Why AI guardrails make your team faster
When leaders hear the word “guardrails,” they usually think of brakes. Something that slows the work down.
I see it the other way around.
Picture driving a winding mountain road at 120 km/h with no guardrails. You’d ride the brakes the whole way down. Put solid barriers along the edge and you can drive at a normal speed without white knuckles.
AI works the same way. If your staff don’t trust the boundaries, they’ll either check every single thing the AI does, which wipes out the time savings, or they’ll quietly stop using it. A reminder to “be careful with AI” won’t change that, but building the limits into the system will.
A prompt isn’t a guardrail
A lot of AI safety today is a line in the instructions. “Please don’t change payment records without permission.”
That’s a suggestion. Language models work on probability. Run something enough times and the model will eventually slip past a soft instruction. You don’t want your donor database relying on a model having a good day.
Real guardrails live in ordinary code around the AI, where a rule works the same way every time. Here are the three patterns I use in my Salesforce engineering work and my own builds:
- A check before anything touches live data. Every action the AI proposes passes through a layer of plain, rules-based code first. If the action isn’t allowed, it gets blocked before it ever reaches your database.
- A sandbox for anything the AI builds. When an agent writes code or a script, it runs in a separate, throwaway environment. If the logic is wrong, it fails there, away from your real records.
- Approval steps the AI can’t skip. High-risk work, like changing payment settings or sending a mass email, moves through fixed stages. The AI can move a task from draft to ready for review. Only a signed-in person can move it to done.
Ground it in your real systems
There’s one more piece, and I see it go wrong all the time in Salesforce work. Someone asks AI to build a flow or write some code without showing it the actual org. It makes up fields that don’t exist, uses outdated features, or breaks a validation rule nobody mentioned.
The fix is to give the AI your real setup before it starts. That means the current list of objects, fields and rules, examples of how similar work was done before, and a place to test its changes before a person reviews them. That grounding does more for safety than switching to a smarter model.
How to set up guardrails your team will actually use
Ask a fundraiser or program manager what worries them about AI and they’ll usually describe a public mistake on their watch, like a made-up detail in a major donor email or a wrong number in a board report.
Here’s how I’d set things up so people feel safe using it:
- Sort your work by risk. Reading data and drafting are low risk. Changing records, sending anything outside the building and touching money are high risk. Write the list down.
- Start with read-only access. Let the AI read what it needs. Give it permission to change things one workflow at a time, once you trust that workflow.
- Put a review gate near the finish line. Let the AI do most of the heavy lifting: gathering the data, finding the trends, drafting the copy, formatting the report. Then it stops and shows a short summary of the data it pulled, the logic it used and its proposed draft. Your staff member spends three minutes reviewing and clicks approve, instead of three hours doing the data entry.
- Check the numbers automatically. Anything with calculations should pass an automated check before a person even sees it. Your reviewer shouldn’t be the only thing standing between a typo and the board.
- Give it your real templates and examples. Share your field list, a few past grant reports and your style guide. The more of your actual setup it starts from, the less it has to guess.
For leaders: ask where the rules live
Guardrails are what give your team the confidence to let AI do real work.
When you’re looking at AI tools, ask where the rules are enforced. If the answer is “in the prompt,” keep asking. Here’s what I’d want to know before signing anything:
- What can the AI read, and what can it change?
- What happens when it tries something it isn’t allowed to do?
- Which actions need a person to approve, and can the AI ever skip that step?
- Where does it test its work before it touches live data?
Get this right and your staff stop acting as data couriers. They spend their time on the calls only they can make.
These same rules are built into Cosly, the AI staff team I’m building for small nonprofits. It drafts, and a person approves anything that goes out.