You Have a Standard for Color. Where's Your Standard for AI?

Your shop has a standard for color. Someone owns the targets, someone signs off on the proof, and nobody ships a job that fails the check. You have safety protocols on the floor, lockout procedures on the press, and quality gates at every stage. You would never put a new operator on a seven-figure press with no training, no limits, and no one to answer to.

Now ask what your team is allowed to do with AI. Who decided? What data can go in? What happens when it gets something wrong? If the honest answer is a long pause and a look at the ceiling, you are not alone. You are the norm.

Grant Thornton's "2026 AI Impact Survey" found that only 7% of manufacturers have a tested AI incident response playbook, the lowest of any industry surveyed. Just 14% say they are extremely prepared for AI privacy and security challenges, compared with 40% overall. And 12% are confident they could pass an independent AI governance audit, against 22% across the full sample.

Those are manufacturing numbers, not print numbers. Nobody has clean print-specific data yet. But I have spent enough time inside print and packaging operations to tell you we are not the exception. We are just less likely to say it out loud.

First, let's define the words, because they are being used badly.

Half the AI conversation in this industry is people using three different terms to mean the same thing. They are not the same thing, and the difference is exactly where your risk lives.

An AI assistant answers when you ask. You type a question, it gives you a response, and a person decides what to do with it. ChatGPT drafting a customer email is an assistant. Copilot summarizing a meeting is an assistant. The human is still the one pressing send.

An AI agent is software built on an AI model that has been handed a goal and given access to your systems so it can act on it. It reads the job ticket, checks substrate stock, drafts the quote, and updates the schedule. It does not just say things. It does things.

Agentic AI is what you have when those agents plan multi-step work, choose their own tools, and carry it out with little or no human in the middle. Think of it this way: The agent is the worker, and agentic describes how the work gets done. Autonomy, multiple steps, real systems touched. When several agents work together, that is a multi-agent system, and it is where things get interesting in a bad way.

Here is why the definitions matter. When an assistant is wrong, you get a bad paragraph. When an agent is wrong, you get a sent quote, a placed purchase order, or a rescheduled press run. Assistant risk is bad words. Agent risk is bad actions. Most governance conversations in our industry are still built for the first one.

Where print and packaging shops are already exposed

You do not need to wait for a big agentic rollout to have a problem. You probably already have three.

Your CSR has a free ChatGPT account and is pasting customer job specs and pricing into it to write faster replies. She is not being careless. She is being efficient. In PagerDuty's 2026 Shadow AI Survey, 66% of office professionals said they had used AI tools at work even though they believed company policy did not allow it. A meaningful share admitted sharing correspondence and customer data. That survey covered larger companies than most print shops, so if you think your people are the exception, I would love to see your evidence.

Your sharpest operations person built an estimating or scheduling agent over a few weekends, and it has write access to your MIS. I respect the initiative. But an agent with that access is closer to a new hire than a new tool, and nobody onboarded it.

Your prepress automation was configured five years ago by someone who has since left, and it now routes jobs, checks files, and triggers production steps in ways nobody fully remembers setting up.

In all three cases, there is no rule. Not a bad rule. No rule.

"But we're on the enterprise version."

I hear this in nearly every conversation about controlling AI. We are on the enterprise version of whatever, so we are safe. It is a reasonable thing to believe, and it is half right.

An enterprise license protects you on the vendor's side of the fence. Depending on the product and the contract, that usually means your data is kept out of public model training, plus single sign-on, admin controls, and some logging. Good. Buy it.

What it does not do is govern your side. It does not decide what your people may put in, what an agent is allowed to touch, who is accountable when an output is wrong, or what happens when the vendor changes its terms. An enterprise license is a locked warehouse. It does not tell you who holds the keys or what they are carrying out the back door.

And people do go around it. In BlackFog's 2026 research, 34% of employees admitted using free versions of company-approved AI tools. IBM's 2026 breach research, as reported by Cybersecurity Dive, found that more than two-thirds of breached organizations lacked governance processes to limit shadow AI. The license was in place. The rules were not.

No license comes with a tested incident response playbook, either. That is the 7%. An enterprise version is a purchase. A standard is a decision.

Three controls for any AI your team touches

This is the core of what we are teaching in our Nov. 10 webinar, and it fits on a sticky note. For every AI tool or agent in your operation, you need to be able to answer three questions.

1. What can it reach? Limit it. Which data, which systems, which customers' artwork. An estimating assistant does not need your payroll file. A scheduling agent does not need to send email to customers. Give each one the smallest set of access that does the job, the same way you would for a new hire in their first week.
2. What did it do? Record it. If an AI-influenced decision goes wrong, can you reconstruct what it saw, what it chose, and why? In a separate Kiteworks report, only 19% of manufacturers said they maintain evidence-quality audit trails. That sample was small, so treat it as directional, but the direction is not encouraging. If you cannot see it, you cannot explain it to a customer, an auditor, or your own CFO.
3. Who can stop it? Be able to stop it. Name a person, by name, with the authority to pull the plug. Not a committee. Not IT in general. If the answer is "we would figure it out," you have already told me what happens at 2 am when a job is in the wrong queue and an agent is still making decisions about it.

Limit it. Record it. Stop it. If you cannot answer all three for a given tool, that tool is running on trust.

You do not need a task force. You need a starter set of rules.

This is the part where most organizations overbuild. They form a committee, commission a 40-page policy, and by the time it is approved the tools have changed twice. Do not do that.

Start with one page. Which categories of information never go into an external AI tool until you have made a deliberate choice: customer pricing, unreleased artwork, estimating logic. Which decisions AI may make on its own, which it may recommend but a person must approve, and which stay entirely with a human. Who owns the list. Then review it quarterly, because the tools will keep changing whether you are ready or not.

Then use the same three questions to judge any AI vendor before you sign. Can I limit what it reaches? Can I see what it did? Can I stop it? A vendor who gets vague on any of those is telling you how the relationship will go when something breaks.

This is not about slowing down. Grant Thornton's own conclusion points the same way: Leaders who invest in governance are moving faster, because they have the confidence to scale. You do not put guardrails on a press to make it run slower. You put them there so you can run them at full speed without losing a hand.

The standard comes before the speed

You already know how to do this. You built a standard for color, for safety, and for quality, and your business is better for every one of them. AI deserves the same discipline, and right now it is the one part of the operation where nobody is watching.

That changes with one page, one named owner, and three questions.

On Nov. 10, we are running "Nobody's Watching," a 75-minute live webinar for print and packaging teams. No software demos and no theory. You leave with a starter set of AI rules you can finish and take to your team. Register for the "Nobody's Watching" webinar here.