AI and the Hard Part of Managing Clients

Describe a wobbly client relationship to AI and ask what's going on, and the response usually lands on something plausible and generic: communication has broken down, expectations may have drifted, consider a check-in call. All reasonable. All the kind of advice that could apply to almost any strained relationship, professional or personal, which is exactly the problem: it isn't diagnosing your specific client, it's pattern-matching to "relationship trouble" in general.

The more useful question isn't "what's wrong with this client relationship." It's "which specific input into their trust in me has actually weakened," a question that produces a genuinely different, more actionable answer than the vague sense that something has shifted.

What does asking AI about a shaky client relationship actually produce?

Describe a client who's gone quiet or started pushing back on scope, and ask AI what's happening, and you'll get a reasonable but generic diagnosis: possible miscommunication, expectations may have shifted, worth a check-in. It's the kind of read that applies to almost any strained relationship, which means it isn't really diagnosing your client's specific situation at all.

The gap isn't inaccuracy. It's that a generic prompt has no structured way to separate the different things that make a client trust an advisor, so it defaults to a single, vague read instead of checking several specific, separately diagnosable causes, each of which points to a different, more useful next step than "reach out and reconnect."

Is AI actually useful for reading client risk, or does it just sound informed?

Unaided, it's genuinely good at organizing a scattered set of observations, late replies, a shrinking meeting cadence, a new CC'd stakeholder, into a coherent pattern once you list them out. It's also useful for generating a calibrated next step rather than a generic "reach out."

It's weak at anything it can't observe directly: what the client's internal politics look like, whether a budget cut is coming from above them, whether their quietness is about you at all or about something happening entirely inside their own organization. It reasons entirely from what you report, so an incomplete or self-flattering account of the relationship produces an incomplete diagnosis back.

What's worth asking AI before you misread this relationship?

These four work in Claude, ChatGPT, or Gemini with no setup required.

1. The trust-input audit. "Here's what's changed in this client relationship recently: [describe it]. Score, roughly, whether the issue looks like a credibility problem, a reliability problem, or an intimacy problem, credibility meaning expertise and track record, reliability meaning consistency and follow-through, intimacy meaning psychological safety and candor." Why it works: forcing a choice between three distinct categories produces a sharper diagnosis than an open-ended "what's wrong," because each category points to a different fix.

2. The self-orientation check. "Be honest: in my recent interactions with this client, is there anywhere I was more focused on looking competent or protecting my position than on their actual problem?" Why it works: this is the input most people skip entirely, since it requires looking at your own behavior rather than the client's, and it's exactly the input the documented equation treats as capable of undermining all the others.

3. The evidence sort. "List every specific, dated example I've given you of this relationship changing. For each, is it clearly about me, clearly about their internal situation, or genuinely ambiguous?" Why it works: it's easy to attribute every signal to your own performance under stress, and separating genuinely ambiguous signals from clear ones keeps you from over-correcting on a read that isn't actually about you.

4. The next-conversation prep. "Based on this read, draft three questions I could ask the client directly that would surface which trust input is actually weakest, without sounding defensive or needy." Why it works: a diagnosis that never becomes an actual conversation with the client doesn't fix anything; this converts the internal read into something you can actually say out loud.

What hasn't this fixed about the account?

This read will sharpen your understanding of this one relationship, more than a generic check-in would produce on its own. It won't catch early trust erosion across your whole client book by itself, the account that looks fine today and is quietly slipping by next quarter, well before it becomes an obvious problem. A single diagnosis isn't the same as a habit of checking routinely, before something already feels wrong.

How does the Trust Equation actually change the diagnosis?

David Maister, with Charles Green and Robert Galford, published the Trust Equation in The Trusted Advisor: Trustworthiness = (Credibility + Reliability + Intimacy) / Self-Orientation. Credibility is expertise and track record, reliability is doing what you said you'd do, consistently, intimacy is the psychological safety that lets a client say something true and uncomfortable to you, and self-orientation, in the denominator, is how much you're focused on your own position rather than the client's actual problem. The documented insight in the denominator specifically is that self-orientation doesn't just fail to add trust, it actively divides whatever the other three have built.

Before, generic prompting: told that a client has started copying a second stakeholder on every email and taking longer to reply, a generic AI response suggests scheduling a check-in call to "realign on expectations," treating the symptom as the whole problem.

After, the Trust Equation applied: the framework asks which input the pattern actually points to. A new stakeholder being copied often signals an intimacy gap, the client no longer feels safe raising concerns with you directly and privately, rather than a credibility problem. The framework then runs the self-orientation check specifically: has communication recently become more about defending decisions you've made than about the client's actual concern. If so, the documented fix isn't a status update call, it's deliberately creating space for the client to say something uncomfortable without you immediately defending your position.

A better-worded check-in doesn't fix a diagnosis problem. Checking each specific input separately does, self-orientation included, the one input most people are least willing to examine honestly in themselves.

How do you get this framework running in Claude, ChatGPT, or Gemini?

Claude runs it as a Skill: Settings → Skills → Add skill, upload the .zip. It auto-invokes when your question matches the skill's description, or gets forced with /david-maister-framework at the start of a message. From Claude Code, the package sits in ~/.claude/skills for use everywhere, or a project's own .claude/skills folder for just that one.

ChatGPT takes the plain .md into a Custom GPT instead: Create a GPT → Configure → Instructions, paste it in. Skip the standard Custom Instructions fields here, they're capped at 1,500 characters; a Custom GPT's Instructions field runs to roughly 8,000, enough room for the full framework.

Gemini has no dedicated upload option, so the content goes straight into the system prompt where one's available, or opens the first message instead: "Here is a thinking framework to apply throughout this conversation. Read it carefully, then answer using this framework's approach," with the framework text below it.

Where should this take you next?

For the specific account that's worrying you right now, Client Risk is a $149 tool built to structure exactly this judgment call, including the retain-or-fire decision, across your whole client book. Maister's method, credited jointly with Charles Green and Robert Galford as the documented history requires, is one of several in the Consultant & Advisor category, alongside Green's own framework and Peter Block's, each an .md file that works with Claude, ChatGPT, or Gemini.