Rolling out AI to the people who do the work

You have a virtual assistant who handles your inbox, a bookkeeper who invoices monthly, and two contractors who do the work you sell. You have been using Claude yourself for six months and it is obviously faster for first drafts. So you mention, in the Monday message, that you are going to start using AI for the first pass on client copy. Nobody objects. Nobody asks a question. Three weeks later the contractor who writes that copy has gone quiet on new work, your VA has not opened the seat you paid for, and someone has started sending you longer, more defensive status updates than they used to.

Nothing broke. What happened is that you announced a change to how people earn money, and they started responding to it immediately, while you were still thinking of it as a tool you were evaluating. This guide is about the gap between those two clocks and how to close it. It is written for someone with between 1 and about 10 people around them, employees or contractors or both. It is not for a company with an HR function, a works council or a union agreement, where the sequence is set by law and by a contract and you need counsel rather than a guide. It is also not for a solo operator with nobody else touching the work, who has no rollout problem, only a tooling one.

The rollout starts the day you mention it, not the day you deploy

The clearest recent illustration is not from a small business. In July 2026, Hyundai’s South Korean union staged 3 days of partial strikes, from 13 to 15 July, with day and night shifts each ending 2 hours early, after more than 34,000 of roughly 40,000 union members authorised industrial action [8]. Alongside pay demands, the union sought “job and income guarantees tied to AI and factory automation”, and warned of an “employment shock” if robots entered workplaces without a labour-management agreement [8].

The Atlas humanoid robot at the centre of it was developed by Hyundai-owned Boston Dynamics and is scheduled to begin working at the group’s Georgia factory in 2028 [8]. No deployment date has been announced for South Korean plants, and no detailed timeline or list of factories has been released [8]. What the union was responding to, then, was a published intention: more than 25,000 Atlas robots across Hyundai and Kia facilities, and annual production capacity of 30,000 units by 2028 [8].

Strip out the scale and the machinery and the mechanism is the one you are dealing with. People do not wait for a deployment to respond to it. They respond to information about a deployment, and they respond on their own timeline, using whatever leverage they have. At 40,000 people that leverage is a work stoppage. At 4 people it is quieter and slower and it costs you more per head: a contractor who takes another retainer because yours looks like it is shrinking, a VA who stops proposing improvements because improvements now sound like arguments for replacing her, an assistant who does the work by hand and does not tell you.

The practical consequence is that the announcement is the intervention. By the time you are choosing between tools, the people around you have already priced in a version of your plan, and it is usually a worse version than the one you have, because the absence of terms gets filled with the worst plausible reading.

Name the decisions the tool will not make

The fear underneath most quiet resistance is not that the AI will do the work badly. It is that the AI will decide something. Whether the retainer renews. Whether the hours get cut. Whether this person is still needed in March. If you never say the tool is not making those calls, you have left the question open, and people answer open questions pessimistically.

You can close it with something you already agreed to. Anthropic’s Usage Policy, effective 15 September 2025, requires that when the product is used for advice, recommendations or subjective decision-making that directly affects individuals, “a qualified professional in that field must review the content or decision prior to dissemination or finalization”, and it names employment among the fields where that applies [4]. OpenAI’s usage policies, effective 29 October 2025, prohibit the “automation of high-stakes decisions in sensitive areas without human review”, with employment on the list of sensitive areas [5]. Two vendors, writing separately, drew the same line in the same place.

So the sentence you can say honestly is that the decision about someone’s work stays with you, that you personally read anything the model produces about a person’s performance or pay, and that using the tool otherwise would breach the terms you signed. It is more convincing than reassurance because it is checkable. It is also a constraint on you, which is the part that makes it worth saying out loud.

In the EU, telling people first is an obligation, not a courtesy

If the EU AI Act reaches you, some of this stops being a choice. Annex III makes high-risk any AI system “intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics or to monitor and evaluate the performance and behaviour of persons in such relationships”, and any system used for recruitment or selection, including filtering job applications and evaluating candidates [2].

Where that applies, Article 26(7) is direct: “Before putting into service or using a high-risk AI system at the workplace, deployers who are employers shall inform workers’ representatives and the affected workers that they will be subject to the use of the high-risk AI system” [1]. The same article requires that deployers “assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support”, and that they take measures to ensure they use the system in accordance with the provider’s instructions for use [1].

Most of what a small team does with Claude or ChatGPT is not high-risk under Annex III. Drafting copy is not evaluating a person. But the boundary is closer than people assume: a tool that allocates tasks based on individual behaviour, or that scores or monitors how well someone is performing, is inside point 4(b) whether you bought it as an HR system or assembled it from a spreadsheet and a prompt [2]. And if you are outside the EU entirely, notice-before-use is still the right default, for the reason the Hyundai case demonstrates rather than because a regulator says so.

Put the terms in the language of the actual worry

“We’re adopting AI to work smarter” is not an announcement. It is a mood. The announcement people need answers three things in their own units: whether scope changes, whether hours or rate change, and what happens to them if this works.

Answer them in writing, in whatever channel you already use, and keep it short. Say which specific tasks the model takes the first pass on. Say what each person keeps, and say it as a positive scope rather than a residual. Say whether their hours or retainer are changing in the next quarter, including when the honest answer is that you do not know yet, plus the date by which you will. Say who reviews the model’s output before it reaches a client, because that role is usually a promotion in disguise and nobody notices unless you name it. And say what happens to the time saved, because “we do more of the work we want” and “we buy fewer hours from you” are different plans and people can tell which one you are refusing to specify.

The written version matters more than the conversation. Spoken reassurance fades, and it is unquotable when a decision later goes the other way. A short written commitment is something you have to either honour or explicitly revise, which is a useful pressure on you and a real piece of security for them.

The rollout is a cost before it is a saving

Seats are the small part. Claude Team is $20 per seat per month billed annually or $25 billed monthly, and is sold for teams of 2 to 150; Pro is $17 per month on an annual subscription or $20 billed monthly [6]. Three seats is roughly the price of a modest subscription and none of the actual expense.

The actual expense is hours, and some of it you owe. Article 4 of the AI Act, applicable since 2 February 2025, requires providers and deployers to “take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf”, weighted by those people’s technical knowledge, experience, education and training and the context of use, while stopping short of requiring you to guarantee any specific level for any individual [3]. That covers your contractors as well as your employees, since they operate the system on your behalf.

At your size that means a scheduled hour where you sit with each person and work through 2 real tasks from their week, not a link to a prompt library. Budget a second hour a few weeks later, when they have hit the first failure and stopped trusting it. When a rollout produces nothing, the tool choice is rarely the first place to look. Look instead at whether anyone paid for the transfer: the owner already knew how to use it, assumed the skill was obvious, and left everyone else to work it out between other jobs.

Quiet non-adoption is the failure you will actually get

You are unlikely to face a strike. You are quite likely to face a seat nobody logs into, or its opposite, which is worse: the work being done in someone’s personal free account, on their own laptop, with your client’s material pasted into it.

That second one is a data problem with a specific shape. On Claude’s consumer plans, conversations are used for model training if the user has chosen to allow their chats to be used to improve Claude, if a conversation is flagged for safety review, or if they have explicitly opted in some other way, such as joining the Trusted Tester programme; feedback data is stored in Anthropic’s secured back-end for up to 5 years [7]. Nothing there is sinister, and none of it is a setting you control on an account you do not administer. That is the point. Shadow use moves your client’s material to a place where your answer to “where did this go” is a guess.

The fix is not a policy document. It is making the sanctioned account the path of least resistance, then asking, about 3 weeks in, one question per person: what did you try this on where it did not help. People will tell you about the failure long before they will tell you they are avoiding the tool, and the failure is usually the reason they are avoiding it.

What still goes wrong

The largest limit is that you cannot promise what you do not control. If your revenue drops in November, the commitment you wrote in September will be tested, and a written commitment you break is worse than one you never made, because it converts a business problem into a trust problem. The way out is to write terms at the resolution you can actually hold: a quarter rather than a year, a scope rather than a headcount, a review date rather than a guarantee. Vague promises feel generous when you write them and dishonest when they expire.

The second limit is that none of this helps if the plan really is to buy fewer hours. Announcements are not a technique for making that palatable, and people read an evasive rollout message faster than you would like. If a role is going to shrink, the useful version of this guide is the same sequence with different content: say it early, say the timeline, and let the person plan around it rather than discover it. That costs you something real, and it is still cheaper than the alternative, where you keep the information, they work it out anyway, and you spend the difference on trust you cannot buy back.

The third is that the legal reading here is deliberately narrow. Article 26(7) and Annex III bite only where the AI Act applies and where a system falls in a high-risk category, and most small-team drafting and summarising work does not [1][2]. Employment rules for AI use are moving in several jurisdictions, and a guide is not advice about yours. What travels regardless of jurisdiction is the mechanism: information about a change starts producing responses on the day it becomes available, and the only variable you control is whether it arrives with terms attached.