How to Use AI in Client Work Without Flattening the Relationship

Someone in every workshop asks whether using AI in my client work means my clients are getting less of me, and my answer is that it depends entirely on where you put the software, because if you put it between you and the client then yes, they are getting less of you and they can feel it, and if you put it around the outside of the relationship — on the preparation, the notes, the resources, the tracking and the follow-through that used to live in my head and regularly fell out of it — then the client gets more of the part they were paying for in the first place.

That distinction runs through everything below. AI prepares, organizes, drafts, remembers and assembles; I read what it produces, decide what it means, and handle the conversation, the challenge, the timing and the care, because those are the reasons a person hires a person.

My clients are not short on information — they can find information in nine seconds like everybody else, and most of them have a Drive folder full of it. What they need is somebody who remembers where we left off, notices the thing that keeps stopping them, and knows which next step belongs to their business rather than to businesses in general, which no model can supply, because the model was not there for the decisions they made before they ever met me.

I use AI every single day, client work included, and I review one hundred percent of what goes out under my name, which I mean literally rather than as a figure of speech. AI writes smoother and creamier and breezier than any of us do, and the sharp edges are the entire point of hiring one specific human being (mine are well documented, and several of them are load-bearing).

If you have not picked a first use case yet, start with the piece that opens this series, because the decisions in it — the pressure point, the described result, the human boundary — are what make any of the ten below work.

What is AI good at in client delivery, and what should it never touch?

AI is at its best in client work when it is assembling context and building materials for a human being to read, correct and sign off on: briefs, recaps, drafts, resource lists, progress summaries, and first versions of tools you already know how to teach. Every one of those has a finished shape you can recognize, source material you already own, and an obvious place for your eyes before anything reaches the client.

It is at its worst the moment it is allowed to decide something about a human being without you, which is a longer list than people expect — whether a client is stuck or stalling, whether a passing comment on a call became an assignment, whether the frustration she expressed at minute forty belongs in a permanent record.

So the rule I use in my own business is boring and it holds up: AI gets read access and draft duty, I get the send button. Almost everything in this article is built on read-only connections to sources I approved, with the requirement that it show me where each claim came from, and with no permission to message anyone, update a record, or take an action a client would notice.

I run most of this through Claude Cowork and Claude Code, where I have connected specific workflows to specific sources and told each one what it may read and what it may prepare. Paying for Claude does not hand the software your email, your Drive, your Trello board or your LinkedIn — those connections get configured deliberately, one at a time, and you decide the scope of each one. (I say this because I keep meeting people who assume the AI can already see everything, and then either panic or hand over far more than they meant to.)

What are the ten places AI fits into client work?

All ten sit on the same side of the line: AI prepares something, I review it, and the client experiences my judgment rather than the software’s. Eight of them I use or have built, and two I have not built yet and am telling you about anyway, because the pattern is the useful part and pretending everything on my list was finished would be its own kind of dishonesty.

1. Arrive at the call already caught up

Every morning, Claude sends me a client debrief, so I walk into a session with the relationship and the loose ends in front of me instead of trying to remember what happened three weeks ago while somebody is already talking. The workflow reads my connected email so I know what a client has asked for or shared, checks her folder in Google Drive for where we left the work, looks at my Trello board for anything in motion with my VA, and pulls approved LinkedIn activity when she has posted something we should talk about.

A brief that is worth reading has the purpose of this particular conversation at the top, then previous decisions, commitments made on both sides, work still open, recent requests, any deadline that is getting close, and a flag on anything that might need a careful conversation rather than a cheerful one.

This is the best place to start, if you are starting, because the agent can prepare the entire thing without making a single client-facing decision. Give it read-only access, limit it to sources you approved, and make it cite where each commitment came from and when — then read it yourself, because the sentence “she agreed to publish weekly” might have been an idea she floated out loud on a Tuesday and immediately talked herself out of.

2. Write the recap the way that particular client needs it

I have built separate Claude skills for each type of call I run, so the recap and action plan come back shaped for that client rather than forced through one all-purpose template that fits nobody particularly well. A VIOS client needs a different shape of recap than a long-term authority client does, and a mastermind needs something different again, so I have a client debrief skill, a WTIC debrief skill, a legacy debrief skill, and so on.

For an individual client, the recap usually leads with the decision she made, the commitments attached to it, and one private next step. A mastermind recap leads instead with the shared themes and the resources the group needs, and it has to be careful about what was said in that room that should stay in that room, which is a judgment the software cannot make for you.

Build one strong base process first, then write down what changes by client type — the purpose, the structure, the level of detail, the resources, the tone, and the material that must never appear in a shared document. AI is good at pulling decisions, commitments, questions and next actions out of a transcript; I still read every recap, because a possibility somebody mentioned is not homework, and a moment of frustration at minute forty is not a finding.

3. Rebuild one of your own workflows for a client

One of my favorite things about working this way is how fast I can take a workflow that already helps me and remake it for a client who needs the same kind of help in a different business. One of my authority clients was overwhelmed and scattered, so we picked up the agent I had already built for my own week — called, with no imagination whatsoever, my “Keep Viveka on Track” agent — and rebuilt it around her business, her priorities, her working style, her commitments and her definition of progress. She is thrilled, and I did not have to invent anything, because the thinking was already done.

The part that takes care is the copy. Start from the process that works for you, make a clean version, and strip out every reference, assumption, connection, private detail and inside joke that belongs only to you, because there is more of you in a working agent than you remember putting there.

Then ask the client what kind of support would help, and listen to the answer rather than assuming. One woman wants a short morning list and nothing else; another needs the agent to tell her that six things do not fit into one day and make her choose which three survive.

4. Turn a dead PDF into something a client can work inside

I can take an ordinary PDF worksheet, the kind I used to email to clients and hand out at webinars, and have Claude Code rebuild it as an interactive HTML worksheet that people can type into, work through, save or print. It can be gated or open, and I can send people to it with a link or a QR code on a slide, which is a far better experience than a fourth PDF slowly going grey in somebody’s Downloads folder.

It takes minutes rather than months, because the hard part was done years ago — the questions, the order, the exercise, the thinking behind why question three comes after question two. What I need from AI is the build, not the content, and that is the difference between using AI and building with it.

Before you send one to a client, test every field and every button on a phone, because that is where most of them will open it, and decide where the answers go before anybody types into it. Say on the page, in plain language, whether the answers save anywhere, so that nobody loses forty minutes of thinking to a closed tab.

5. Reshape a program you already teach for a new audience

I say yes to more webinars now, because I can take a program I developed for one client, add the parameters for a different room, and come out the other side with a presentation, a workbook and an interactive guide in a fraction of the time it used to take. This week alone I have three presentations going out: one on AI for entrepreneurs and two on LinkedIn, and the two LinkedIn sessions are for audiences so different that the examples, the language, the exercises and the call to action all have to change. Total prep was probably about an hour.

What makes that possible is that AI is starting from material I already built and tested in front of real people, and all it is doing is mapping it to a new audience and a new outcome. Starting from a blank prompt gets you a competent generic webinar, and nobody needs another one of those.

Give it the original program plus a short brief on the room: what these people already know, what they are stuck on, and what they should walk out able to do. Then check every slide, example, exercise, source and promise, because faster production only helps if the new version still fits the people sitting in front of you, and because AI will happily invent a statistic to make a slide land better.

6. Find resources, then check whether they say what they claim

Research for a client used to cost me half a day on Google, and now an agent can start that work the moment I have the idea, which has changed how often I bother to do it properly. For client work I tell it to search my own approved library first and explain why each piece might help this particular person, and only go outside when there is nothing in my material that fits.

Then I open the citations, all of them, because AI still makes things up with total confidence and a URL that loads is not proof that the page behind it supports the sentence sitting next to it. I ask for primary sources and direct links whenever current information is involved, and I check the dates, because a statistic from 2019 about anything AI-related is a museum piece.

Do not send everything it finds. Thirty links lands on her desk as a second research project when she is already behind, and part of what she pays you for is knowing which five of them matter.

7. Track progress so the conversation starts somewhere real

I built an Excel sheet — which, if you know me, is its own small miracle — that tracks client progress across the whole relationship, and it changed how our conversations open. I can see what we agreed to, what got done, where things slowed down, and which pattern keeps showing up in month after month, which is a better starting point than “so, how do you think it’s going?” and the answer we all give to that question, which is “fine, I think?”

Decide what progress means before you ask anything to track it, because the number that matters is different for every client. It might be decisions made, work published, conversations started, systems that exist now and did not before, confidence, or revenue — and the wrong measure will make a client who is doing beautifully feel like she is failing.

AI prepares the update and flags the gaps and the contradictions. I decide what any of it means, and the client gets to correct the record, because she knows things about her own month that no connected source ever captured.

8. Answer routine questions between sessions

I have not built this one yet, and I want it: a clearly disclosed AI assistant that can answer routine client questions between calls from my approved material, so that a client who needs to remember which framework we used in session two does not have to wait nine days for me to surface. VivekAI is coming, and when it arrives it will be built from my methods, my resources, my frequently asked questions and a very short leash.

It answers only from the material I approved, shows the client where the answer came from, and hands anything personal, sensitive, strategic or uncertain straight to me. It does not diagnose, does not promise, does not improvise, and does not pretend to be me — every client knows they are talking to an assistant, because the alternative is a trust problem I have no interest in creating.

Before something like this touches a real client, test it with ordinary questions, vague questions, emotional questions, questions built on a wrong premise, and questions that are nowhere near its scope, and write down the exact moment it stops and gets you. The refusal language matters as much as the answers, because a client who gets brushed off by your robot is not going to mention it to you, she is going to mention it to somebody else.

9. Turn a method you have repeated for years into a tool

The most valuable thing I have built with AI is intellectual property that works without me in the room, and the biggest of those is VIOS, which took three years of my thinking and turned it into a tool that walks a client, in a bit over two hours, through work that used to fill three to six months of coaching calls. Clients show up to our calls having already done the thinking, so I spend my time on the judgment and the strategy they hired me for rather than on the intake questions I have asked more times than I can count.

Look for the assessment you keep running, the question set you always ask, the decision path you walk people down, or the planning exercise that shows up in every third session. Write down the process and, more importantly, write down the places where it depends on your judgment, because those are the places the tool has to stop and hand back to you. If your method is already this repeatable, your IP is probably already an app and you have been giving it away as a PDF.

Start smaller than you want to. The first useful version is almost always an interactive worksheet, not a platform with accounts and dashboards, and you will learn more from ten clients using the scrappy version than from six months of building the beautiful one.

10. Collect the evidence for case studies while the work is happening

This is the other one I have not built, and it is in my sandbox for the next few months: a simple tool that collects approved notes, milestones, starting conditions, changes and the client’s own language as the work goes along, so that at the end of something excellent I am not reconstructing six months of results out of scattered calls, messages and documents.

What it would produce is a private evidence summary and a set of interview questions, not a finished testimonial, because the difference between a documented result and a result a client attributes to your work is a real difference and it belongs to her, not to you. Separate the two, cite where each claim came from, flag what needs confirming, and ask her in her own words rather than handing her a paragraph to approve.

Then get it in writing before anything goes anywhere: what she is happy to have shared, and whether she wants her name on it. I have watched people skip that step out of enthusiasm and spend the next year unable to use the best story in their business.

What should you tell clients about the AI?

Tell them, plainly, before they find out some other way. I say that I use AI for preparation, recaps, research and building resources, that I read and edit everything before it reaches them, and that the thinking, the advice and the judgment are mine — and nobody has ever reacted badly to that sentence.

What people react badly to is discovering it, and the shape of the discovery matters more than the fact. A recap that sounds like a corporate memo instead of the conversation they remember having, a resource that quotes a study that does not exist, a response that reads like it was written by somebody who was not on the call — that is how trust goes, and it goes over the sloppiness, not over the software.

Disclosure also gets easier the more specific you are, because “I use AI” means almost nothing to a client and can mean anything from spell check to a bot answering her emails. I wrote about how this plays out in public writing in the piece on AI transparency, and the same principle applies in private client work: say what it did, say what you did, and let the specificity do the reassuring.

What client information should stay out of an AI tool?

Use the smallest amount of material that can do the job, and leave everything else where it is. One recap needs one transcript; it does not need the client’s entire folder, her financials, her contract or the note you made after the hard call in April.

Before anything sensitive goes in, know where it lands: which provider is processing it, whether it is used for training, who on your side can see it, how long it stays, and what happens when you delete it. I know that reading vendor documentation is nobody’s favorite Tuesday (mine either), but you are the one who told your client her information was safe with you.

Test with old, anonymized or invented information before you point anything at current client records. Every workflow in this article got built and broken on made-up clients first, which is also how I found out that one of my early recap prompts was cheerfully including the asides people make when they think the recording has stopped.

When is AI the wrong answer in client work?

When the client is stuck, because stuck is not an information problem and AI will treat it like one. This is the counter-case to everything above, and it is the reason I am not worried about being replaced by a workflow I built myself.

When somebody tells me she is stuck, there are usually three different things happening and they need three different responses. Sometimes she is stuck because she does not know the next step, which is the easy one, and a resource or a framework or twenty minutes of explanation solves it. The second kind is stuck because there are forty next steps and she cannot see past the pile, which needs someone to help her take things off the list rather than add a tool to it. And the hardest one is when she does not want to go where the next step leads — the corporate client she should fire, the price she should raise, the offer she has outgrown and does not want to bury — and no prompt on earth is going to touch that.

An AI assistant handed all three of those gets it wrong in the same direction every time, which is toward more information, more options, more suggestions, a tidy numbered list of things to try. That is precisely the wrong medicine for the second and third kind, and a client who is already drowning does not need a longer list, she needs someone to notice that the list is the problem.

The software cannot tell which one you are looking at, and it cannot hear the difference in someone’s voice between “I have not gotten to it” and “I have not gotten to it,” which are two completely different sentences and I promise you have heard both.

What do you do instead when it is a coaching moment?

You slow the call down instead of speeding it up, which is the opposite of everything a productivity tool wants for you. When a client is in the third kind of stuck, I stop working on the business entirely, and sometimes — and I know exactly how this sounds coming from the woman who builds agents for a living — I run a short meditation with her right there on the call.

Five minutes, eyes closed, no laptop, and then I ask her what she noticed. What comes back is usually the thing she has been steering around for a month, and it comes out in about nine words, and then we can work. I have been doing this for years, it is a strange thing to be known for in a room full of tech people, and it has done more for my clients’ businesses than any workflow I have ever built.

The hard part of this work is almost never the information, and normalizing that is part of the job. When a client tells me she has had a resource open for three weeks and has not touched it, the useful response is that this is extremely normal and we should look at why, not a better-organized resource. AI would send her a better-organized resource. (It would be gorgeous. She still would not open it.)

How do you choose the first one to build?

Pick the piece of your client work that repeats, that you already understand well, and that currently depends on your memory, because that combination is where the setup pays for itself fastest. For most consultants and coaches that is session prep or the recap, which is why both sit at the top of my list.

Then do it by hand, in whatever AI you already pay for, before you build anything. Paste in the material, describe the job and the boundary, look hard at what comes back, fix your instructions, and run it again on a different client with messier notes — five runs on five real inputs will tell you more than any amount of planning, and it will show you the specific way this particular task fails.

The first one takes longer than doing the work yourself would have taken, and that is not a sign you chose wrong. You are writing down decisions you have been making by instinct for years and never had to say out loud, which is slow and slightly uncomfortable, and it is also why the second one takes twenty minutes.

What changes when the preparation is already done?

The conversation gets better, which is the only outcome I care about here. I arrive knowing where we left off, the client arrives having already done the thinking, the recap goes out the same day instead of the following Thursday, the resource she needs exists by Friday, and the progress we made in March is still visible in September.

None of that replaced a single thing my clients hired me for. It cleared the space around it, and what filled the space was attention — more room to listen, more room to push back, more room to notice the thing she is not saying and ask about it anyway.

If you want help deciding which part of your client work should go first and which part must stay yours, that is a good chunk of what I do. Otherwise, pick the one thing you rebuild for every single client, and start there.