↓ Skip to main content

AI Needs an Employee Handbook

·3236 words·16 mins
CyberBunny
Author
CyberBunny
Over 30 years of experience in technology, making security practical and easy to understand.

Starting a new job usually means getting handed an employee handbook—basically the company’s greatest hits collection of rules, policies, and the all-important leave policy. Most folks give it a quick once-over, click ‘I acknowledge,’ and then promptly lose track of the PDF somewhere in the bottom of their downloads folder. Still, the handbook does serve a real purpose: it lays out what’s expected and what’s off-limits at work. Even if nobody can recite it from memory, just having it around gives everyone a place to look when disagreements or disputes pop up. Think of it as the workplace’s anchor, quietly holding things together while it collects electronic dust.

Let’s be real: nobody is going to win a trivia contest by memorizing the employee handbook, and for most of us, the details fade into the background faster than last year’s password. The real value isn’t in remembering every rule, but in knowing there’s a written record to dig up when things get murky. When confusion hits, at least there’s something to point to—even if it’s buried somewhere on the company intranet, right next to the lunch menu from 2017.

The real win is just having those expectations written down somewhere, so when confusion inevitably strikes, there’s at least a document to point to—even if it’s gathering more virtual dust than your old MySpace page. (And yes, I know you still remember your login.)

AI deserves the same treatment. If we’re going to trust it with our data, our decisions, and maybe even our coffee orders (because who wants to risk decaf?), it needs a rulebook too. Just like employees lean on a handbook for understanding and accountability, AI systems should have their own set of ground rules. As we transfer more responsibilities to AI and weave it deeper into our daily work, not having clear expectations is just asking for confusion, risk, or the kind of unintended consequences that make for awkward all-hands meetings.

In the previous post, I imagined a future organizational chart where human employees and AI workers operate side by side. In my last post, I pictured a time when humans and AI sit side by side on the org chart—no, the AI doesn’t get a cubicle, but you get the idea. Once an AI has a job title, a manager, an identity, the right access, and a to-do list, the organization needs to lay out clear rules for what the AI should and shouldn’t do. Without those boundaries, both people and AI are left to make up the rules as they go, which is a surefire way to end up resulting in confusion and risk. data-handling rules. IT has platform standards. The business has prompts. The vendor has a responsible-use guide.

And somewhere, lurking in a network drive, there’s a steering committee presentation called AI_Governance_Final_v3_REALLYFINAL.pptx. Because let’s face it, naming things is harder than configuring a firewall.

So the organization might have plenty of AI guidance floating around, but what’s missing is a single, coherent set of expectations everyone can actually find and follow.

That’s where the employee-handbook analogy comes in handy.

An AI handbook doesn’t have to be a literal PDF that someone tries to upload into the model and hope it reads. Think of it more as a governance framework—a way to turn organizational expectations into real policies, technical controls, system instructions, monitoring, architecture, and operating procedures.

The key here is consistency.

If the organization says unapproved AI services can’t touch confidential data, that rule should show up everywhere: in policy, technical settings, employee training, access controls, and monitoring. And if some actions need a human to sign off, that shouldn’t just be buried in a governance document nobody opens after launch.

Ideally, the workflow itself should enforce these rules, not just leave them buried in a policy document that everyone ignores after orientation and only remembers when something goes wrong.

Rules only work if the technology actually enforces them. Otherwise, it’s like putting up a ‘Do Not Enter’ sign and then holding the door open with a chair for good measure.

The first chapter of the AI handbook should tackle acceptable use.

Organizations need to spell out what jobs AI is allowed to do, which AI services are approved, what kinds of information employees can feed into those services, and which situations need an extra set of eyes. Can AI write customer-facing content? Can it generate production code? Can it analyze employee data? Can it make recommendations about hiring, finances, cybersecurity, or legal issues? These are the kinds of questions that need clear answers.

There does not need to be one universal answer for every situation.

But there absolutely needs to be a clear boundary. Otherwise, things get messy faster than a toddler let loose with a bowl of spaghetti—and trust me, that’s fast.

If you don’t set those boundaries, employees will happily invent their own. And from experience, their version is usually more creative than compliant—think life hack meets compliance nightmare.

That’s how you end up with shadow AI, which is basically the Wild West of workplace technology: lots of energy, not a lot of rules, and the occasional tumbleweed rolling by.

Most employees aren’t using unapproved AI because they’re plotting against corporate governance. They’re just trying to get their work done faster. If the organization doesn’t give them a clear or practical path, they’ll find one themselves—sometimes with results that make Security reach for the aspirin.

Someone will paste a document into a public chatbot because it saves twenty minutes. Someone else will spin up a personal API key. Another person will connect an AI tool to a cloud app because the connection takes three clicks and looks harmless—until it isn’t.

Eventually, Security stumbles across it—usually right after someone brags about their new productivity hack in the break room, totally unaware they’ve just set off a compliance alarm loud enough to wake the auditors.

This is the classic enterprise technology lifecycle, only now the speed dial is cranked up to eleven and everyone’s wearing roller skates. Welcome to the modern workplace.

Clear acceptable-use rules help employees know where they can experiment and where they’re crossing the line. If people know which services are approved, what information they can use, and when they need to ask for a review, they’re a lot less likely to make up their own policies as they go.

The next big chapter is all about data.

AI is, at its core, a data-consuming machine. Its usefulness—and its risk—depend on what information it can get its digital grasp on. Organizations already classify things like confidential records, customer data, intellectual property, regulated info, and internal-only material.

AI needs those same kinds of data boundaries.

Which data classifications can a particular AI process? Where can that processing happen? Can the service provider keep prompts? Can the vendor use your organization’s info to train their models? What can connectors pull in? What information might show up in generated responses? How long can prompts, outputs, logs, and agent memory stick around? These are the issues that require solutions.

These questions shouldn’t have to be reinvented every time someone spins up a new AI system.

That’s one of the main reasons we have governance in the first place.

A Finance AI and a Marketing AI may have very different access requirements, but both should operate within the organization’s broader data-classification and handling rules. The individual implementation changes. The governing principle does not.

The handbook should also spell out where a human needs to step in and approve something.

Not every AI action needs a person hovering with their finger over the approval button. That would defeat the whole point of automation. But some decisions are important enough that a human still needs to be in the loop.

An AI may draft an external communication, and another may draft an external email, but a person should approve it before it goes out. A security AI might recommend deactivating an account, but an analyst gives the green light. A financial system might identify something weird, but a human makes the final call.This is where we draw the line between advice and authority.

AI can be a fantastic decision-support tool without actually making the decisions. Organizations need to decide where that line is, because if they don’t, automation has a way of quietly taking over more than anyone planned.

First, the AI drafts the message.

Then it drafts and sends the message because humans approve it every time anyway.

Then someone removes the approval step because it slows things down.

Six months later, the AI is making decisions nobody remembers officially handing over.

Governance rarely fails because someone stands up and says, ‘Effective immediately, we’re abandoning all controls.’

It usually unravels one little convenience at a time, because who can resist making things just a bit easier?

An AI handbook also needs clear rules for when to escalate a problem.

Employees are expected to know when to escalate a problem to a manager, Security, Legal, HR, Privacy, or someone else. AI systems need the same kind of guidance.

If information conflicts, what should the AI do? If a source can’t be verified, should it guess or stop? If a request involves restricted info, what process should kick in? If an action goes beyond the AI’s authority, should it refuse, ask for approval, or hand it off to a person?

The answer shouldn’t be to just make it up as you go.

AI needs to know when it’s okay to be uncertain and when it’s time to call for backup.

Sometimes, the best answer an AI—or a person—can give is a simple, honest ‘I don’t know.’ That shouldn’t just be allowed, it should be encouraged.

That might feel strange, since organizations buy AI to answer questions. But an AI that always gives an answer—even when it shouldn’t—is a lot more dangerous than one that knows when to say, ‘I don’t have enough info.’

Let’s face it, humans who confidently answer every question are a risk too—though, ironically, those folks sometimes get promoted.

Maybe we can skip building a robot that always has an answer, especially since we already have enough of those in the office.

Let’s not forget, security requirements deserve a spot in the handbook as well—right alongside the dress code and the reminder not to microwave fish in the break room.

AI identities need to log in using the same approved methods as everyone else—no secret handshakes or back doors allowed. Sensitive information, or ‘secrets’ in tech speak, should be locked down tight. Access should be limited to just what’s needed, which is the fancy way of saying ‘don’t give the AI the keys to the whole building if it only needs to water the plants.’ Anything important or risky should have extra controls, and if the AI wants to connect to the outside world, someone should sign off on it. And of course, we need to keep good records of what the AI is up to, so we can trace things back if something goes wrong.

I realize that list sounds like it was taken directly from a technical manual, not your average employee handbook. But every now and then, you have to let your inner geek have its moment in the spotlight.

And that’s perfectly okay. Sometimes the handbook. Of course, the AI handbook isn’t a single, ancient volume sitting on a shelf somewhere. language I’m pretty comfortable with.

The AI handbook is not one literal document.

Instead, it’s really a collection of all the expectations and rules that guide how our digital coworkers behave.

Some of these expectations show up in policies that actual humans can read without a code key. Others are baked into the technical architecture, identity systems, lines of code, or tucked away in monitoring tools and workflow engines.

The goal is to make certain everything lines up—so the rules on paper match what actually happens in practice.

If the policy says one thing but the technology lets you do another, guess which one wins? Hint: it’s not the policy. If the policy says one thing but the technology lets you do something else, guess which one wins? Spoiler alert: it’s not the policy. Technology has a method of sneaking past the rules if you’re not careful. It can determine where AI shouldn’t be used, even if it technically could be. Should AI make final hiring decisions? Determine disciplinary actions? Approve high-value financial transactions? Make safety decisions? Communicate autonomously during a major crisis?

The answers will differ depending on the organization, industry, legal requirements, and risk threshold.

The key is that someone actually makes a conscious decision, instead of just letting the technology run wild and expecting a positive outcome.

Just because we can do something with AI doesn’t mean we should automatically call it ethical or make it the new normal.

Transparency earns a place in the handbook too, because nobody likes surprises—especially from their software.

When should people know they are interacting with AI? When should machine-generated content be identified? When should employees disclose that AI contributed to an analysis, recommendation, or decision?

There’s no universal answer here. An AI-generated document that’s just for internal use is a completely different beast from something you send to a customer, post online, file as a legal document, or use to decide who gets the coveted corner office.

Organizations need to decide when AI’s involvement is important enough to get a shout-out—kind of like finally giving credit to the intern who did all the heavy lifting.

Retention is a entire different can of worms, since AI can crank out data just as quickly as it devours it.

All those prompts, responses, logs, and even the AI’s own memory can end up packed with your organization’s information—sometimes in places you never even thought to look.

So what actually sticks around? For how long? Where does all that data live? Who gets to peek at it? Can the AI hang onto something forever, like that embarrassing office party photo that refuses to disappear? And if you delete the source, does the AI still have a copy hiding out somewhere?

Suddenly, all those records-management questions get a lot more interesting when your systems can remix, summarize, and spread information more quickly than you can say ‘version control.’

Just because you delete the first document doesn’t mean you’ve tracked down every place its information has wandered off to—sometimes data has a way of hiding in the digital equivalent of the office storage closet.

The AI handbook therefore needs to comply with existing legal, privacy, retention, and business requirements.

Otherwise, you could end up with the corporate version of an elephant: never forgetting, never letting go, and always ready to remind you about that one email you sent in 2017 when you definitely needed another cup of espresso.

Exceptions also need a defined process.

No policy can anticipate every legitimate business situation. A research team may need temporary access to data that is normally restricted from AI processing. A pilot may require an integration that is not usually approved. A particular business process might justify expanded capabilities for a limited period.

The answer does not always need to be “no.”

It should be, “There is an exception process.”

That process should identify the owner, business justification, duration, additional controls, review requirements, and expiration date.

The expiration date matters.

Temporary exceptions, much like those so-called temporary permissions, have a sneaky way of becoming permanent. It’s a bit like that ’temporary’ folder on your desktop that’s been quietly collecting virtual dust since 2012.

At least with an expiration date, there’s hope we can clean things up before they become part of office legend.

The handbook should also describe what happens when expectations are violated.

For human employees, policy violations might cause training, restricted access, disciplinary action, or eventually termination.

AI, of course, is not going to lose any sleep over a sternly worded memo in its personnel file—assuming it even knows what sleep is.

But there are still equivalent controls.

An AI can be reconfigured. Its access can be reduced. You can require additional human approval for its actions. A connector can be removed. Its permissions can be narrowed. The agent can be disabled entirely.

Think of it as providing the AI a digital time-out. No need for a lecture—just a few clicks and its privileges are gone.

If an AI repeatedly operates outside its defined expectations, the solution shouldn’t always be a further sentence added to the system prompt: “Please do not do that again.”

At some point, technical enforcement is appropriate.

The same principle applies to employees using AI.

If people misuse AI services, the organization needs an answer that coordinates training, education, monitoring, and enforcement. Many problems will come from misunderstanding rather than malicious intent, especially as AI capabilities and business policies evolve rapidly.

Training really does matter, because nobody is born knowing how to use AI responsibly—not even the AI itself.

And let’s remember monitoring. It’s basically the digital version of keeping an eye on the office cookie jar—just to make sure nobody’s sneaking an extra treat.

That means the AI handbook really has two audiences.

The AI worker needs technical guardrails.

The human workforce needs clear rules for how to use, supervise, verify, and interact with it.

Both sides matter.

Over time, organizations may formalize these expectations through an AI acceptable-use policy, AI security standard, development requirements, data-governance controls, model-risk processes, agent lifecycle procedures, and other documents.

The names matter less than the coverage.

What matters is whether people throughout the organization can answer the important questions. What AI can we use? What information can it access? What actions can it perform? When is human approval required? How is it monitored? How long is information retained? Who owns it? What occurs when something goes wrong?

If those answers are scattered across ten documents, three committees, four SharePoint sites, and the memory of someone who left the company last year, you don’t really have an AI handbook—you have a scavenger hunt.

What you really have is folklore: tales handed down from one IT veteran to the next, usually around coffee and a sigh.

AI governance needs something better.

The good news is that most of the concepts organizations need already exist.

Acceptable use. Security. Privacy. Records management. Access control. Change management. Incident response. Ethics. Identity governance. Vendor management.

AI does not replace those policies.

It crosses all of them.

The purpose of an AI handbook is to connect those existing principles to a new kind of worker and keep expectations consistent, whether a human performs the work, an application does, or an autonomous agent does.

And with that, we’re almost at the end of this series.

We’ve treated AI like a real employee: gave it a job description, assigned it a manager, set up permissions and performance expectations, handed out a badge number, applied Zero Trust principles, and even found it a spot on the org chart.

Now, we’ve even given it an employee handbook.

At some point, we have to step back and admit what we’ve created here.

We’re not only buying software anymore.

We’re actually introducing a whole new kind of worker to the organization.

Maybe it’s time to stop thinking about buying AI. The mindset is too small.

We should stop buying AI.

And start thinking about hiring it instead.

And that’s where we finish this series.