What should you say when a client asks whether you use AI?

Tell them yes, and tell them before they ask. Then demonstrate it: name what you use it for, what you refuse to use it for, and where it fails. Underneath the question the client is working out whether you understand these tools better than they do, and whether that difference is worth paying for.

Why does this question feel loaded?

Because a lot of people in professional services are still faintly embarrassed about it, and the client can hear the embarrassment before they can hear the answer.

I am not embarrassed. I think understanding how these tools actually work, and getting very good at using them, is one of the more important things a consultant can do right now. The awkwardness in that conversation almost always comes from a professional who has not decided what he thinks, being asked to defend a position he never took.

So take the position first, in private, before anyone asks you in public.

When should the conversation happen?

Early. Possibly on the discovery call, before you even know whether this person is a real lead.

That timing sounds aggressive and it does the opposite of what you would expect. Bringing it up first means you are the one framing it, and you get to frame it as competence instead of defensiveness. Waiting means it arrives as a challenge halfway through a proposal, usually phrased as do we even need you anymore, and now you are answering from your heels.

Share a couple of specific ways AI is working in your own practice. Not the brochure version. The real one: this is what I run through it, this is what I would never run through it, this is where it produced something confident and wrong last month.

Specificity is the whole demonstration. Anyone can say they use AI. Very few people can describe, concretely, where it fails them, and that description is the thing a prospect cannot get from a competitor's website.

How do you say you are better at it without sounding arrogant?

Plainly, and accept that it will be a little uncomfortable.

There is still a widespread idea that anybody can do AI, and it is wrong. These tools take a particular set of skills and a body of accumulated knowledge about where they break. Your client needs to understand that not everyone uses AI equally, and that the difference between a competent operator and an average one is large and consequential for their business.

That sentence is awkward to say about yourself. Say it anyway. You have to be confident in your own ability, because the alternative is competing on price with someone who is going to do the work badly and quickly.

And then let some of them go. If a prospect decides the poor AI version is good enough, and there are real consequences to using these tools badly that they are willing to live with, that is a legitimate choice. You will watch work you could have done get done worse. Fine. They did not need what you do in the way somebody else needs it. A firm that cannot name the client it is willing to lose does not really have a position, it has a pipeline.

Why does agreement from a machine feel like a second opinion?

Worth understanding, because it shapes what your client brings you.

The behaviour has a name and a research literature. It is called sycophancy: a model's tendency to produce the response a user will approve of over the response that is true. In October 2023 a team at Anthropic published a study of five state-of-the-art AI assistants across four free-form generation tasks and found it in all five. The mechanism is the part that matters. When the researchers examined the human preference data these systems are trained on, one of the strongest predictors of a response scoring well was whether it matched the reader's existing view. Both human raters and the preference models trained on them preferred a well-written sycophantic answer to a correct one a meaningful share of the time.

It showed up at scale in April 2025. OpenAI shipped a GPT-4o update on 25 April, watched it begin praising more or less any idea a user brought it, and rolled it back on 29 April, four days later. Their own account is precise about the cause: they had weighted short-term feedback too heavily and the model had skewed toward responses that were supportive but disingenuous.

So when a client asks a model whether their plan is good, they are asking a system optimised to agree with the person holding it. Useful to know, and a poor thing to be smug about. What it gives you is a job: be the person in the room who can say this part is already right, leave it alone, which is a sentence the machine will almost never produce.

What happens when you skip the buy-in?

Something I got wrong, and it cost eighteen months.

For two years of my time in ministry I was the resource development pastor at a church running around 1,500 kids, birth through grade eight. Every age group had its own director, and each of them had gone out and found teaching material they liked. Some of it did not line up with the church's own position. Almost none of it lined up with the age group above or below, so a child moving up a level got a different framework every couple of years.

I proposed writing all of the curriculum myself, for every ministry that used one. Leadership agreed, gave me a clear mandate, and I went away and did it. Writing and teaching are what I am best at, which is why I was asked.

A year and a half later I found out that several of the directors simply were not using it.

I was angry, and not only about the work. They had gone against a direction we had agreed on. But the honest post-mortem is less flattering to me than that. Large ministries and large nonprofits build silos between departments, which is a real and well-known problem, and I knew it. I also went in a little arrogant. I knew I could write it better than anyone else there, so I put my head down and wrote, and I never brought the directors to the table, never trained them properly on using it, never sat in their rooms watching how it actually landed, never talked to their volunteers, and never collected the feedback that would have told me any of this at month three instead of month eighteen.

The mandate was explicit. It was signed off at the top. It still did not hold, because the people who had to live with it were never really part of the decision.

That maps directly onto client work. A clause the founder signs does not govern the marketing lead who never agreed to it. If the person who will actually use the tools was not in the conversation about how you work together, you have an agreement with one person and a negotiation with everyone else.

Putting it to work

Putting it to work

Three things, and the first one takes twenty minutes.

Write your AI answer down before you need it. Three sentences: what you use it for, what you refuse to use it for, and one specific instance where it gave you something confident and wrong. The third sentence is the one that convinces people, and it is the one nobody prepares.

Move it to the front of the conversation. Put it in the discovery call, in your own words, before the prospect raises it. Once they have brought it up first, you are explaining yourself on their terms.

Get the actual user in the room. Before the engagement starts, ask who on their side will be doing the work day to day, and get that person on a call. Ask them what they are already using, what they have tried, and what they think should be automated. You are looking for the gap between what the founder signed and what the operator believes, because that gap is where the relitigating happens six months later.

One test worth running on yourself: can you name the client you would be willing to lose over this? If some prospect is going to choose the cheap, poorly-executed version and you have no one you would let walk, you are about to compete on price against a machine. That is not a fight anybody wins.

Sources

• Sharma, M. et al., "Towards Understanding Sycophancy in Language Models", arXiv:2310.13548, 2023. Five state-of-the-art AI assistants across four free-form generation tasks; human preference data rewards agreement with the reader's stated view.
• OpenAI, "Sycophancy in GPT-4o: what happened and what we're doing about it", 29 April 2025. Postmortem on the 25 April 2025 update and its rollback four days later.
• Van Dycke, "If AI can produce the deliverable, what are clients paying you for?" — the companion argument on diagnosis as the scarce part of professional work.

Frequently asked

Should I disclose that I use AI in client work?

Yes, and lead with it instead of waiting to be asked. Disclosure framed as competence lands completely differently from disclosure extracted under questioning. Tell them what you use it for, what you keep away from it, and where it has failed you.

What if the client thinks that means they can do it themselves?

Some will, and a few of them are right. For most, the gap between using these tools occasionally and using them well is large, and your job is to make that gap visible with specifics. Arguing about it will not do the work. A client who genuinely does not need the difference was never your client.

Isn't saying "I'm better at AI than you" arrogant?

It is uncomfortable, which is not the same thing. You are stating a difference in skill, the way you would about any other part of your craft. Say it plainly, back it with something concrete, and let them decide whether it is worth paying for.

How do I stop a client's AI suggestions from becoming endless free revisions?

Decide the revision terms in the agreement, and get the person who will actually be using the tools into that conversation. An agreement the founder signs does not bind the operator who was never in the room, which is where most of these disputes actually start.

Should I be using AI in my own practice?

Almost certainly. The risk sits in failing to notice when a task has moved outside what the tool is reliable at, because the output reads as equally confident on both sides of that line.

Ready to look at the architecture honestly?

If the AI conversation with prospects still feels like something to survive, the problem is usually a position you have not written down rather than a script you are missing. Twenty minutes, and we will tell you what we see. You can read how we work through the four frameworks, or what an engagement involves, first.

Book the conversation →