Solutions · Ready for agents
Agentforce data readiness: find out what the data would get wrong first
An AI agent acts on your data as it is, quickly, and it won’t notice what a person would. Before it starts, measure what it would meet. After it starts, keep a record of what it changed.
The job, and why it hurts
The agent pilot has a start date, and the design review turns up three questions nobody can answer. Which of the two accounts for the same company will it read? Who owns the contacts it’s meant to email? What happens to a field it overwrites?
A person works around bad data by shrugging and opening both records. An agent reads the duplicate, the blank field or the stale record and acts on it as confidently as on a clean one. So the pilot stays read-only, because nobody can say what a write would do or how to take it back.
That makes two jobs. Before the agent starts, find out what it would get wrong. After it starts, know what it changed and be able to put it back.
The apps for Agentforce data readiness
Where things stand. Sounding is pre-release: built and in pilot. Tiller is in design and may never be built. Sluice and Wake are in early access, and Watchbill is pre-release.
Sounding: a score, a fix list and an answer on agent readiness (pre-release)
Sounding, by GTM Cloud, scans five standard objects plus up to three of your own, with ten checks covering duplicates, completeness, staleness, orphans, ownership and agent readiness. You get a score from 0 to 100 and a fix list in priority order, and every figure opens the records behind it. It only reads. See what Sounding measures.
Tiller: a ledger of what an agent changed (in design)
Tiller, by GTM Cloud, is designed to keep one entry per agent action that goes through GTM Cloud’s apps, with before and after values, an undo, and approval for the actions you list. Whether we build it depends on what the platform already records. It won’t catch changes made another way. Read the Tiller design.
Sluice: one record per customer
Sluice, by GTM Cloud, resolves the duplicates Sounding finds. Review each group side by side, merge with a backup taken first, and undo from History if the call was wrong. See Sluice’s merge review.
Wake: a history of changes
Wake, by GTM Cloud, logs old value, new value, who and when for the fields you choose, beyond the native 20 per object. A dedicated agent view for changes an agent’s user makes to those fields is planned, not built. See Wake’s field history.
Watchbill: what the agent’s user is allowed to do (pre-release)
Watchbill, by GTM Cloud, is an access review tool. Open the user an agent runs as and see every permission and what grants it. Snapshots and review campaigns are in the app. See how Watchbill shows access.
How they hand off to each other
- Sounding measures duplicates, undocumented fields, untracked changes and thin descriptions.
- Sluice resolves the duplicates it finds.
- Wake starts a history on fields Sounding flags as untracked.
- Watchbill shows what the agent’s user can read and change.
- Tiller, if it’s built, would record what the agent did.
Run Sounding again after each fix to see whether it held. Every app installs alone, and none needs another.
See it first. Take it back. Keep the receipt.
For agents, that means three habits: preview, keep the old values, and say afterward who did what, as in the promise behind every app. Purser’s two Agentforce actions dry-run first (they’re off until you turn them on), and Stow’s agent actions always preview and ask before filing.
When a person should help
The Agentforce Readiness Sprint takes two to four weeks. We look at your data and permissions through an agent’s eyes, pick two or three use cases worth doing, and design the guardrails before anything goes live. Want the number first? Start with a Data Health Check.
Questions people ask
Is my Salesforce data ready for an AI agent?
Check four things: one record per customer, fields that say what they hold, a history of changes, and enough text on a record to ground an answer. A scan that links every figure to its records shows what the agent would get wrong.
How do I prepare Salesforce data for AI agents?
Measure first, then fix in order. Resolve duplicates, document fields people misread, start history on fields an agent might change, and review the access of the user it runs as. Then measure again, to see whether the clean-up held and whether new records still arrive incomplete.
What access should an AI agent have?
An agent runs as a user, so treat that user like any other: review what it can read and change, and give it no more than the task needs. Watchbill’s explorer shows each permission and what grants it.
How do I audit what an AI agent changed in Salesforce?
The platform’s tracing explains what an agent did and why, but not what each record looked like before. A history on the fields in scope fills that gap. Tiller is designed to add a ledger with before and after values and an undo, and it isn’t built.
Should an agent pilot start read-only?
It’s a sensible default while nobody can say what a write would do or reverse it. Move to writes one action at a time: preview first, approval where it matters, a ceiling on records per call, and a way to put things back.
Find out before the pilot does
Book a Data Health Check and learn what your data would get wrong before an agent does.