Where I Let AI Near a Practice, and Where I Do Not

If your voice is your product, the question of where AI belongs in your business is not a technical question. It is a question about whether the relationship survives contact with it.

I want to answer it properly, because the two available answers in public are both useless. One says AI will handle your marketing while you sleep. The other says a practitioner should never let it near anything. I do not believe either, and there is enough real research now to do better than believe.

Three findings that set the boundary

People cannot tell. Researchers at Cornell and Stanford ran six experiments with 4,600 participants across 7,600 self-presentation texts, in three settings where trust genuinely matters: job applications, online dating profiles, and Airbnb host profiles. Detection accuracy came in between 50% and 52%. A coin flip. Worse, the cues people used were wrong. They treated first-person pronouns, contractions and mentions of family or personal history as proof of a human author, and the researchers were then able to use those same cues to write AI text that was rated as human more reliably than genuinely human text was.

But they punish it anyway. Pew Research Center surveyed 5,023 US adults in June 2025 and asked how they would feel on discovering AI had been involved in something. The results lean negative across every scenario tested. 71% said they would like a political candidate less if they learned AI helped write a speech they had liked. A news article, 56%. A doctor's treatment recommendation, 45%. A song, 38%. Note that the speech in the question was one the respondent already liked. The judgment is not about quality. It is about discovery.

And telling them first does not help. This is the finding that surprised me. Schilke and Reimann ran thirteen preregistered experiments on what happens when someone voluntarily discloses that they used AI. Trust went down in every one. In the study closest to a practitioner's daily life, participants read a scheduling email from a colleague. When she mentioned using an AI tool to help manage her calendar and send scheduling emails, trust dropped measurably against the identical email with no disclosure. The mechanism they identify is perceived legitimacy rather than perceived competence.

Their own summary of it is worth sitting with: people who try to be trustworthy by transparently disclosing AI use end up trusted less for it. The one consolation in the data is that being caught by somebody else is worse than disclosing.

There is a fourth result that completes the picture and is genuinely unsettling. In an earlier experiment using Airbnb host profiles that were all, without exception, written by real humans, participants who were told the set was mixed rated the profiles they suspected of being AI as less trustworthy. The profiles were human. The suspicion alone did the damage.

What that adds up to

The naive reading is that AI is radioactive and should be kept out entirely. That reading fails, because the research is not about AI. It is about identity and warmth.

Look at what the punished scenarios have in common. A speech. A song. A painting. A doctor's recommendation. A colleague's personal email. Every one is a case where the point of the artifact is that a specific human meant it. When the machine writes those, something is genuinely lost, and the audience is right to mind.

Now look at what is missing from that list. Nobody has ever been betrayed by an automated calendar reminder. No client has felt their relationship cheapened because the booking confirmation was generated. The scheduling-email result is instructive precisely because it sits on the boundary: the task was administrative, but it arrived as a personal message from a named colleague, so it read as identity.

So the boundary is not the technology. It is whether the thing carries a claim of personal presence.

The line I actually work to

AI carries your message, in your own tone of voice. You set the standard it works to, and you choose how much of it runs without you.

That third clause is the one that matters and it is the one most vendors will not say out loud, because it means the answer differs per practitioner and per message type. Some people want every re-contact message reviewed before it leaves. Some want the routine ones running and the sensitive ones held. That is a setting, and it is theirs.

What does not vary is the source material. The words come from your body of work, your language, the way you actually explain things. That is not a nicety. Given the detection research, a system trained to write in a house style rather than yours would produce text that reads as competent and generic, which is precisely the signature people are scanning for.

And there is a category that never gets delegated, no matter the setting. The conversation where your reading of a situation is the work. The moment somebody discloses something difficult. The judgment call about whether this person is ready. Those are not tasks with an efficiency problem. They are the service.

What the professional bodies actually say

Worth clearing up, because it gets misquoted.

The ICF Code of Ethics, Standard 2.5, requires a coach to fulfill their ethical and legal obligations to clients "directly and through any technology systems I may utilize", and explicitly names artificial intelligence among those systems. It does not, contrary to a lot of confident blog posts, contain a general duty to disclose AI use. What it does is refuse to let technology be a place where responsibility gets laundered.

Separately, the ICF's AI Coaching Framework and Standards is strict about one specific thing. Element A.1.1 requires that an AI coaching system "shall indicate to the Client that it is not human prior to the provision of Coaching Services". That is a rule about systems that deliver coaching, not about a human coach whose follow-up is scheduled by software. The distinction is exactly the one I have drawn above, arrived at independently.

The American Psychological Association went further in a November 2025 health advisory, telling people not to use chatbots and wellness apps as a substitute for care from a qualified professional, and telling clinicians to ask their patients about AI use rather than wait to be told.

None of those bodies say keep it out. All of them say know where the human is.

The part I will not pretend about

Roughly one in five US businesses reported using AI in any business function as of May 2026, according to the Census Bureau's Business Trends and Outlook Survey, and among firms with four or fewer employees the rate is under 20% and has not moved significantly in six months.

So most solo practices are not doing this yet, and the ones that are have mostly not thought about where the line sits. I would rather you drew it deliberately than arrived at it by accident, and I would rather draw it with you than hand you a policy.

Sources: M. Jakesch, J. T. Hancock and M. Naaman, Human heuristics for AI-generated language are flawed, PNAS 120(11), 2023. Pew Research Center, From political speeches to songs, 17 September 2025, 5,023 US adults. O. Schilke and M. Reimann, The transparency dilemma: How AI disclosure erodes trust, Organizational Behavior and Human Decision Processes 188, 2025, thirteen preregistered experiments. M. Jakesch et al., AI-Mediated Communication, CHI 2019, reported by Cornell. ICF Code of Ethics and ICF AI Coaching Framework and Standards. APA health advisory, November 2025. US Census Bureau Business Trends and Outlook Survey, May 2026.