A Salesforce org tends to grow the same way a business does. Field by field, over years, until the list outgrows anyone’s ability to describe how it came to be. The descriptions rarely keep pace.
The custom field habit that works against you
Custom fields build up in a CRM because businesses keep adding them, one requirement at a time. A sales rep may have needed to track a special discount on one deal, a field got added, and it stayed long after that deal closed. Multiply that across years of use, and few people left in the business can explain what half the list is for.
Hubbl’s audit of Salesforce orgs found around 10% of custom fields on core objects like Account, Opportunity and Case sit empty, a figure that climbs as high as 32% for fields installed through packages. We see this show up in a few ways, for example:
- Two fields end up capturing the same detail under different names.
- A picklist collects options nobody uses anymore. A description still refers to a campaign that wrapped up two years ago.
For an AI system to use a field properly, it needs a clear, current description of what that field holds. When the description is missing, or was written for a business process that no longer exists, the output built on that field will be wrong, because the field never told it the truth.
Gartner estimates poor data quality costs businesses $12.9 million a year on average, and contributes to 40% of failed business initiatives.
Less is more, and the reason goes deeper than tidiness
The instinct in a CRM build is to add coverage. If there’s a chance the data might be useful one day, capture it. That instinct works against AI readiness. Ten fields with accurate, current descriptions are more useful to an AI system than fifty where half haven’t been reviewed since they were created. A smaller, well-documented data model earns your trust. A sprawling one just looks complete.
A good field description answers three questions clearly:
1. What does the field mean?
2. When should it be populated?
3. Who’s responsible for keeping it current?
Without this information, even a human struggles to work out what the field means, let alone an AI system reading it without local context.
Where statistical models and generative AI both depend on this
Inside a platform like Salesforce, statistical models sit alongside generative AI, doing much of the work of deciding what to recommend to a customer next. Those models draw on structured fields, and their output is only as reliable as the fields feeding them.
Generative AI needs the same discipline, for a different reason. An agent asked to reason over a messy set of fields will still produce an answer, one that sounds plausible because that’s what generative AI is built to do. Research from Technion, Oxford and Hebrew University found that models can sound equally certain whether they’re right or wrong, even when the correct answer was available to them. A review of OpenAI’s research on the topic traces the cause back to training. Standard methods reward confident guessing over admitting uncertainty, so models learn to bluff rather than flag doubt.
Whether that answer is correct depends on whether the fields it drew from were accurate and up to date. Clean, minimal fields give both kinds of AI solid ground to work from. A sprawling, half-documented field list just gives them more room to be wrong with total confidence.
The same fields also feed reports, dashboards and forecasts, so getting this right pays off well beyond any single AI project.
The audit worth doing first
Before building an agent, go through the existing custom fields in your Salesforce instance and ask two questions of each one:
1. Does it still earn its place?
2. Is its description current?
Working out whether a field still earns its place comes down to a few practical checks:
1. Is it consistently populated across records?
2. Does it feed a report, automation or page layout anyone uses?
3. When was it last reviewed?
Checking whether a description is current takes less effort. Read it as a new starter would, and ask whether it still matches what the field holds today. While this is a less exciting task than building the agent itself, it’s also the step that determines whether the agent is worth building at all. Gartner’s own forecast backs this up. By the end of 2026, 60% of AI projects lacking AI-ready data are expected to be cancelled.
The benefits of field discipline
Staying disciplined about custom fields pays off before you build an agent. It means a trustworthy data model, clean reports, and AI working from solid ground. Holding that line is hard when a business wants a new field to solve this week’s problem, but having strong governance will accelerate your AI capability.
If you’d like a second set of eyes on your field list before you build an agent, let’s talk.
