Minimum viable AI: how to use AI at work when you don't want to

I write a lot about AI and use it constantly.

So I am probably the wrong person to tell somebody who hates AI that they should learn to love it.

I don’t think they should have to.

I wrote recently about the strange position employees are being put in. Increasingly, businesses are encouraging people to use AI, experiment with it, increase productivity and work out how it fits into their jobs. At the same time, much of the public conversation around AI tells them that using it is lazy, environmentally irresponsible, professionally dangerous, creatively bankrupt or some combination of all four.

Most employees aren’t spending their evenings comparing models, testing workflows or thinking about where human judgement belongs.

They just have a job.

So what do you do if your employer wants you to use AI and you would rather have nothing to do with it?

I think there is a useful position between enthusiastic adoption and outright refusal.

Call it minimum viable AI.

Learn enough to work competently in an organisation where AI exists. Use it where there is a clear reason. Know where you won’t use it. Keep responsibility for your work. And don’t pretend enthusiasm is a professional requirement.

That may become one of the more useful forms of AI literacy.

You are allowed to dislike this

Start there.

There are perfectly reasonable reasons to be sceptical of generative AI.

You might worry about what it does to jobs. You might dislike the way creative work has been used to train models. You may be concerned about energy use, misinformation, privacy, surveillance or employers using productivity gains mainly to demand more production.

You might simply think a lot of AI-generated material is terrible.

Those positions do not make somebody technologically illiterate.

Acas found in 2025 that 26% of British workers surveyed were most concerned that workplace AI would result in job losses. Others named errors and a lack of regulation. The concern is hardly fringe.

The mistake is turning the workplace question into a referendum on whether AI is good.

Your employer probably doesn’t require you to believe Microsoft Excel was a positive development for civilisation. It requires you to know enough about spreadsheets to do the parts of your job that depend on them.

AI may be heading towards a similar, although much messier, position.

You can object to the technology while recognising that professional competence may increasingly require understanding it.

Those things can coexist.

First work out what “use AI” actually means

This sounds obvious, but I suspect many organisations haven’t answered it.

There is a huge difference between:

• AI tools are available if you find them useful.
• You are encouraged to experiment.
• Certain workflows now include an AI component.
• Your performance is expected to improve because AI is available.
• AI use is effectively mandatory.

All of these can be described internally as “AI adoption.”

They place completely different obligations on an employee.

So before reluctantly reorganising your whole working life around Copilot, ChatGPT or whatever else has appeared in the company software stack, work out what is actually expected.

Which tools are approved? Which tasks are they intended for? What information can go into them? Is use optional? What remains your responsibility? Is anybody measuring adoption, output or productivity differently because AI is now available?

If nobody can answer those questions, that itself is useful information.

The ambiguity belongs partly to the organisation, not to you.

Learn the minimum before deciding where your maximum is

Outright refusal has a problem.

It makes it difficult to distinguish a principled objection from an uninformed one.

You don’t need to become a prompt engineer. You do need enough familiarity to understand what the system does, what it gets wrong and what risks arise when you use it.

The government’s AI skills for the workforce guidance is interesting here because its definition of workforce capability is much broader than knowing how to type prompts. It covers technical skills, responsible and ethical use, and non-technical skills, and says people need to understand when AI should and should not be used, with confidentiality, consent, data protection, transparency and human oversight built into everyday practice.

That is a much healthier definition of AI literacy than “uses AI a lot.”

A person who knows when not to use AI may be more AI-literate than somebody pumping every task through it.

So learn enough to make that judgement.

Then stop wherever the additional value stops.

Start with work you can easily reverse

If you dislike AI, don’t begin by handing it the part of your job you care most about.

Start somewhere boring.

Ask it to reorganise notes. Summarise material you already understand. Suggest alternative structures. Turn a rough list into categories. Help retrieve something from an approved internal system. Generate questions you can decide whether to use.

Then look at what happened.

Was it useful? Did it save time? Did it introduce errors? Did checking the result take longer than doing the work yourself? Did it remove a tedious part of the job or a part you actually valued?

There is no rule saying every successful AI experiment has to become permanent.

This is where I think some corporate adoption programmes get the psychology backwards. They treat experimentation as a route towards increased use.

It can also be a route towards justified non-use.

“I tried it three times on this task and the checking cost outweighed the saving” is a far stronger position than “I don’t want AI anywhere near my work.”

The first is evidence.

Keep the judgement you would need if the AI disappeared

This is the Editorial Intelligence part of the argument.

My own working principle is editorial judgement before AI scale. AI can support discovery, synthesis, drafting and reuse, but a person still has to decide what matters, what is credible and what should survive.

That matters just as much to somebody reluctantly using AI as it does to an enthusiast.

If AI writes the first version of every report you produce, you still need enough reporting skill to recognise when the report is wrong.

If it summarises research, you still need to understand the evidence well enough to recognise when the summary has flattened the important qualification.

If it writes code, somebody needs enough technical understanding to know whether the code works.

If it prepares a recommendation, somebody has to own the decision.

This is where “human in the loop” becomes too easy a phrase.

A human being sitting somewhere in a workflow does not automatically provide useful oversight.

They need enough capability, time, authority and confidence to disagree with the system.

That skill is worth protecting.

Decide your boundaries before convenience decides them for you

AI has an unusual characteristic: once it becomes convenient, the boundary tends to move.

You use it for something trivial. That works. Then something slightly more important. That works too. Soon the tool is involved in work you would previously have thought required much more care.

So decide some boundaries while you are still thinking deliberately.

What information will you never enter? Which decisions will always require you to inspect the underlying evidence? Where do you need another human? Which parts of your work do you deliberately want to keep practising yourself? Which uses would you want disclosed if you were on the receiving end?

Your answers do not have to match mine.

They probably shouldn’t.

Different professions carry different obligations, and different people value different parts of their work.

The important thing is that the boundary exists because somebody chose it rather than because the software made crossing it frictionless.

If you object, object specifically

There is another practical problem with saying “I hate AI.”

Organisations can’t do very much with it.

They can disagree, acknowledge the sentiment or classify you as resistant to change.

A specific objection is much more powerful.

“This customer material cannot go into that system under our data policy.” “The summary repeatedly removed qualifications that matter to the decision.” “Using this for the task saves ten minutes and adds twenty minutes of checking.” “We are measuring increased output but not whether the additional output creates value.” “This part of the workflow requires professional judgement and nobody has defined who remains accountable for the decision.”

Now you have something that can be examined.

The TUC’s July 2026 report on AI at work makes a related point at organisational level. It argues that AI adoption is widespread but shallow, and that workers are routinely excluded from decisions about how it’s introduced. Whatever your view of the union’s wider policy recommendations, the basic point is difficult to dismiss: adoption works differently when the people doing the work have some influence over how the technology enters it.

Employees know things implementation teams don’t.

Especially about all the awkward exceptions that make a supposedly simple workflow difficult.

Your employer has responsibilities too

This is the bit I think gets lost when businesses talk about becoming AI-first.

Buying licences is the easy part.

Telling people to experiment is also fairly easy.

What happens next is harder.

If an organisation wants employees to use AI, it should be able to explain:

• which tools are approved
• what data can be used
• where AI use is encouraged
• where it is inappropriate
• what employees remain accountable for
• what needs human review
• how quality will be measured
• whether AI use changes performance expectations
• what training is available
• how employees can challenge an AI-enabled process that is not working

That isn’t an exotic anti-AI position. It’s remarkably close to the direction of current mainstream guidance, which recommends baseline training before workplace AI use in areas involving organisational data, confidential material, regulated activity, safety or professional judgement — and finds that many organisations are still developing AI capability informally, through trial and error and peer support, rather than structured training.

Acas similarly recommends clear workplace AI policies and open conversations with employees, treating adoption as a genuine two-way discussion in which affected staff can raise concerns and have them considered.

That makes “just use AI” look increasingly inadequate.

If an employer receives the productivity benefit, it cannot outsource all the uncertainty to individual employees.

Complete refusal still has a cost

There is an uncomfortable part of this argument for people who want nothing to do with AI.

The technology isn’t waiting for consensus.

The ONS found in January 2026 that a quarter of UK businesses were using AI in some form — rising to 44% among large businesses with 250 or more employees — while the government is working towards giving 10 million UK workers foundation AI skills by 2030 through its expanded AI Skills Boost programme.

You can think the speed of that adoption is reckless and still recognise what it means professionally.

If tools become embedded in the ordinary systems used to perform your job, never learning how they work eventually stops being a statement about AI and starts becoming a limitation on what work you can do.

I don’t think telling people otherwise does them any favours.

The useful distinction is between adoption and surrender.

You may need to learn a system. You don’t have to hand it every task.

You may need to use AI where your role genuinely changes around it. You don’t have to describe every change as progress.

You can benefit from something while remaining critical of it.

And you can refuse a particular use for a reason that survives scrutiny.

Minimum viable AI

So this is the version I would suggest to somebody who hates AI but works somewhere increasingly determined to use it.

Learn what your organisation actually requires. Understand one approved tool well enough to know its capabilities and failure modes. Use it first on low-consequence, reversible work. Check whether the claimed benefit appears in your actual job. Keep the underlying skills necessary to challenge its output. Set boundaries around data, judgement and work you deliberately want to keep human. Raise objections with evidence. Ask your employer to define the rules it is asking you to operate within. And remain responsible for whatever leaves your hands.

That’s enough.

You don’t need an AI personal brand. You don’t need five models open. You don’t need to spend your weekend learning prompting tricks. You definitely don’t need to pretend to be excited.

The goal isn’t maximum AI. It is enough understanding to make deliberate choices about a technology that is increasingly becoming part of work.

Perhaps that is the bargain organisations should be offering employees too.

We will give you the tools, training, boundaries and right to question how they are used. You give us informed professional judgement rather than either blind adoption or blind refusal.

That seems a more mature definition of an AI-first workplace than simply asking everybody to use more AI.