The Captivity Well

A framework for reading retention when users cannot leave

Framework paper · September 2026

Summary

Product teams read retention as a verdict on quality. That reading only holds when users could realistically leave. In public services, heavily integrated enterprise platforms and, increasingly, AI products, users often stay not because the product earns it, but because leaving costs more than staying.
This paper extends the Gravity Well concept, which explains why users stay with products that are good enough, into a second mechanism: the Captivity Well, which explains why users stay with products that are not. It also:

Describes how AI builds captivity faster than earlier software;
Shows where the cost of poor design goes when users cannot leave;
Proposes a method for separating earned retention from trapped retention;
Closes with a worked example and practical guidance on designing the exit.

The models and the scoring method in this paper are conceptual. They synthesise established research on switching costs, exit and voice, and administrative burden, together with recent findings from competition authorities. They have not been empirically validated. Their purpose is to change which questions a product organisation asks, not to replace judgement with a score.

There was no dedicated design thinking process in place, which meant designers were expected to produce outputs without a clear structure for influencing direction earlier. As a result, many solutions were delivered at a service level instead of being shaped at an experience level.

I saw an opportunity to change that by building a discovery-led way of working that could strengthen design culture, improve collaboration across product teams, reduce costly late-stage refinements, and give design a stronger role in shaping product quality before execution began. This work did not just improve how design worked. It improved how product work started.

1. The problem: usage without exit

Some years ago, a service provider charged me twice for the same service, and the call to customer service went nowhere. My mentor, sitting across the desk, pointed out what I had missed. The agent on the phone was the only person who could solve it, and walking away would not undo the charge. The problem was eventually solved. But the situation stayed with me: an unhappy customer with nowhere else to go.
From the provider’s side, that call looked like a retained customer. That is the problem this paper addresses. When exit is unavailable, the signal that normally tells a provider its product is failing goes quiet. Dashboards keep reporting success. Dissatisfaction does not disappear. It moves somewhere the dashboard is not looking.

2. Key terms

Term

Meaning in this paper

Service-driven mindset

Users stay with a product that is good enough, because a better alternative is not worth the effort of switching

Captivity Well

Users stay with a product that is not good enough, because switching is not realistically available

Contestability

Whether a dissatisfied customer can take their usage, data and workflows elsewhere at a justifiable cost

Exit feasibility

How realistically a specific user or organisation could leave, assessed across six factors (section 10.1)

Adoption moat

Retention protected by continuing value: users keep choosing the product

Exit moat

Retention protected by switching costs: users would leave but cannot economically do so

Quality floor

The minimum quality a provider must maintain, set by whichever pressures act on it

Preference retention

Users who would choose the product again if switching were effortless

Friction retention

Users who would not choose it again, could realistically leave, but find moving not worth the effort

Constraint retention

Users who would not choose it again and cannot realistically leave

3. Starting point: the Gravity Well

The Gravity Well explains why specialist AI products struggle against general-purpose assistants. A tool that is familiar, always available and good enough across many tasks becomes the default. A specialist alternative may be clearly better at one job, but the improvement is smaller than the effort of finding, learning and adopting it.

As a conceptual model, a user switches only when:

Value of the alternative − Value of the current product > Switching cost + Learning cost + Coordination cost + Risk

In the Gravity Well, the left side is too small. The alternative is better, just not better enough.

The model rests on one assumption: that the user could leave. It describes a choice not to switch. It does not cover situations where switching was never realistically available.

4. The Captivity Well

A second mechanism produces the same visible behaviour, the user stays, for the opposite reason.

Gravity Well: “This is good enough, so switching is not worth the effort.”
Captivity Well: “This is not good enough, but switching is not realistically available.”

In the Gravity Well, the left side of the equation is too small. In the Captivity Well, the right side becomes disproportionate, or no viable alternative exists at all.

Two questions, whether the product is good enough and whether exit is realistic, give four situations:

The Gravity Well and the Captivity Well look identical in behavioural data. In one, staying means preference. In the other, staying means constraint. When leaving is expensive, usage stops being evidence of preference.

Captivity has several sources, and most products combine more than one:

Source

Why users stay

Typical examples

Legal or institutional

No authorised substitute exists

Tax filing, passports, permits, public benefits

Technical

Leaving means rebuilding integrations, code or data

Cloud infrastructure, enterprise software

Contractual

Commitments and terms make exit expensive

Multi-year enterprise and cloud agreements

Network

The alternative loses value without the same ecosystem

Communication and platform ecosystems

Organisational

The organisation has standardised around one system

ERP, identity, productivity platforms

Data and workflow

History and processes live inside the incumbent

CRM, healthcare, analytics

Context

Accumulated knowledge about the user does not travel

AI assistants with memory, personalised agents

Risk

Migration is possible in theory but dangerous in practice

Core banking, healthcare, critical infrastructure

The pattern is not limited to one industry. Changing banks means rebuilding salary arrangements, standing orders, cards and payees. Changing payroll systems means re-validating every employee record. Changing an AI platform means rebuilding everything the platform has learned. Wherever the experience is poor and leaving is expensive, users absorb the poor experience rather than pay the price of moving.

5. Contestability, not competitor count

Captivity does not require a monopoly. Farrell and Klemperer showed that switching costs and network effects can give a supplier market power over customers it has already won, even after competing hard to win them. Many markets move through two phases: competition for the customer, then lock-in after adoption. On paper, competitors still exist. In practice, the installed customer negotiates with the only supplier it can realistically use.

The relevant variable is therefore contestability, not the number of competitors.

Hirschman’s Exit, Voice, and Loyalty explains what happens when contestability is low. People respond to decline by leaving or by speaking up. When exit is blocked, voice has to carry the full load. The meaningful comparison is not private companies against governments. It is between three kinds of system: those disciplined by exit, those disciplined by something else, and those disciplined by neither. The third group is where quality erodes without anyone noticing.

6. The AI lock-in stack

The claims in this section are propositions. They are grounded in regulatory evidence at the infrastructure layer and in observable product patterns above it. The personal, organisational and agentic layers have not yet been studied empirically, and section 13 sets out how they could be.

Most software builds switching costs slowly, through data and integrations accumulated over years. AI products appear to build them faster, in layers that are harder to see:

The personal layer is the newest and least discussed. An assistant that learns a user’s context becomes more useful every week. That is the point. It also means that value mostly stays behind when the user leaves. The agentic layer is the deepest. Agents carry keys to internal systems and embody decisions about who approves what. Replacing them is an organisational change, not a software change.

The evidence at the bottom of the stack is already clear. In July 2025, the UK Competition and Markets Authority concluded its cloud services market investigation. It found that fewer than 1% of customers switch cloud provider in a year. The barriers were transfer fees, technical differences, integration difficulty and a shortage of transferable skills. The CMA linked customers’ ability to switch directly to providers’ incentive to improve. Separately, the US Federal Trade Commission published a staff report in January 2025 on partnerships between cloud providers and AI developers. It noted that these arrangements can raise both contractual and technical switching costs.

Regulation is beginning to target exit directly. The EU Data Act, applicable since September 2025, includes provisions intended to make switching between cloud and data-processing providers work in practice. The regulatory question is moving from “is the incumbent charging too much?” to “can the customer leave?”

6.1 The captivity lifecycle

AI products tend to follow a recognisable path:

The shift from the first question to the last is rarely deliberate and never announced, and the metrics look the same throughout. An adoption moat and an exit moat both appear as low churn. Few strategy reviews separate them.

6.2 Agents cut both ways

The same agentic capability that deepens lock-in can also dissolve it. An agent can compare offers, move data, rebuild configurations and handle the paperwork of switching. That lowers the effort of leaving dramatically. Much of captivity has always rested on effort rather than rules. Nobody forbids a customer from changing banks; it is simply tedious. Tedium is what agents are best at removing.

There is a second shift. As agents act on behalf of users, the user may stop experiencing products directly. The agent becomes the customer. Agents do not form habits or feel attached to familiar interfaces. They evaluate outcomes. The competitive frontier has already moved from goods to services to experiences. In the agentic era, it moves towards systems where AI acts for the user. A provider that relies on friction to retain customers is betting that users will keep doing the switching work themselves. That bet weakens every year.

7. Where the cost of poor design goes

In a competitive product, poor experience costs the provider: a confusing checkout loses the sale. In a product users cannot leave, the transaction usually still happens, and the cost moves to the user.

Provider complexity → User time + cognitive effort + errors + support needs

Public services face the hardest version of this problem, because many are the only provider by design. The OECD’s work on “sludge” treats unnecessary friction as a substantive barrier to access rather than a cosmetic issue. The research supports that position. Bhargava and Manoli ran a field experiment with the US Internal Revenue Service. It involved around 35,000 tax filers who appeared eligible for the Earned Income Tax Credit but had not claimed it, together worth roughly $26 million. Simpler, clearer notices increased claiming. The study shows how evidence can locate the point where design stands between people and what they are entitled to.

When users cannot easily leave, dissatisfaction does not appear as churn. It appears as unclaimed value, errors, delays, support contacts, workarounds and reliance on intermediaries.

8. The quality floor

If competition is not setting the minimum standard, something else has to. As a simplified model, the provider’s quality floor can be treated as the strongest of several pressures acting on it:

Quality floor = the highest of (competition, regulation, reputation, mission, user voice, operational need)

Operational need is the quality required simply to keep the service running at acceptable cost: error rates, support volumes and manual rework the organisation cannot afford to carry.

The model is a simplification. In practice, pressures can reinforce or offset each other. A strong mission can be undermined by weak funding, and reputational pressure can amplify regulation. But the simplification makes one point clear. A product’s quality floor is set by whichever pressure is actually binding, not by the one the organisation talks about.

In a competitive consumer market, competition usually sets the floor. In a public service, the competition term is close to zero, so regulation, mission, accessibility obligations and user voice must do the work. They can. The UK Government Digital Service Standard assesses services on six things: understanding users, solving a whole problem, simplicity, accessibility, iteration and performance measurement. The assessments are published. HMRC’s “View and change your tax account” service was rated Red at alpha assessment in June 2024. The assessors acknowledged real strengths and identified journeys that needed more testing. The team worked through the findings, and the February 2025 reassessment moved the service to Amber. No competitor drove that improvement. A deliberately designed feedback loop did.

The risk zone is a captive product where every pressure in the model is weak at once. Competition is one way to create a quality feedback loop. It is not the only way. But some loop must exist, and when exit is weak, it has to be designed on purpose.

9. Three types of retention

Observed retention can be decomposed into three components:

Observed retention = Preference retention + Friction retention + Constraint retention

Two questions assign each retained user to exactly one component:

Would they choose the product again if switching were effortless?
Could they realistically leave?

Benign lock-in counts as preference retention. A satisfied user is satisfied, whatever their exit options.
The Gravity Well contains both preference and friction retention, depending on whether users would actively choose the product again.
The Captivity Well is constraint retention.
Competitive correction does not appear in retention at all, because those users have left.

The same logic explains the three kinds of “good enough” the phrase currently blurs together:

Equilibrium

What the user thinks

Retention component

Satisficing

“It solves my problem. Something better isn’t worth the effort.”

Friction retention

Locked-in

“I’d prefer something better, but switching costs too much.”

Constraint retention

Mandatory

“I have to use this.”

Constraint retention, with no viable substitute

These are different kinds of advantage:

Preference retention indicates a strong product.
Friction retention indicates a defensible incumbent.
Constraint retention indicates dependency.

They should not be reported as one number. Constraint retention also carries a hidden risk. It looks like strength until something lowers the cost of leaving: a regulation, an open standard, a migration tool, an agent that does the switching. An incumbent that mistook captivity for satisfaction can discover quickly how shallow its well was.

10. Measurement toolkit

10.1 Exit feasibility assessment

The assessment below is a working method, not a validated instrument. It is meant to make the conversation about exit concrete and comparable across a team, not to produce a score that stands on its own. It can be applied to a whole customer base or, more usefully, to individual segments. Each of six factors is rated from 0 (exit effectively blocked) to 3 (exit straightforward).

Factor

Guiding question

0