or two versions of the same customer, and gives a wrong answer with total confidence. It also sees whatever its user is allowed to see, which is often more than anyone intended.
The good news is that both problems can be found before you buy anything. Below are the checks we run first, and a short checklist to tell whether you're ready.
Why Agentforce pilots stall after the demo
The demo runs on clean sample data. Your org doesn't.
The pattern shows up in the market numbers too. In July 2026, KeyBanc analysts reported that many Salesforce customers' data is not in good enough shape for meaningful AI work, and that partners were only starting to turn proofs of concept into real deals. A month later, a TD Cowen partner survey found partners weren't yet seeing meaningful revenue from Agentforce projects.
That matches what we see on scoping calls. Teams that stall are rarely stuck on the agent. They're discovering, mid-project, that the data can't support what they bought. Fixing it after you've signed costs more than fixing it before, every time.
The data checks
An agent reasons over whatever is already in your org. These four checks tell you whether that's good enough.
Duplicate rate on Accounts and Contacts. If one customer is three records, the agent picks one and answers from it. Run a duplicate report on your core objects and look at the percentage, not just the count.
Required fields that are actually filled. Fields marked required on the page layout but blank on older records are common. Check fill rates on the fields your first use case depends on, like Industry, Product or Case Reason.
Knowledge freshness. A service agent is only as accurate as the articles it reads. Find the date your Knowledge base last had a real update, and count articles nobody has touched in a year.
Data that lives outside Salesforce. Order history, inventory, billing and usage often sit in an ERP or data warehouse. If the agent needs them, plan how it reaches them: an integration into Salesforce, or a zero-copy connection through Data 360 that leaves the data where it is.
None of these needs a consultant to answer. If you can't get a number on any of them in an afternoon, that is the answer.
The access check nobody runs
The fifth check isn't about data quality. It's about who, or what, can see the data.
Salesforce's own AI plumbing handles where the model runs. Inside Agentforce, the Einstein Trust Layer covers masking, audit logging and zero data retention with the model provider. When Claude reaches Salesforce from outside through Salesforce's hosted MCP servers, it runs as the signed-in user, with object permissions, field-level security and sharing rules all applied.
What doesn't come in the box is a tidy permission model. The AI inherits yours exactly as it is. A customer-facing agent runs as a dedicated agent user with whatever access someone granted it. An over-permissioned profile or a sharing setting left wide open during setup becomes something the AI will faithfully use.
So before anyone turns on an agent or connects Claude, check three things:
What your most permissive profile and permission sets can see, and who holds them
What the agent user or integration user has access to, object by object
Org-wide defaults on sensitive objects, especially anything exposed to a portal or Experience Cloud site
The trust boundary answers "is it safe where it runs?" Your org answers "what can it see?"
Ready or not: a quick checklist
Use this before you scope an agent or sign for one. Most orgs land on "fix first" for at least one row, and that row is where to start.
Duplicates. Ready if one customer is one record and you know your duplicate rate. Fix first if Sales and Service each keep their own version of an account.
Required fields. Ready if the fields your use case needs are filled on most records. Fix first if key fields are blank on older records or filled with placeholders.
Knowledge. Ready if articles are reviewed on a schedule and have owners. Fix first if nobody has updated the Knowledge base in a year.
Outside data. Ready if the agent's data is in Salesforce or connected to it. Fix first if order, billing or inventory data only lives in another system.
Access. Ready if you know what the agent user and your broadest profiles can see. Fix first if sharing defaults were set during setup and never revisited.
Ownership. Ready if someone is named to review the agent's conversations after launch. Fix first if the plan ends at go-live.
FAQ
Do we need Data 360 or Snowflake to use Agentforce?
Not always. Many first use cases run on core Salesforce data. If the agent needs data from your ERP or warehouse, connect to it or bring in only what the use case needs.
Is it safe to connect Claude to our Salesforce?
It can be, once your access settings are right. The connection respects your Salesforce permissions, so it exposes whatever those permissions already allow.
How long does a readiness check take?
You can answer the checklist above yourself in an afternoon. A full review of data quality, access and integrations usually takes two to three weeks.
Should we fix the data before buying Agentforce?
Fix what your first use case depends on. You don't need a clean org everywhere, just clean data under the one process you're starting with.
Find out before you pay for an agent
Our AI Readiness Audit runs these checks across your data, access and integrations in two to three weeks, with read-only access, and gives you a prioritized fix list. You can take the list in-house or have us do the work.
