Writing an AI Policy for Your Company: What It Needs to Say, and Why the Long Ones Get Ignored
Someone at your company is already using AI. Probably several people, probably on accounts you do not know about, probably with customer information pasted into them. That is not a discipline problem and it is usually not a security breach. It is what happens when a useful tool is free and nobody has said anything one way or the other. The question a policy answers is not whether to allow it, but who gets to decide, and what you can show about that decision later when a customer, an insurer, or opposing counsel asks you.
This is written for Washington business owners and the people who run their operations, not for lawyers. It is about the document itself: what goes in it, what stays out, who it binds, and why the long ones fail. I have written separately about which Washington AI laws actually reach an ordinary business, and the short answer there is that most of them do not reach you yet. That is exactly why this piece is not about statutes. The exposure that is live today comes from the agreements you have already signed and the obligations you already owe, not from anything that takes effect in 2027.
Do we actually need a written AI policy?
Not every company needs a document, though a regulator or a customer contract may require one. If you have four employees and you make every decision yourself, you may not need a policy so much as a habit. But every company needs an answer to three questions, and the document is simply the cheapest way to hold the answer in a form you can hand to somebody.
The three questions are: which tools are people allowed to use, what information is allowed to go into them, and who decides when someone wants to use something new. A company that can answer those three in a conversation is in better shape than one with a fourteen-page policy nobody has read.
What has changed in the last year is who asks. Customer security questionnaires increasingly include AI questions. Some carriers have started asking at renewal. Larger customers are adding AI terms to contracts they send you, and some of those terms require you to represent that you have a policy. The document is increasingly a commercial prerequisite rather than an internal nicety, and that is often the reason a company finally writes one.
What does an AI policy actually need to say?
A working policy is short and answers a handful of practical questions in language an ordinary employee can act on. It names what it covers. It lists the tools that are approved right now. It says what categories of information may and may not go into them. It names the route for getting something new approved, and the person at the end of that route. It says that a person remains accountable for anything the company sends out or relies on, regardless of what produced the draft. And it says what happens if someone ignores it.
The test is not whether the policy is thorough. The test is whether a new employee can read it in five minutes and know what to do on Monday morning when they want to summarize a customer call. If it does not answer that, it is a compliance artifact rather than a policy.
What should stay out
Resist the urge to put technical detail or vendor-specific instructions in the policy itself. Those change every few months, and every change forces a re-adoption and a re-acknowledgment. Keep the policy to the stable part and push the volatile part into a separate list.
Which tools is this policy even about?
This sounds like a definitional footnote and it is actually the part that decides whether the policy works. If you define AI as any system that uses machine learning, your policy now covers the spam filter, the spellchecker, the fraud screening on your card terminal, and the recommendation engine in your shopping cart. Nobody can comply with that, so nobody tries.
A more usable line is narrower. The policy covers tools where an employee sends company information out to a system that generates a response, and tools that produce content or recommendations the company then relies on. That captures the general-purpose chatbots, the writing assistants, the meeting transcribers, the code assistants, and the industry-specific products built on top of them. It leaves out the plumbing nobody is making judgment calls about.
The features inside software you already bought
This is where most companies are wrong about their own footprint. You did not buy an AI tool, so you believe you do not have one. Meanwhile your CRM added a summarizer, your help desk added suggested replies, your video conferencing added an automatic notetaker, and your accounting package added anomaly detection. None of those went through a purchasing decision, because they arrived as updates to software you were already paying for.
Those features are frequently processing exactly the information you would have been most careful about if someone had asked you first. A policy that only addresses tools people go out and sign up for will miss them entirely. Say explicitly that the policy covers AI features turned on inside existing systems, and that enabling one is a decision, not a settings change.
Who decides whether a new tool gets used?
Name a person. Not a committee, and not a department. A committee meets monthly, and an employee who needs an answer on Tuesday will not wait three weeks for one. They will use their personal account instead, and now the tool is in your business with none of the terms you would have negotiated and no record that it is there.
Speed is a governance control, not a convenience. If your approval route cannot produce an answer in a few days, the route will be bypassed, and the policy will be a document describing a company you do not have. Build the thing people will actually use.
The three questions the approver asks
What information goes into it. What does the vendor do with that information, specifically whether it is used to train models, how long it is retained, and who else can see it. And what is the consequence if the output is wrong, which is a different question for a tool drafting marketing copy than for one screening job applicants or flagging transactions.
Write the answer down. One short paragraph per tool, dated, saying what it is for, what may go into it, and who approved it. After a year you have an inventory that was built as a byproduct of working rather than as a project. That inventory is what you will reach for first when a customer, a regulator, or a plaintiff asks what you are running.
What should the policy say about company and customer data?
Sort information into a small number of buckets and attach a rule to each. Most businesses need three or four. Information you have already published can generally go anywhere, unless a contract says otherwise. Internal operating information can go into approved tools. Customer information, personnel records, anything covered by a confidentiality obligation, and anything regulated should go only into tools that have been approved in writing for that specific purpose, and in many businesses should not leave the company at all.
Keep the categories concrete and name real examples from your own business. A policy that says employees must protect sensitive information has told them nothing. A policy that says customer lists, pricing, anything from a personnel file, and anything a client sent you under an NDA do not go into a general chatbot has told them what to do.
The part that is genuinely urgent
Here is the exposure most businesses have not considered, and it has nothing to do with AI law. You have probably promised somebody that their information stays in particular hands. It might be a confidentiality clause in a customer contract, a data processing agreement that controls which subprocessors may touch the data, a nondisclosure agreement from a deal, a government contract, or a professional obligation if you are in a licensed field.
Most of those promises were written before anyone was thinking about AI, and many of them are worded broadly enough to cover sending that information to an AI vendor. An employee pasting a customer file into a chatbot may be breaching a contract you signed three years ago. That is a live, present-day risk with a real counterparty who can enforce it, and it is usually closer to hand than a statute that has not taken effect yet. It is worth having someone read your top handful of customer agreements with this specific question in mind.
If this is describing your situation, tell me what is going on. I read what comes in myself, and I will tell you whether it is something I can help with.
Tell me about your situationDoes anyone have to check what the tool produces?
Yes, and the policy should say who. The useful formulation is that the person who sends it, signs it, or acts on it owns it. What produced the draft does not change who is answerable for it. That single sentence does more work than a page about responsible use.
Then name the categories that require a check before anything moves. Anything that goes to a customer. Anything with a number in it that someone will rely on. Anything concerning an employee or an applicant. Anything that becomes a business record or goes into a filing. Anything that states a commitment, a price, or a date.
Verification has to be built into how the work moves, not stated as a value. An instruction to review AI output carefully tends to fade within weeks. A step the work cannot get past without someone signing off tends to hold. If the check only exists in the policy and not in the workflow, assume it is not happening.
Why do long AI policies get ignored?
Because they are written to survive a review rather than to be used. A fourteen-page policy gets acknowledged once during onboarding, filed, and never opened again. It makes the company feel covered while changing nobody's behavior, which is the worst of both outcomes: you have the obligations the document created and none of the protection it was supposed to provide.
Split the document in two. The policy proper is two pages and holds the parts that do not change: scope, the data rules, the approval route, human accountability, consequences. The approved tools list is a separate page, dated, owned by the person who does approvals, and updated whenever something is added or dropped. Changing the list does not require re-adopting the policy, which means the list stays current, which means people trust it enough to check it.
Then set a date to revisit it. Not a vague commitment to review periodically. An actual date on an actual calendar, with a name next to it. Six months is about right in current conditions.
What if people are already doing it?
They are. By the time a company gets around to writing a policy, the thing the policy was meant to prevent has usually already happened somewhere in the building, and often for a year. Writing the document as though you are starting from zero produces a policy aimed at a company that does not exist.
The more productive opening move is to ask, once, with no consequence attached. Tell people you are putting a policy together, that you want to know what is actually being used, and that nobody is in trouble for answering. You will learn more in a week than a technical audit turns up in a month, because most of this is running on personal accounts and browser tabs that no audit of company systems will ever see.
You will also usually find that some of it is genuinely good. People have solved real problems with these tools. A policy that starts from what is working and puts boundaries around it will be followed. A policy that arrives as a crackdown on something people have been quietly doing for a year will push it further out of view, which leaves you with the same exposure and less visibility than you had before.
Where this goes wrong
A policy copied from a template written for a company ten times your size, imposing approval steps you have no one to staff.
Defining AI so broadly that the policy covers the spam filter, which teaches everyone that the policy is not meant literally.
Addressing only the tools employees sign up for, while the AI features switched on inside your existing software process customer records untouched by any of it.
An approval route with no named person at the end of it, or one slow enough that going around it is the reasonable choice.
Confidential customer information in a general chatbot, in breach of a contract signed years before anyone had this conversation.
A policy adopted, acknowledged, filed, and never checked against what people are actually doing.
Treating the 2027 Washington statutes as the deadline. Many businesses fall outside them, while the obligations that can already be enforced against you are in force now: the ones in your contracts, and the consumer protection, privacy and employment rules that do not mention AI at all.
What a lawyer does here
Reads the agreements you have already signed and tells you what they permit, which is the question that decides most of this and the one a policy template cannot answer.
Writes the policy to the size of the business, in language your employees can follow, rather than to the size of the risk somebody is selling you.
Sets up the approval step so that it produces a decision quickly and leaves a record, because the record is what you will need later.
Works out what you can truthfully say when a customer questionnaire or an insurer asks whether you have a policy, which is a question with legal consequences for the answer.
Works through it with you when something goes wrong, when the question becomes not only what the policy said but what you can show about what you actually did.
About my practice
I am an independent attorney in downtown Spokane, at 601 W. 1st Ave., Suite 1400, admitted in Washington and in Oregon. Alongside business, real estate, land use and civil litigation work, I advise companies on AI governance, technology and SaaS vendor agreements, and the policies and contracts that decide who carries the risk when an AI system gets something wrong. I build and run AI systems inside my own practice. I represent clients throughout Washington.
If you want to talk
The first conversation is short and mostly diagnostic: what your company is already using, what is written down, what your customers and insurers have started asking, and what you are being sold. That is usually enough to say whether what you need is a policy, a contract read, an inventory, or a decision.
Disclaimer
This article is general information about Washington law, written in November 2026. It is not legal advice, and reading it does not create an attorney-client relationship. Whether any of it applies to your company depends on your contracts, your industry, and facts this article does not know. Laws and obligations change. Get advice on your own situation before relying on any of this.
This article is general information about Washington law, current as of publication. It is not legal advice, it does not create an attorney-client relationship, and the law changes. For advice on your own situation, speak to a lawyer.