Dev.to AI 🤖 Ai 👁 0 📖 5 min read

Design AI onboarding around five levels of customer effort

A developer who wants an API and an owner who wants a finished report may need the same underlying AI capability. They do not need the same product. My working customer framework has five buying modes: builders, guided

A developer who wants an API and an owner who wants a finished report may need the same underlying AI capability. They do not need the same product.

My working customer framework has five buying modes: builders, guided adopters, casual chat users, outcome buyers and non-adopters. These are practical product-design labels, not mutually exclusive populations. Someone can occupy several modes for different tasks.

This article proposes an onboarding and evaluation approach. It does not report results from a system we have built or tested.

Give builders a working interface, not an adoption course

For builders, the first useful experience may be a successful API call, a command-line task or an agent connecting through MCP. Let them inspect the permissions, inputs, outputs and failure modes.

An infrastructure product should show a minimal example, a realistic error and a route to debugging. Make logs and source provenance available where they matter. Avoid forcing a technical customer through a consumer-style wizard before they can understand the integration.

A useful first metric is time to a successful, authorised action on a realistic task. A copied quickstart that cannot survive a changed input is not enough.

JetBrains' 2026 developer survey documents broad coding-agent adoption. It is evidence about developers and tool use, not a census of the self-starting automation builders this product targets.

Let guided adopters copy a result and understand its dependencies

A guided adopter wants a skill, template or demonstrated workflow. Design the starting point around a concrete job, such as sorting incoming enquiries into a review queue.

Show the accounts required before asking for credentials. Explain which actions the workflow can take and where approval is needed. Provide representative test data and a way to run without affecting real customers.

Test whether a new user can finish setup, recognise an error and complete a second run. Count support effort. A template that only its author can configure is closer to a consulting lead than a self-serve product.

The recommendation to bundle setup is a product hypothesis to test, not a claim that all template buyers want the same service.

Remove orchestration work for casual chat users

A casual chat user may be comfortable asking questions but unwilling to invent a workflow. The product should supply the sequence.

For a document-preparation tool, ask the few questions needed for the specific document, identify missing information and show a reviewable output. For a tutor, structure practice and feedback instead of opening on an empty prompt box.

Measure task completion and correction, not conversation length alone. Ask whether a user can tell what happened, what remains uncertain and what needs their decision.

Bain's consumer research distinguishes enthusiasts and dabblers, but its December 2024 survey is US-focused. It does not provide the global size of this product segment. OpenAI's February 2026 report of more than 100 million weekly ChatGPT users in India is a platform-use figure, not paid demand for your application.

Give outcome buyers scope, accountability and a handover

An outcome buyer prefers delegation to another setup project. A managed agent service can present a clear brief, delivery boundary and escalation path.

A hypothetical reporting service might agree on the data sources, a delivery schedule, review requirements and the format of the finished report. It should not promise an improvement in business performance that the provider cannot attribute or control.

Test delivery reliability, exception rate, human review cost and customer acceptance. A successful model call is not the unit of value. The accepted deliverable is.

Keep permissions explicit. Buying a service does not automatically authorise every email, payment or account change that could help achieve its goal.

Keep a familiar front door for people who do not want AI

Some buyers prefer a person, phone call or existing agency relationship. Others use AI embedded in familiar software without identifying it as AI.

Pew found that 51% of US adults said they never used chatbots in its February 2026 survey. The figure is US-only and about chatbot use. It cannot be used as a worldwide non-adoption rate.

For these customers, reduce the visible complexity without misrepresenting material data handling or human involvement. Let them buy the service they understand. Test whether the familiar interface actually improves completion and trust, rather than assuming age or occupation determines preference.

Keep a sizing notebook separate from your customer database

The research behind this framework included rough estimates. They are useful only if their methods and limits remain attached.

19.5 million developer scenario: 50 million active developers estimated by SlashData multiplied by JetBrains' 39% Claude Code workplace-adoption share. Different populations and survey dates make this an illustrative cross-source calculation, not a count of power users or your reachable buyers. Active developers are not identical to the professional-developer sample.

3.3 million paid-plan equivalents: assume $1 billion annual revenue and $25 monthly revenue per subscription, then divide by $300 a year. The result is about 3.3 million subscription equivalents, or about 1.7 million at $50 monthly. The revenue assumption was informed by a secondary compilation of vendor-claimed app-builder run rates from different dates. It mixes business models and can include enterprise or usage revenue. Customers can use several tools. This is not unique paying prosumers or an established minimum market size.

7.2 billion remainder, not a cohort estimate: the draft research subtracted 900 million ChatGPT weekly users from an illustrative 8.1 billion population input. That misses other AI products, occasional users and reachability. The arithmetic is visible here to identify the mistake, not to endorse it as an estimate of non-adopters.

Adjacent spending: Mordor's estimated $3.61 billion outcome-as-a-service software market in 2026 and forecast $15.91 billion in 2031 include industrial and enterprise uses. Its separate $473.6 billion marketing-agency estimate for 2026 covers the whole agency market, including AI adopters. Neither measures the outcome-buyer or non-adopter group.

Do not add these figures together. They have different units and overlapping scope. A product's own qualified pipeline and cohort retention will be more actionable than a fabricated global total.

Test the boundary between modes

A template buyer may later pay for managed execution. A casual user may become a builder for one valuable task. An expert may still buy concierge delivery when time matters.

Offer a clear upgrade path between those modes where your service can support it. Do not force every customer toward maximum configuration just because the product team enjoys building tools.

The design question is how much effort and accountability this customer wants to keep for this job. Answer that before deciding how much interface to build.

Sources

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.