AI Won't Shrink the Therapy Market. It Will Expand It.

Every fear about AI in mental health rests on the same unstated assumption: that the market for therapy is fixed. That there is a fixed number of people who will ever seek professional mental health care, and every conversation they have with a chatbot instead of a clinician is a client lost forever.

That assumption is almost certainly wrong. And the history of every other resource that got cheaper and easier to access says so.

The paradox that keeps repeating

In 1865, the economist William Stanley Jevons noticed something strange about coal. As steam engines became more efficient, using less coal to do the same amount of work, total coal consumption didn't fall. It rose. Cheaper, more efficient use of a resource didn't shrink demand for it. It expanded the range of things people used it for, and total consumption went up.¹

This isn't a quirk of coal. It's a pattern that shows up whenever the cost of doing something drops sharply. Cheaper air travel didn't shrink the airline industry, it created a market of flyers who never used to fly. Cheaper computing power didn't reduce demand for engineers, it created entire industries that needed more of them. Cheaper legal templates didn't kill the legal profession, they expanded who could afford to start engaging with legal problems at all, and a share of those people went on to need a real lawyer.

Right now, the barrier to entry for engaging with your own mental health has been cost, time, availability, and stigma. Conversational AI is lowering all four at once. Under the fixed-pie assumption, that should shrink the addressable market for therapists. Under Jevons Paradox, it does the opposite: it expands the total population of people willing to look at their own mental health in the first place, and a portion of that expanded population will need more than AI can give them.

The real competitor isn't the therapist

Clayton Christensen spent his career studying why big, capable incumbents lose to smaller entrants that look laughably inferior at launch. His answer was almost always the same: the incumbent was busy competing with other incumbents, while the entrant was actually competing with nothing, with the enormous population of people who weren't buying any version of the product at all.² He called this competing against non-consumption, and it explains a huge amount of what looks like disruption from the outside.

Apply that lens here. A licensed therapist's real competitive set isn't a $20-a-month AI chatbot. It's the roughly 60 to 70 percent of people with a diagnosable mental health condition who receive no treatment at all in a given year.³ That's the market AI is actually entering. Recent research led by John Torous and colleagues, reviewing nineteen surveys published since 2024, estimates that around 27 percent of AI users have used AI for some form of mental health support, and around 40 percent say they'd be open to it.⁴ Those are not people switching from a therapist to a chatbot. The overwhelming majority are people for whom the honest alternative was doing nothing, at 2am, alone, with a problem they'd never said out loud before.

That's the first-order effect: AI is currently expanding total engagement with mental health, not cannibalizing existing care.

The second-order effect is the one that matters for your caseload

The more important question is what happens next to the people AI reaches. A chatbot can help someone name what they're feeling, learn a coping skill, and get through a bad night. What it consistently cannot do is hold a therapeutic relationship over months, work through trauma that requires a trained clinical judgment call about pacing and safety, or notice the kind of pattern that only shows up across dozens of sessions with someone who remembers who you were a year ago.

So the second-order consequence of millions of first, low-stakes conversations with AI is not fewer people needing therapists. It's more people discovering, often for the first time, that they have something worth working on with one. Reduced stigma and a larger vocabulary for describing internal experience are consistently associated with higher, not lower, rates of formal help-seeking in the public health literature on stigma reduction. AI conversations, at scale, function as an enormous, informal stigma-reduction campaign, whether or not that was anyone's intention.

This is also visible in where the money has actually gone. US spending on mental health and substance use disorder treatment grew from $40.9 billion in 2000 to $139.6 billion in 2021, according to a Health Affairs analysis of Bureau of Economic Analysis data. Critically, 87.3 percent of that growth came from more people entering treatment, not from treatment getting more expensive per person.⁵ Every previous wave of expanded access (parity legislation, telehealth, reduced stigma) has produced the same shape of curve: access goes up, and the number of people who go on to want real treatment goes up with it. There is no historical precedent for expanded access shrinking a care market. AI-driven access is very unlikely to be the exception.

Why the therapist relationship gets safer, not smaller

None of this works, however, if the higher volume of people entering the system through AI just burns out the clinicians on the other end of it. This is where the Pareto Principle earns its place in the argument. A large share of a therapist's cognitive load between sessions isn't the clinical work itself, it's reconstruction: catching up on what happened this week, re-establishing context, triaging what's urgent versus what can wait. If AI absorbs the volume of that reconstruction work (structured check-ins, symptom tracking, between-session support) it frees the clinician to spend a much higher proportion of session time on the small set of moments that actually require a trained human: the breakthrough, the disclosure, the intervention that needs real clinical judgment.

That same daily volume of AI-mediated contact also becomes a margin of safety the weekly model never had. A once-a-week cadence means a crisis on Tuesday isn't seen until Thursday. A hybrid model with supervised daily check-ins and crisis screening catches the same signal in hours, not days, and surfaces it to the clinician before the next scheduled session. Rather than making the relationship more fragile, the volatility of a patient's week becomes something the system learns from and adapts to. The relationship gets more resilient under stress, not less, precisely because more of the noisy, unscheduled parts of a patient's life are now visible instead of invisible.

What this means for the profession, not just the market

If this thesis holds, the practical implication for therapists isn't defensive, it's closer to Robert Axelrod's Red Queen framing: you have to keep adapting just to hold your relative position, because the baseline of what patients expect is moving. A patient who has spent six months talking to an AI at 2am will arrive at their first real session with a different, higher baseline expectation of continuity than a patient did five years ago. They will expect their therapist to already have context. They will expect care to persist between sessions, not restart from zero every Tuesday at 4pm.

Therapists and group practices who treat this shift as a threat to be resisted will find themselves competing on yesterday's terms against practices that have learned to safely absorb AI-mediated volume, triage it, and turn it into better-prepared sessions. Those who upskill now, on how to safely supervise AI-assisted care rather than avoid it, are positioning for a market that is expanding toward them, not shrinking away from them.

The bet underneath all of this

The honest version of this thesis has a condition attached: it only holds if the AI layer is built to route people toward real care when they need it, rather than optimizing to keep them talking to a bot indefinitely. A product that maximizes engagement time with AI, at the expense of referring someone into human treatment when the evidence says they need it, would break the whole mechanism and genuinely could cannibalize the therapy market instead of expanding it. That's a design choice, not an inevitability, and it's the reason safety architecture and crisis escalation are not a peripheral feature of AI mental health products. They're the entire hinge on which this argument turns.

Built the right way, AI in mental health is not competing with therapists for a fixed number of clients. It's competing with silence, with 2am, with the version of someone's week where they never said anything to anyone at all. Every person that competition wins is, eventually, good news for the people qualified to do what AI cannot.

Explore how Citt.ai's approach to AI safety keeps that hinge intact, and read more on why AI should make therapy more human, not more efficient.

---

Citt.ai is an AI co-pilot for therapist-led care, built to extend clinical context between sessions without ever replacing the clinical relationship. Explore features or book a session to see it in practice.

Footnotes

1. Jevons, W.S. (1865). The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. The original formulation of what became known as Jevons Paradox: efficiency gains in resource use tend to increase, not decrease, total consumption of that resource. ↩

2. Christensen, C.M., Hall, T., Dillon, K., & Duncan, D.S. (2016). Competing Against Luck: The Story of Innovation and Customer Choice. HarperBusiness. Christensen's "Jobs to Be Done" framework argues that the most significant competitor to a new product category is frequently non-consumption, not existing competitors. ↩

3. Estimates of mental health treatment gaps vary by country and condition, but US and global surveys consistently find that a majority of people with a diagnosable mental health condition in a given year receive no formal treatment. See SAMHSA National Survey on Drug Use and Health and WHO World Mental Health Survey Initiative data on treatment gaps. ↩

4. Torous, J., et al. (2026). Systematic review of 19 surveys (2024–2026) estimating the prevalence of AI use for mental health support. Nature Mental Health. Available at nature.com/articles/s44482-026-00025-7. Individual survey estimates ranged from 3% to 70%, reflecting inconsistent definitions of "AI use for mental health"; the authors' pooled approximation was 27% of AI users, with roughly 40% expressing openness to it. ↩

5. Saloner, B., et al. (2025). US National Spending on Mental Health and Substance Use Disorder Treatment Driven by Case Growth, 2000–21. Health Affairs. doi.org/10.1377/hlthaff.2025.01351. Analysis of Bureau of Economic Analysis Health Care Satellite Account data found 87.3% of the spending increase was attributable to growth in the number of people treated, versus 12.7% from increased cost per case. ↩