AI for Small Business 10 min read

When to use AI in your business, and where it stops paying

You have already found the easy yes. The invoice reminder writes itself. The meeting notes get tidied in seconds. The first pass at a proposal comes back in a shape you can argue with, which beats the blank page you have been avoiding since Thursday.

What you have not settled is where it stops.

Most advice on when to use AI in your business arrives in one of two flavours. Use it everywhere, because the businesses that hesitate will be gone by next year. Or keep it away from anything that matters, because clients can always tell. Neither of those is any use at four o’clock on a Tuesday with a client email open in front of you and a decision to make about who writes it.

There is a line. It sits in a more specific place than most people assume, and it has research underneath it rather than instinct.

In this article

Two studies on AI and trust that reach opposite conclusions

Oliver Schilke and Martin Reimann ran thirteen experiments on what happens when someone discloses that they used AI. Supervisors and subordinates. Professors. Analysts, creatives, investment funds. Across all of it, the people who disclosed were trusted less than the people who did not.

The mechanism the authors identify is legitimacy. Disclosure signals a departure from what people expect of a person in that role, and the departure sets off an alarm of its own. The effect held above and beyond general suspicion of algorithms, across different framings of the disclosure, and whether disclosing was voluntary or required. It softened among people who already held favourable views of the technology and rated AI as accurate. It did not disappear.

One detail from that paper matters later on. Disclosing was penalised, and being exposed by somebody else was penalised harder.

Now the second study, which appears to contradict the first.

Researchers put 1,637 people through incentivised trust games. You have to persuade a stranger to hand over real money on the strength of a written message. Half the participants had AI writing assistance available, half wrote unaided. Some had that assistance disclosed to the recipient, some did not. The trust effect was close to nothing across every condition, measured both by what people actually did with their money and by what they said afterwards. The assisted writers reached the same result in less time. Linguistic analysis found their messages scored slightly lower on authenticity while scoring higher on warmth, complexity and confidence.

Thirteen experiments finding a consistent penalty. Two preregistered experiments with a large sample finding almost none. Both are competent, which means they were measuring situations that differ in some way that matters.

What the trust penalty attaches to

Look at what the Schilke and Reimann scenarios share.

A professor marking your essay. A candidate’s application. A founder in front of investors. In each case the recipient is reading the document as evidence about the person who produced it. The mark tells you what your professor made of your argument. The application tells a hiring manager who this candidate is. The pitch tells an investor whether this founder has judgement worth backing.

AI involvement leaves the document intact. What it alters is the question of what the document proves.

The trust game runs on different rules. A stranger wants you to send money in an exchange you will never repeat. Their character has no bearing on anything, because there is no relationship for it to have a bearing on. The message is a mechanism. Whether a machine helped shape it makes no difference to the job it does.

The researchers say as much themselves. They describe their setup as emotionally neutral and their participants as strangers meeting once, and they draw a line between the kind of trust they measured, the reasoned sort that operates in formal and transactional exchanges, and the kind they did not, the sort that develops between people with a history. Their own conclusion limits the finding to one-shot transactional interactions and flags that emotionally rich or higher-stakes relationships may behave differently.

There is a second candidate explanation worth putting on the table. In the trust game the disclosure was imposed by the system rather than chosen by the writer, and the authors suggest that a disclosure everyone knows was mandatory may read as more legitimate than one a person volunteered. That would fit the legitimacy mechanism rather than contradict it.

Either way, the useful question about any task in your business is what the output is being read as proof of.

An invoice reminder proves nothing about you. Nobody has ever received one and formed a view of the sender. Booking confirmations, shipping notifications, receipts, calendar invites, the automated follow-up that fires when a form gets submitted. These are mechanisms, understood as mechanisms by everyone involved, and they were never carrying a claim about your character in the first place.

Then there is the reply to a client who has just told you something difficult. The message explaining why a project has run over. The note to a customer who has been with you six years. The email you send when something has gone wrong and you are the one who has to say so.

Those documents are doing a second job underneath the obvious one. The obvious job is conveying information. The second is demonstrating that a specific person, with specific judgement, paid attention to this specific situation. The second job is where all their weight comes from, and it is the one that empties out the moment it gets handed over, because handing it over is the opposite of what it is claiming.

There is a further distinction worth holding onto, and it has nothing to do with writing.

Plenty of AI use in a business never produces anything a client sees. Pulling structure out of a messy spreadsheet. Sorting enquiries by type. Summarising a call recording so you can find the bit you need. Drafting something you then take apart and rebuild. In all of those, the output lands in front of you, and you are the last stop before anything reaches another person. Your judgement is still in the chain. Nothing has been claimed on your behalf.

The category changes when the output goes out unattended. The moment a system sends something to a client without you having read it, a claim about your attention has been made in your absence. That may be entirely fine, depending on what the message was carrying. An automated receipt carries nothing. A response to a complaint carries everything. The question is the same question, applied at the point of sending rather than the point of drafting. It is also the point at which the automation you built and never switched on starts to make sense.

This holds even when nobody finds out. The penalty Schilke and Reimann measured only fires when the recipient knows, and the version where somebody else tells them costs more than the version where you do. Which leaves the option of saying nothing and hoping. But you know. What you know changes how you hold the relationship, which is slower and more corrosive than being caught, and it is the ground covered in the piece on why you will not let automation near your clients.

What Google actually says about AI content

The other version of this question is about search. Will Google punish you for publishing AI writing?

Google’s own documentation answers it more clearly than most of the commentary built on top of it.

The guidance on generative AI content says the technology is useful for researching a topic and for adding structure to original content, and that using it to generate many pages without adding value for users may fall foul of the spam policy on scaled content abuse. The spam policy defines that abuse as generating many pages primarily to manipulate rankings rather than to help people, and states that it applies no matter how the content was created.

No matter how it was created. Volume with nothing in it gets treated the same whether a person produced it or a model did. Method was never the criterion.

Which puts the search question on the same line as the trust question. What fails is publishing that makes no claim of its own. Twenty articles assembled from what has already been said about your subject will sink whether you wrote them by hand across three months or produced them in an afternoon. They fail with search for the reason they fail with a reader, which is that there is nobody in them.

That conclusion happens to suit the argument I am making, so it is worth being explicit that it comes from Google’s published policy rather than from anyone’s theory about what search engines ought to reward.

The writing you most want help with

Which brings the problem home.

The place solopreneurs reach for AI hardest is their own website. The About page. The services descriptions. The positioning line that has been sitting in a draft since March. Those are the documents that consist entirely of a claim about who you are, and they are the first ones to get handed over.

Laziness has nothing to do with it. Writing about yourself is the hardest writing there is, for reasons examined properly elsewhere on this site. You can talk about what you do for an hour over coffee and go blank in front of a text field, and that gap is a well-documented feature of how self-knowledge works rather than a personal failing.

AI happens to be extremely fluent at exactly this. Ask for an About page for a business like yours and something competent, warm and readable comes back. There will be nothing wrong with it.

There will also be nothing of you in it. A model can only write from what it is given, and if what you can give it is a general sense that you help people and you take it seriously, it fills the remaining space with the average of everyone in your category. What comes back reads like your competitors because it was assembled from the same material your competitors were assembled from. The borrowed voice problem, arriving by a faster route.

The failure here has nothing to do with detection. Nobody is going to catch you. The failure is that what you actually needed was to reach a position on what you think, and the sitting there unable to say it was the process that would have taken you there. Removing the difficulty removed the outcome attached to it.

That pattern is worth watching for beyond writing. Friction comes in two kinds. Some of it is pure cost, the fifteen minutes you spend reformatting a spreadsheet, and removing it is a straightforward gain. Some of it is the mechanism by which something gets formed, and removing it leaves you with a faster route to a place you did not want to arrive at. Deciding what your business stands for runs on the second kind. There is no version of it that happens on your behalf. If you want to find out which one you are sitting on, the Brand Mirror takes five minutes.

The boundary itself keeps moving. Automated appointment confirmations once read as impersonal and now read as normal. Most of that movement runs in one direction, towards more automation being unremarkable, and there is no reason to expect it to stop.

Common questions

Does Google penalise AI-generated content?

No. The spam policy targets scaled content abuse, defined as generating many pages primarily to manipulate rankings rather than to help people, and it states this applies no matter how the content was created. Volume with nothing in it fails whether a person or a model produced it.

When should I not use AI in my business?

When the output is being read as evidence about you. Replies to a client who has raised something difficult, explanations when a project has gone wrong, your About page, your positioning. Those messages carry a second job underneath the information, which is demonstrating that a specific person with specific judgement paid attention to a specific situation.

Is it safe to use AI for admin and internal work?

Yes, where the output reaches you before it reaches anyone else. Cleaning up a messy spreadsheet, sorting enquiries, summarising a call, producing a draft you then take apart. Your judgement stays in the chain. The category changes when a system sends something to a client unattended.

Does disclosing that you used AI damage trust?

Thirteen experiments found that people who disclose AI use are trusted less than people who say nothing, and that the penalty runs through perceived legitimacy rather than any judgement about output quality. Being exposed by a third party costs more again. The workable response is to keep AI away from anything where disclosure would matter, rather than to use it and stay quiet.

What holds still underneath the movement is the test. Somewhere in your business there is a document whose entire job is telling people who you are. You already know which one it is, because it is the one you keep not finishing.

Keep reading

More on AI for Small Business