The VAULT Framework: How to Use AI Safely at Work
VAULT is a five-part framework for responsible AI use at work: Verify, Augment, Understand why, Loop humans in, Transparency. Here's how it works.
What is the VAULT framework?
VAULT is a five-part set of principles for using AI responsibly in business workflows: Verify the output, Augment instead of replace, Understand the why before you build, Loop humans in at the right checkpoints, and keep the whole system Transparent. The framework was distilled from practices used at Goldman Sachs, where engineering leadership (including CIO Marco Argenti) has spoken publicly about treating AI outputs with skepticism, keeping humans in control of consequential actions, and requiring that automated systems be explainable to regulators, clients, and internal teams.
TL;DR
- VAULT stands for Verify, Augment, Understand why, Loop humans in, and Transparency, five checks meant to catch the ways AI systems quietly fail in real workflows.
- Verification has to start before the AI even runs, because a confident, well-formatted answer built on messy or outdated data is still wrong, just wrong faster.
- Augment, don’t replace, means AI isn’t always the right tool: deterministic automation is often cheaper, more testable, and more reliable for tasks with a known, fixed set of steps.
- Every AI project should start with a one-sentence problem statement, not with a tool or a trend, because a well-built system nobody needs solves nothing.
- Human oversight should scale with the stakes of the action, not apply uniformly, so low-risk tasks can run autonomously while anything touching money, data, or large audiences gets a checkpoint.
- A real-world failure illustrates the risk: an autonomous agent misinterpreted a task and sent a discount code to roughly 200,000 people, the kind of mistake that “loop humans in” is designed to prevent.
- Transparency means being able to show your work: sources, assumptions, tools used, and decision points logged well enough that someone else could pick up the system and understand it.
One coffee. One working app.
You bring the idea. Remy manages the project.
How does the “Verify” step actually work?
Verification in VAULT isn’t a single spot-check at the end. It starts with the data feeding the system and continues through the final output.
The core idea is that a model’s reasoning process and its final answer are two different things to evaluate. A model can walk through a problem in a way that’s genuinely useful, breaking down assumptions or surfacing a helpful angle, even if the conclusion it lands on is wrong. That means you get value from watching how it thinks, but you still can’t take the final answer at face value.
Practically, this means building verification into the process itself rather than trusting the output because it looks polished. That can be as simple as prompting a model to recheck every number and factual claim, cite a source for each one, and flag anything it isn’t fully confident about. You don’t need to manually verify every sentence a model produces. In something like a business report, there are usually only a handful of numbers that could actually change a decision, so those are the ones worth spot-checking directly. For higher-stakes work, a separate AI reviewer, or a small team of reviewing agents, can check outputs before a human ever sees them, adding a layer of iteration before anything ships. But important claims still need original sources, deterministic tests, or a human review at some point. No amount of AI-on-AI checking replaces that.
Why does “Augment, don’t replace” matter?
Not every workflow problem needs an AI agent. Some of them need a well-built, deterministic automation instead, and confusing the two is one of the most common mistakes people make right now.
Deterministic automation runs the exact same process the same way every single time. That doesn’t guarantee correctness (bad logic repeated perfectly is still bad logic), but it makes systems far easier to test, audit, and fix when something breaks. Humans, by contrast, get tired, skip steps, and drift into slightly different versions of “the same” process over time. AI models introduce a different kind of unpredictability: judgment and flexibility, which is valuable for messy or ambiguous problems but overkill (and riskier) for tasks with a known, fixed set of steps.
The practical rule: if a task has clear rules and a known correct answer, a normal automation is probably the better choice. Use AI when the work actually requires interpretation, flexibility, or handling messy, unstructured information. The strongest systems often combine both: a deterministic pipeline pulls and cleans data, runs calculations, and verifies the numbers, while an AI layer explains what changed in plain language on top of that verified foundation. That division of labor lets automation handle facts and AI handle communication, rather than asking one tool to do both jobs at once.
Why does “Understand the why” come before building anything?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
It’s easy to open an AI tool because you want to build an agent, or because a workflow sounds interesting, without first defining the problem you’re solving. That produces impressive systems nobody actually needs.
The fix is a simple forcing function: write one sentence describing the problem you’re trying to solve, and one sentence describing what a good result looks like. If you can’t fill in both blanks, you’re not ready to start prompting yet, because a model can’t figure out what you want if you haven’t figured it out yourself. When the problem isn’t fully clear, AI can still help as a brainstorming partner, working through context and constraints to help find the simplest path forward, rather than being handed a vague goal and asked to guess.
This matters most when the goal is practical impact, whether that’s saving time internally or building something people will pay for. Starting with a specific, painful problem and building one system that solves it tends to outperform building an impressive but unfocused stack of AI workflows.
What does “Loop humans in” mean in practice?
AI systems act like a megaphone: whatever you give them gets amplified across a workflow, mistakes included. A slightly unclear instruction in a one-off chat might produce one wrong answer you can immediately catch. That same instruction handed to an autonomous agent can send the wrong email, update the wrong record, or contact an entire client list before a human ever sees it.
That’s not hypothetical. One documented case involved an autonomous agent that reviewed a task list, proactively pulled an item off it, misinterpreted the task, and sent a discount code to nearly 200,000 people. The lesson isn’t that every AI action needs manual approval. It’s that the level of human oversight should match the consequence of the action. Letting an AI tool organize personal notes with no oversight is fine. Letting it send client communications, touch important data, or spend money without a checkpoint is not.
A practical technique is setting automations to draft rather than send by default: AI writes the email into a Gmail draft, proposes the reply in a document, or shows the plan before changing files or deploying anything. That preserves most of the speed benefit while giving a human a chance to catch a bad output before it becomes a real problem. The guiding assumption should be that if an agent can technically take a risky action, it eventually will, even if it behaves correctly the vast majority of the time.
Why does transparency matter for AI systems?
A human can’t approve or trust work they can’t see the reasoning behind. In regulated environments, anyone building something that touches important data, reporting, or risk needs to be able to explain where information came from, what happened to it, and how a result was produced, because regulators, clients, and other teams may need to review those decisions later.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The same standard is useful outside finance. If an AI system works but nobody understands how, it becomes a serious liability the moment something breaks, the underlying data changes, or someone else has to take over the project. The fix is building documentation into the system itself: logging data sources, assumptions, tools used, validation checks, and key decisions for every run. A report should show which source supports each conclusion. An automation should make clear what triggers it, what happens at each step, and where a human needs to sign off.
Transparency here doesn’t mean exposing a model’s raw internal reasoning. It means producing inspectable evidence, the sources, inputs, actions, and checks, so the system is easier to trust, easier to improve, and easier to hand off.
Frequently Asked Questions
What does VAULT stand for?
Verify the output, Augment don’t replace, Understand the why, Loop humans in, and Transparency. It’s a five-part framework for using AI responsibly in workflows where mistakes carry real cost.
Is the VAULT framework specific to Goldman Sachs?
The principles were drawn from practices and public commentary associated with Goldman Sachs, including remarks from CIO Marco Argenti about model reasoning versus output, and about keeping humans in the loop until AI systems are consistently reliable. The framework itself is meant to apply to any AI tool or workflow, not just financial services.
When should you use automation instead of AI?
When a task has a clear, known set of steps and a predictable correct answer. Deterministic automation handles that kind of work more reliably and is easier to test and audit. AI is better suited to tasks requiring judgment, interpretation, or handling ambiguous information.
How much human oversight does an AI agent actually need?
It should scale with the consequences of the action, not be uniform. Low-stakes, reversible tasks can run with little oversight. Anything involving money, sensitive data, or communication with a large group should include an approval step, such as drafting instead of auto-sending.
Why is transparency important if an AI system already works?
Because “it works” isn’t the same as “it’s trustworthy.” If nobody can explain how a system reached its result, it becomes very hard to fix when it breaks, audit when questioned, or hand off to someone else.