AI Policy for Employees: Which Tasks AI Drafts, Which It Runs, and Which Always Need a Person

Two-thirds of office professionals in a June 2026 survey said they had used AI at work even though they believed company policy did not allow it, and 88% had shared work information with public AI tools.¹ The survey covered 1,250 people at organizations with $500 million or more in annual revenue in four countries, so read it as a signal, not a measure of your company. The signal is still plain: a policy that exists on paper does not decide what happens in the work, so people decide for themselves.

An AI policy, in practice, is one decision repeated for every task: does AI draft it, does AI run it, or does it stop for a person. This article gives you four questions that sort any task into one of those three groups, a sign-off standard that makes the third group real, and a one-page policy you can write this week. If you would rather talk it through, a free 30-minute Strategy Call with an AI adoption specialist ends with the first two or three workflows worth automating.

01. When the policy exists on paper and not in the work

The scene you may already know. A controller pastes month-end variance notes into a public AI tool to tidy the wording. An estimator asks AI to summarize a spec section before a tender closes. Neither was told yes. Neither was told no.

Both did something reasonable with the information they had, and neither has any idea whether it was allowed. If the rule is not where the work happens, each person is writing it for you.

The gap shows up in every sample. The samples differ, so the figures are not comparable with each other, but they point the same way.

• Resume Now, a career platform, surveyed 1,020 employees in May 2026. Of them, 41% said their employer had given no tools, training or guidance for AI, and 21% said they had clear, role-specific guidelines.²
• HR Partner, which sells HR software, surveyed HR professionals and business owners in the USA, UK and Australia, with data from teams of 20 to 500 people. It found 80% of the HR professionals using AI tools and 23% of businesses with a formal AI policy.³
• Pax8, a marketplace that sells technology to small and midsized businesses, surveyed 1,150 leaders at companies of 5 to 499 employees in the US, UK, Germany and Sweden for its third-quarter 2026 report. Among businesses it counted as AI-engaged, 35% had found AI running without approval in the previous six months.⁴

What leaks, and who sets the example. In the first survey, policies existed in most respondents' minds: 86% worked at organizations they believed had AI policies in place.¹ Even so, 31% had shared financial information or confidential company documents and strategies with AI tools, and 81% believed leadership operates under a different set of rules than everyone else.¹

The 81% figure is the one for an owner or CEO to sit with. A rule that leadership skips is not a rule, and employees notice. Again, the sample is large companies in four countries, and the question asked what people believed, not what policy said.

What follows. The rest of this article is a method. First, three groups and why the sort is the policy. Then four questions that do the sorting, what the sorting looks like in your kind of company, what makes a person's sign-off real, and a one-page policy that holds all of it.

02. Three groups, and why the sort is the policy

The definition, plainly. The policy is a sort of tasks into three groups.

• First group: AI drafts and a person finishes.
• Second group: AI runs within limits you set and a person checks a sample.
• Third group: a person always does the task or makes the decision, and AI at most helps with preparation.

The policy is different from a list of banned tools, from a legal document and from an enterprise governance programme. It is closer to a decision you make once per kind of task and write down. We work to the same three groups on our own projects.

The rule first appeared in the introduction to AI adoption, alongside the point that your AI policy will show up in your partner's contract. This article is the longer version.

Why a sort beats a list of rules. Nobody can map "never paste confidential data into an AI tool" onto a Tuesday afternoon task. People can answer "which group is this in." The Canadian Centre for Cyber Security advises organizations to identify the policies that say how AI is used and what content it may generate, including the oversight and review processes required.⁵

The Government of Canada's guide for federal institutions makes the same point from the other side: the risk of generative AI depends on what it is used for and the mitigations in place. It gives a low-risk example, editing a draft that will go through further review, and higher-risk ones, including a chatbot used by the public and a summary of client information.⁶ That guide is written for federal institutions, so treat it as a worked model, not a rule for your company.

Who answers for the output. In February 2024 the Civil Resolution Tribunal of British Columbia decided Moffatt v. Air Canada. The airline had argued that it could not be held responsible for what its website chatbot said. The tribunal called that "a remarkable submission," held the company responsible for all the information on its website, and ordered it to pay $812.02 including damages, interest and fees.⁷

Moffatt involved a customer-facing chatbot on a public website, not an internal tool. The point carried over is narrow: the company answers for what its tool says. CPA Ontario's guidance, published in June 2026, says the same about professional services. Accountability stays with the CPA, and every AI-generated output must be reviewed, verified and understood before it is used.⁸ Two settings, one point: accountability does not move to the tool.

If you outgrow one page. A standard exists. ISO/IEC 42001 specifies requirements for an AI management system, and it is written for organizations of any size that use AI-based products or services.⁹ You do not need it to start. It is useful to know that the fuller version exists and what it is called.

03. Four questions that sort any task

How to use them. Pick one task. Ask the four questions in order. A "yes" to any danger question moves the task up a group. This is a working method, not a standard, and no source below endorses it. It draws on how the sources below reason about risk.

Question 1: What data does it touch? Personal, confidential or financial data changes the group before anything else does. The Cyber Centre says to avoid providing personally identifiable information or sensitive corporate data in queries or prompts.⁵ CPA Ontario's confidentiality rule, Rule 208, says confidential, restricted or personally identifiable information goes into AI tools only where the environment is verified as secure and compliant.⁸

The exposure is not hypothetical. In the first survey above, 31% had already shared financial information or confidential company documents and strategies.¹ If the task touches restricted data and you cannot say where the data goes, it belongs in the third group. Who may see which data is also the fifth check in the data readiness checks, so start there.

Question 2: Who sees the output? Internal scratch work and anything a client, lender, board, regulator or customer sees are different tasks. Moffatt is the customer-facing case.⁷ The federal guide's low-risk and higher-risk examples map onto the same line: a draft that will be reviewed again is low risk, and something the public or a client relies on is not.⁶ The rule is short: anything that leaves the building has a person on it.

Question 3: Can it be undone? A draft email, a payment and a posted accounting entry carry different consequences. The Government of Canada's Algorithmic Impact Assessment, a federal tool for automated decision systems, ties impact levels in part to how long a decision's effects last and whether they can be reversed. It says requirements are proportionate to the impact level, with peer review and human involvement increasing at higher levels.¹⁰

Article 14 of the EU AI Act, which applies to high-risk AI systems, says human oversight measures must be commensurate with the risks, level of autonomy and context of use.¹¹ Neither binds your company. Both state the same idea: the harder a mistake is to reverse, the more a person has to be involved.

Question 4: How fast can a person tell it is wrong? ConstructConnect, a company that sells takeoff and estimating software, published a review guide for AI takeoffs in August 2026. It sorts outputs by review level:

• A spot-check for standard counts and areas on clean plans.
• A full review for complex geometry and low-quality scans.
• An always-review rule for scope boundaries and specifications.

ConstructConnect says its own figures are benchmarks, not averages.¹² The general version is useful beyond estimating. If you can check an answer against a number you already trust in a few minutes, it can sit in a lower group. If checking takes an afternoon, it cannot. That is the idea behind the data readiness test where you check an AI answer against numbers you trust.

Putting it together. If the task touches restricted data, leaves the building, is hard to undo or is hard to check, it goes in the third group: a person always. If you are comfortable on all four, start in the first group, where AI drafts and a person finishes. Move a task to the second group, where AI runs, only after a trial with sampled checks. The default is drafts.

04. What the sort looks like in your kind of company

These are patterns, not benchmarks, and not a measure of your company.

Professional services. Typical tasks: the weekly client update, a meeting summary, a proposal first draft, client analysis. A first draft and a summary fit the first group. Anything that states a fact to a client goes through a named reviewer. CPA Ontario's guidance is the clearest current statement of that standard in a professional setting.⁸

The cautionary case is Deloitte Australia, which agreed to partially refund the AU$440,000 paid by the Australian government for a report with apparent AI-generated errors, including a fabricated quote from a federal court judgment and references to nonexistent research papers, according to the Associated Press.¹³ Only what AP reports is stated here, and "apparent" is AP's word.

On our own work, our agents draft the weekly client report and save it as an email draft, and a person reviews and sends it. The full walk-through is how we run weekly client status reports. See also AI for professional services.

Construction and specialty contractors. Typical tasks: a first-pass takeoff, a tender estimate draft, an RFI summary, a subcontractor scope comparison, change order wording. AI can draft the takeoff, and the estimator signs off. Scope, specifications, pricing and risk are always a person, which is the same line ConstructConnect draws for always-review items.¹²

The ConstructConnect guide comes from a vendor with an interest in the answer, so treat it as one practitioner view. Our own estimating work has the same shape: a draft estimate that the estimator reviews and signs off. See AI for contractors.

Finance teams. Typical tasks: variance commentary, an ad hoc report from the ERP, invoice coding, journal entry proposals, month-end notes. Commentary fits the first group. A report fits the first group too, with a check against the trial balance before anyone uses it. Anything that posts to the books is a person's act, and AI may propose.

Pax8's US data shows why this matters. Among the US businesses that had built their own AI solutions, 36% had a formal, documented process for reviewing that AI, and 41% had a consistent but undocumented one.⁴ An undocumented review is not a sign-off. For finance work, CPA Ontario's Rule 205 also says a CPA must not be associated with false or misleading documents, which is the standard financial reporting is held to.⁸ See AI for finance teams.

What the three share. The same four questions, different groupings. This is not a template to copy. It is a method to run on your own tasks. The COO or head of operations keeps the list: which tasks sit in which group, and what happens when a task does not fit any of them.

05. What a real sign-off looks like

Why reminders fail. A policy line that says "review carefully" is not a control. Parasuraman and Manzey's review in Human Factors in 2010 found that automation bias occurs in naive and expert participants alike, cannot be prevented by training or instructions, affects individuals and teams, and produces both omission and commission errors when decision aids are imperfect.¹⁴

The EU AI Act's Article 14 names this tendency and asks that the people overseeing a system stay aware of it.¹¹ If a person reviews AI output but has no reason or room to disagree, the sign-off is a rubber stamp: a signature with no check behind it. The term is not from a source. The way to avoid it is design, not reminders.

Judge first, then look. In two experiments reported in Cognitive Research: Principles and Implications in 2024, Matute and colleagues used a simulated legal decision task. People who gave their own judgment before seeing an incorrect AI assessment ended up closer to the right answer than people who saw the AI's assessment first, and correct AI support helped less than it appears.¹⁵ It is an experiment, not a business study. Read as a practice, not as the study's conclusion: the reviewer writes their own number, answer or decision before opening the AI's.

Five parts of a sign-off.

1. A named person, not "the team."
2. What they check, written down per task.
3. A record of what was checked, which tool was used and whether the output was changed. The Government of Canada's guide suggests saying which system and version you used, whether and to what extent you reviewed, validated or modified the output, and telling your manager.⁶
4. Time to do it. In the Pax8 survey, 55% of decision-makers spend three or more hours a week reviewing, editing or fact-checking AI-generated work, so budget for it.⁴
5. The authority to say no. Article 14(4) describes the ability to decide not to use a system, to override or reverse its output, and to stop it.¹¹ A sign-off that cannot say no is a stamp.

For the group where AI runs. Where AI runs on its own, the check is a sample, chosen by the reviewer and not by the AI, with a stop rule written in advance. Article 14(4)(a) speaks of monitoring a system's operation to detect anomalies and unexpected performance.¹¹ The sample and the stop rule are a method, not a standard.

In our own work. Anything ambiguous is flagged and not written, and we log a known bias where the next person will see it, which is the lesson in a confidence score is not calibration. The review step is where that log gets read.

On NetSuite? Houseblend's NetSuite practice is separate and applies the same principle, with a person signing off, in its agentic delivery pipeline.

06. The one-page policy, and who decides what

What goes on the page. Seven lines, in this order.

1. The purpose, in one sentence.
2. The three groups, each with a list of your own tasks.
3. The data rule, meaning what may never be entered and where data may go, drawn from the same guidance as question 1.⁵⁸
4. Which tools are approved, by category and by who holds the licence, not by brand.
5. The sign-off standard from section 05.
6. Who to ask when someone is unsure.
7. A review date and a trigger. The Algorithmic Impact Assessment requires published assessments to be reviewed and updated on a schedule and when a system's functionality or scope of use changes, and the Government of Canada's guide describes itself as a living document updated regularly.¹⁰⁶

Add one line on consequences, decided in advance. In the first survey, 53% of people who used unauthorized tools were given informal feedback to stop, and 48% faced formal consequences, so an unwritten consequence is an inconsistent one.¹

Who decides what.

• The owner or CEO sets the line on what leaves the building and applies it to leadership too, since 81% of people in that survey saw a double standard. They are also the person the board asks.
• The COO or head of operations owns where in the workflow the sign-off happens, and the time budget for it.
• The CFO or controller owns anything touching money, the books or regulated data, and the check against trusted numbers, as set out in the data readiness checks.

In a smaller leadership team one person may hold all three roles, and that works too.

The bottleneck trap. If every sign-off routes to one senior person, you have rebuilt the bottleneck you were trying to remove. Reviewers in the first group should be the people who own the task. The third group goes to the senior person. That is reasoning, not a finding.

How Houseblend helps. The AI policy and governance service sits in the Plan stage and is described on the AI policy and governance page. It sorts your tasks into what AI drafts, what it runs and what it never touches alone, and it includes a security and privacy review. A person at your company stays accountable for the result.

Where this fits in the series. This is the third article in our AI adoption series. The first, our introduction to AI adoption, explains Plan, Build and Run. The second, is your data ready for AI, gives five checks on the numbers behind an AI answer.

Where to read next. What a company knowledge assistant is covers the Build service where AI answers only from your own documents, and where the sign-off rules above apply to every answer.

This content may be AI-generated and may contain inaccuracies. It is not professional, legal, or financial advice. Verify independently before relying on it.