The three fixes that make AI safe and useful at work
Nick Dreyfus on the data rules, skills, and ownership every business needs before they hand employees an AI login.
“Confidence does not mean correct.” — Nick Dreyfus
Nick starts with a simple, uncomfortable test: do you know how your people are actually using AI, or only what your policy says they should? That gap is where most of the damage — and the missed value — lives. His answer is not a single policy or a tool, but three connected fixes: data, skills, and ownership. Get those three right and AI stops being a risky experiment and becomes a dependable part of the job.
Data: stop trusting the tab the person shows you
Files are bigger than the window your employee opens. Nick walks through the obvious but often-ignored risk: a workbook can contain pricing, customer data, hidden sheets and comments that travel with an upload. Asking a model to “look only at the first tab” doesn’t undo the upload.
He recommends a traffic-light rule every employee can remember:
Green: approved, public content employees can use with approved tools.
Yellow: stop and check — permission, destination, exceptions.
Red: restricted — credentials, proprietary costs, personal data. Don’t paste into unapproved tools.
Beyond labeling, you need a governed environment: who accesses what, where inputs are stored, whether the data may be used for model training, and what happens on employee exit. Nick also points to the Hugging Face incident to underline a second point: containment must be engineered. An agent’s “six‑minute deadline” in an evaluation didn’t create a real human permission; it revealed weak boundaries. Limit access, isolate credentials, monitor actions, and give somebody the authority to stop the system.
Skills and adoption: start with the V-bars people already hate
Nick names the work you should automate: V-bars — volume, boring, annoying repetitive tasks. But repetitive doesn’t mean simple. Employees know the exceptions; they know the edge cases. Make them part of the development, and give them protected time, not a homework assignment to “learn AI after hours.”
Find internal champions who understand the actual workflow and pair them with a specialist or partner to build a tested workflow. Use examples tied to real work so the next person doesn’t start from a blank screen.
Prompting is a job skill, but it’s not a substitute for design. Nick reports concrete performance claims from his Fortify AI demos: prompt coaching and shared workflows reduced token usage by up to 54%, reduced bad answers by up to 73%, and — depending on which demo you read — saved up to five hours per employee (he mentions both per month and per week in the conversation). The point isn’t the exact number; it’s that coached prompts plus checkpoints cut both costs and errors.
Include checkpoints: before a file is submitted, does the user know what they’re sending? Before a customer message goes out, who reviews it? Ask the system to identify missing facts rather than invent them.
“A confident answer has to earn the right to become a business action.” — Nick Dreyfus
Ownership: name the person who can pause the system
This is the most neglected part. Nick is blunt: give AI an owner. Not a committee. A named person who approves workflows, maintains them, handles error reports, and can pause a workflow when results become unreliable.
If you’re evaluating a partner, demand specifics about the relationship and the data flow. Ask them to walk you through an actual example using safe data: employee input → storage → connections → final customer output. Get answers on deletion, exportability, and what you retain if the relationship ends. Put those requirements into the contract — verbal assurances don’t count.
Measure early, measure simply. Choose one V-bar, document how often it happens and how long it takes, include time spent correcting errors, then run the workflow and measure the same things. Don’t count time saved unless quality holds. Track serious failures separately — a wrong price or an unauthorized action is not just noise in the average.
Monday moves you can make
Ask the team what AI tools they’re already using. Shadow use is real; Nick cites surveys and real user workarounds (screenshots, personal accounts, phone uploads).
Define your traffic-light rules with examples tied to real job tasks and publish who to ask when uncertain.
Name the owner and two champions, give them time, and pick one V-bar to measure and improve.
Make the fixes small, specific, and measurable. Don’t redesign the company in a day; move from hope to evidence.
If you want the recorded examples and the full checklist Nick uses when evaluating partners, listen to the full conversation — it includes the Hugging Face example, the detailed partner questions, and the practical tests you can run this week.
What you’ll repeat tomorrow: tidy answers are easy; audits, checkpoints, and a named owner are what make them safe. That’s where the real advantage will come over the next three years.