Tiller, by GTM Cloud · Ready for agents · In design
A hand on the tiller when an agent takes the helm
Tiller is designed to help you audit what an AI agent changed in Salesforce: a ledger of its actions through GTM Cloud’s apps, with values before and after, an undo, and approval before the actions you choose.
Status: In design, and possibly never built. Whether we build Tiller depends on a check of what the platform already records. We’re doing that check first.
You know this one
Two weeks into the agent pilot, a sales director stops you in a corridor. “What exactly did it change?”
Which opportunities? What did the Amount say before? Did a person approve any of it?
You open the agent’s trace. It explains what the agent did and why, in careful detail. It doesn’t show what each record looked like before. So you say you’ll get back to them. The pilot stays read-only, because nobody can answer the question, and you’d rather not guess.
A way to audit what an AI agent changed in Salesforce
Tiller is designed to keep one ledger entry for every action an agent takes through GTM Cloud’s apps: which action, which records, preview or commit. For each changed field it keeps the old value and the new one.
An entry can be undone from the ledger, and the actions you list wait for a person’s approval before they commit.
The tiller is the lever you steer by. This is meant to keep a hand on it.
See it first. Take it back. Keep the receipt.
| Preview | Undo | Prove |
|---|---|---|
| Agent actions are designed to dry-run first: the agent shows what would change, then commits. | Undo an entry from the ledger; the app that made the change reverses it. | Every action, every record, and the value before and after. |
What it’s designed to do
Every action, on the record
Each agent action will be one ledger entry. Filter by agent, action, app and day.
Before and after
For every changed field, the ledger will keep the old value and the new one.
Undo from the ledger
Choose an entry and undo it. The app that made the change reverses it, and refuses if the record has moved on since.
Approval and ceilings
List the actions that must wait for a person. Set the most records one call may change: an action over the ceiling is refused, and the refusal is logged.
How it will work
- Connect the apps. Name Tiller’s ledger class in each GTM Cloud app’s settings.
- Set the rules. Choose which actions wait for approval and the record ceiling for each.
- Let the agent work. Previews and commits appear in the ledger as they happen.
- Review. Read the day’s entries, approve what is waiting, and undo what shouldn’t have happened.
Two readers, two reasons
If you’re the admin
You’ll answer “what did the agent change?” from one tab. Each GTM Cloud app works exactly as before until a ledger is connected.
If you’re the one signing off
An agent is valuable because it acts, and that’s also the risk. Tiller is designed to let a pilot move from read-only to real work, with a record of every change.
Today, and with Tiller
| Today | With Tiller |
|---|---|
| “The agent did it” with no detail | A ledger entry for every action |
| Old values gone the moment they change | Before and after for every field |
| Every action allowed, or none | Approval where you want it, speed where you don’t |
| A pilot paused over one bad night | An undo for the bad night |
Where it’s designed to be strong, and what to know first
We’d rather you find the limits here than after you’ve planned around it.
Where Tiller is designed to be strong
- One ledger entry per agent action, with the records it touched
- The value before and after for every changed field
- Undo from the ledger
- Approval before commit for the actions you list, and a record ceiling per action
- Inside your org, with no dependency between apps
Good to know first
- The ledger records actions that go through GTM Cloud apps. A change an agent makes another way isn’t in it.
- It doesn’t show the agent’s reasoning or conversation. The platform’s own tracing does that.
- Undo is done by the app that made the change, and is refused if the record has changed since.
- It’s a design. Whether it’s built depends on a check of what the platform already records.
Questions people ask
Does it capture everything an agent does?
No. It records actions that go through GTM Cloud apps. For field changes an agent’s user makes by any route, the planned agent view in Wake covers the fields Wake tracks.
Does it slow the agent down?
Only where you ask it to. Actions on your approval list wait for a person, and the rest are recorded and carry on.
How is this different from the platform’s agent tracing?
Tracing explains what the agent did and why. Tiller is designed to keep what the records looked like before and after, and to offer to put them back.
Do our apps need Tiller to use agent actions?
No. Each GTM Cloud app’s actions work without it. Name Tiller’s ledger class in an app’s settings and entries start arriving. Until you do, the app works exactly as before.
How do I audit what an AI agent changed in Salesforce?
For changes made through GTM Cloud’s actions, Tiller is designed to give you one entry per action, with the records touched and the before and after values. Changes an agent makes by other routes aren’t in it.
Works well with
- Wake: a planned agent view will cover field changes an agent’s user makes by any route, for the fields Wake tracks.
- Watchbill: the agent runs as a user, so open that user and see what it can read and change.
Tiller sits in the agent readiness stage of the lead-to-renewal map. Planning a pilot? Our Agentforce Readiness Sprint is where to start. Every app installs by itself. None needs another.
Built on the platform
- Native to Salesforce. Entries, changes and approvals will be records in your org.
- It complements the platform’s own agent tracing and its confirmation and hand-off options. It doesn’t replace them.
- Three permission sets are planned: Admin, Approver and Auditor.
Planning to let an agent write?
Tiller is in design, and we’re first checking what the platform already records. Tell us what you’d need to see before you let an agent change a record.