Most organizations are eager to put AI agents to work. Far fewer have a clear picture of whether their data can support one.
That gap matters more than it sounds. Duplicate records, inconsistent fields, and loose permissions don’t stop a person from getting their work done. People notice the oddities and work around them. An AI agent doesn’t. It acts on whatever it finds, so a weak foundation shows up as wrong answers, missed actions, and pilots that look great in a demo and stall in production.
The good news is that the foundation is knowable. There are ten things that tend to decide whether an organization is ready, and most of them have nothing to do with the AI itself.
The ten checkpoints
The data foundation
| Checkpoint | What good looks like |
|---|---|
| Data quality | Customer records are trusted, and data is monitored and cleaned on a regular schedule |
| System integration | CRM, ERP, Support, and Marketing share data automatically, with no rekeying |
| Data pipeline | An enterprise tool (ETL or similar) moves and connects data, not one-off scripts |
| Data platform | A data lake, warehouse, or similar cloud solution is the governed home for your data |
The first two – data quality and system integration – are the ones most people think of. The second two get overlooked more often. A pipeline and a central, governed platform are what keep the first two from regressing. Without them, clean data tends to stay clean only for a brief moment in time.
Rules and guardrails
| Checkpoint | What good looks like |
|---|---|
| Data governance | Clear rules for who can see and use data, with defined ownership and regular audits |
| AI usage policy | A defined policy on how AI can be used across the organization |
An agent inherits the access of the user or process it runs under. If permissions have sprawled over the years, the agent inherits that sprawl too. A written AI usage policy is the other half. It settles what’s allowed before someone has to ask in the middle of a project.
People and plans
| Checkpoint | What good looks like |
|---|---|
| Team readiness | Strong data literacy across the org, with ongoing training |
| Use case clarity | A prioritized roadmap tied to specific business outcomes |
| Leadership alignment | Executive sponsorship, with AI tied to business strategy and budget allocated |
Technology that nobody is ready to use, aimed at no particular problem, even if it is funded, doesn’t get far. Of the three, use case clarity is the one that saves the most money. A short list of well-defined use cases beats a long list of ideas every time.
Proof it works
| Checkpoint | What good looks like |
|---|---|
| Agent deployment | An AI agent is already live for external use, such as with customers or partners |
This one is different from the rest. It isn’t a prerequisite, and plenty of ready organizations haven’t done it yet. But it is a good signal that the rest of the list is properly managed and will hold up under pressure.
Where AI implementations go wrong
Most agentic AI pitfalls come back to hallucination stemming from missing context, messy underlying data architecture, or both.
A common issue stems from overestimating an agent’s out-of-the-box capabilities. It’s given a task and assumed that it can decipher poorly named fields or implicit object relationships without training or proper inputs.
In practice, an agent should be treated like a new hire. If calculating a metric like website lead conversion rate relies on ambiguous field names or missing definitions, a new team member would struggle; and an agent will too, resulting in confident answers grounded only in assumptions.
When auditing data readiness, two underlying gaps behind those faulty guesses appear repeatedly:
1. Uncaptured History: Leaders want AI to explain trends and changes, but fields rarely track historical values unless history tracking or snapshotting is intentionally enabled. Certain objects, like campaign members, don’t support history tracking at all. The only solution is to establish logging early and allow data to accumulate over time.
2. Conflicting Definitions: If key stakeholders disagree on metrics like “revenue” versus “total sales,” an AI model won’t know either. It will generate responses anyway, and once users spot inaccuracies, trust in the agent quickly vanishes.
Where to go from here
You don’t need to answer all ten perfectly; very few organizations can. What matters is knowing which of them is holding you back, so the next dollar goes toward fixing the right thing.
If you want a quick read on that, we turned these ten checkpoints into a short self-assessment. It takes about three minutes, gives you a score from 1 to 5 and a stage, and points out the one area to look at first. If your results raise questions, there’s an option at the end to talk them through with someone who works with Salesforce, business intelligence, and Agentforce every day.
