Can You Build a Working Alliance with an AI Coach?

Ask a seasoned coach what actually produces change in coaching, and you will rarely hear « my method ». You will hear « the relationship ». Researchers gave that relationship a name — the working alliance — and half a century of evidence says it predicts outcomes better than technique, school or tool. Which sets up the most interesting question in coaching today: can that alliance exist with an AI? Our intuition screams no. The early data politely disagrees. Let's take both seriously.

What is the working alliance, exactly?

The concept comes from psychotherapy research. In 1979, Edward Bordin published a paper whose influence far outgrew its modest title: « The generalizability of the psychoanalytic concept of the working alliance ». His move was to strip the alliance of its psychoanalytic baggage and define it as something any helping relationship contains, in three components:

1. Agreement on goals — coach and client want the same change.
2. Agreement on tasks — both buy into what will be done to get there: the questions, the exercises, the homework.
3. The bond — mutual trust, respect and care between the two.

Bordin's sharpest claim was that the strength of the alliance, more than the kind of therapy or coaching wrapped around it, drives the change achieved. Decades of research across therapy — and later coaching — kept confirming the pattern: when alliance ratings are high, outcomes follow. When they are not, the world's most elegant methodology underperforms.

Notice something about those three components. Two of them — goals and tasks — are matters of structure and agreement, not of soul. Only the third, the bond, sounds like it should require a human on the other side. That decomposition is exactly why the AI question is not absurd.

What happens when you test it?

For years, the assumption that people cannot form an alliance with a machine went untested. Then researchers started testing it.

The most direct test to date is a 2024 study published in Frontiers in Psychology. The design: 52 graduate students, all with prior coaching experience — a deliberately demanding audience — were randomly assigned to a 60-minute leadership coaching session with either a human coach or what they believed was an AI coach, embodied as a talking avatar. (In reality the « AI » was simulated by the researchers — a standard method for isolating what perception alone does. Hold that caveat; we will come back to it.) Afterwards, everyone rated the working alliance on the standard inventory.

The result: human coaches averaged 74.50; the AI condition averaged 72.73. Statistically, no difference (p = 0.48). Clients who believed they were being coached by a machine reported goal agreement, task agreement and — most surprisingly — bond at levels indistinguishable from those sitting with a person.

This converges with what the broader chatbot literature has been finding for years in adjacent fields: users of well-designed conversational agents report genuine working relationships, often helped by a disinhibition effect — it is easier to say the embarrassing true thing to an entity that cannot judge you at the next partners' dinner. And it fits the outcome-side evidence too: in Terblanche and colleagues' two 10-month randomized trials, a goal-focused chatbot matched human coaches on goal attainment — hard to explain if no functional alliance existed. We review that full evidence base in does AI coaching work?

One underappreciated point makes these results less paradoxical than they look: the working alliance has always been measured from the client's side. It is the client's experience of agreement and bond — not the coach's inner life — that decades of research link to outcomes. Coaching never required the coach's feelings to be the active ingredient; it required the client's experience of being understood and usefully challenged. Once you frame it that way, the question shifts from the metaphysical (« can a machine care? ») to the practical (« can a system produce the experience of goal agreement, task agreement and being known? ») — and the second question is testable.

The honest caveats

We build an AI coach, so let us do the skeptic's job properly. Four reservations are warranted.

One study, one hour. The Frontiers finding is a single session with 52 people. Alliances deepen, rupture and repair over months; nobody has yet shown an AI managing a rupture — the moment a client feels misunderstood and the relationship must be actively mended — the way skilled humans do.

The AI was simulated. Participants' belief was tested, not current technology. That is precisely what makes the result clean — it isolates the human side of the equation — but claims about your actual chatbot cannot lean on it directly.

Alliance is not friendship. High bond ratings with an AI measure trust and felt understanding within a working frame. Nobody is claiming the machine cares. The claim is narrower and more useful: the client's experience of being understood, which is what the alliance inventory measures and what predicts outcomes.

The bond may have a ceiling. Graßmann and Schermuly's review of AI coaching capabilities suggests machines handle goals and tasks well but hit limits with strong emotion and deep confrontation. It is plausible the alliance with an AI is real but shallower — sufficient for structured development work, insufficient for identity-level upheaval. That is where a human coach earns the fee, as we detail in AI coach vs human coach.

Why does an alliance with an AI form at all?

Run Bordin's three components through what a decent AI coach actually does, and the mystery shrinks.

Goals: an AI coach can hold your goals with perfect fidelity — restated at every session, never forgotten, never blurred by the coach's own agenda. This is, frankly, a machine's home turf.

Tasks: agreement on tasks requires proposing relevant work and adjusting when it does not fit. Structured questioning, reframing, rehearsal, commitments — the validated toolbox of goal-focused coaching — are exactly what AI systems execute consistently.

Bond: this is where design quality separates products. A bond requires feeling known. A generic chatbot that greets you like a stranger every Tuesday will never get there. This is why we built Lumia around persistent context: it knows your Octagon profile — your self-assessed patterns across eight leadership pillars — and it remembers what you committed to last Thursday. When a coach, silicon or carbon, says « you said the same thing about your CFO three weeks ago », the feeling of being seen is not simulated. The observation is real.

What this means for you in practice

If the alliance drives outcomes, then your job as a user is to build one deliberately — the research suggests it is largely in your hands:

• Negotiate the goal explicitly. Tell your AI coach what you want from the next month, not just from the next message. Alliance component one does not form around vague topics.
• Push back on tasks. If an exercise feels off, say so and ask for another angle. Task agreement is built through exactly this negotiation — and it is a place where machines, unoffendable, take feedback better than some humans.
• Use the candor advantage. The judgment-free channel is only worth something if you put the true, uncomfortable material into it. Alliance follows disclosure, in both directions.
• Watch for the ceiling. If sessions keep touching something heavy — grief, panic, despair — the honest move is a human professional. A good AI coach should tell you the same.
• Close the loop. End sessions by telling the coach what helped and what missed. With a system that retains context, that feedback compounds into a working relationship tuned to you — the closest digital equivalent of an alliance maturing over time.

Our practical guide, how you should use your AI coach, turns these into a weekly routine.

Where we land

Can you build a working alliance with an AI coach? On current evidence: yes — a functional one, formed faster than intuition predicts, resting on the two structural pillars machines handle well, with a bond that is real in its effects even if its depth over the long run remains an open research question. The alliance is not the obstacle people assumed. The quality of the coach — its memory, its context, its honesty about its own limits — is.

Stop reacting. Start seeing. If you want to test the alliance question on yourself, start with a coach that already knows your profile: Lumia begins from your Octagon, not from zero. Practice it with Lumia, your AI coach.