Research-backed calculator guide

AI Add-On vs Bundled Feature Pricing: Test Packaging Against COGS and Adoption

Compare an AI feature add-on with bundled pricing using attach rate, incremental COGS, adoption, cannibalization, and a worked break-even calculation.

Add-on vs bundled AI pricing: what changes economically?

An add-on makes the customer opt in. That protects customers who do not value the feature and gives you a clean adoption signal, but it creates a second purchase decision. The price must overcome that friction. The add-on also concentrates usage among the users who care most, which can be good for value capture and bad for cost concentration. If power users attach first, the average COGS per attached account may be much higher than the average COGS observed in a free trial.

A bundle moves the feature into a plan or tier that customers already understand. It can reduce checkout friction and expose more users to the capability. The trade-off is cross-subsidy: light or non-users fund the users who consume the most. A bundle can also force a customer to buy unrelated features to obtain the AI capability. That is a packaging problem, not a model-price problem.

There is a third option worth testing: a mixed bundle. Include a bounded allowance in the base or higher tier, then offer an add-on or usage-based overage for customers who exceed it. This keeps the first experience predictable while preserving a meter for the tail. Stripe’s usage-based billing documentation describes usage meters, prepaid credits, threshold monitoring, and adding usage pricing to flat-rate subscriptions as separate implementation needs.^[4] That is a useful reminder that “bundled” and “metered” are not mutually exclusive.

Use an add-on first when the AI job is optional, has a distinct buyer, or has volatile usage. Use a bundle when the AI feature completes the core job for most customers, is now a table-stakes expectation, or has a narrow and predictable cost distribution. Use a mixed model when adoption should be broad but a small group can create disproportionate spend.

Microsoft’s business Copilot page is a dated example of the distinction rather than a universal benchmark. On 2026-11-02, the page showed Copilot included in Microsoft 365 Business plans with yearly prices of $23.50 and $32.00 per user/month, alongside a Microsoft 365 Copilot Business add-on displayed at $18.00 promotional pricing and $21.00 per user/month. The page also states that selected advanced AI experiences use usage-based billing.^[2] Those are vendor-published terms for that product and promotion window, not a recommended price for your feature.

Measure the right inputs before choosing a package

Start with the denominator. “Adoption” can mean the percentage of all paid accounts, eligible seats, activated users, or users who complete a qualifying workflow. Pick one denominator and use it in both variants. An add-on attach rate measured against eligible accounts cannot be compared with a bundle activation rate measured against weekly active users.

Next, separate revenue from cost:

• Price or uplift: the monthly amount the customer pays for the add-on or bundle tier, stated in the same currency and billing period.
• Incremental COGS: model calls, input and output tokens, retrieval, storage, tool calls, image or audio units, and any other customer-triggered delivery cost attributable to the feature.
• Support and payment cost: incremental support contacts, refunds, payment processing, and account-management work caused by the package.
• Usage distribution: p50, p90, and p99 feature consumption. A mean alone can hide the users who make a bundle unprofitable.
• Cannibalization: customers who would have bought a higher tier, an existing paid module, or another usage product without this package.
• Retention effect: accounts lost or downgraded because the bundle makes the plan too expensive, or because the add-on creates surprise or checkout friction.

For provider cost, retain the unit and date. OpenAI’s pricing page, checked on 2026-11-02, lists model rates per million input, cached-input, and output tokens and separately lists other meters such as voice, image, web search, and containers.^[1] Do not collapse those into a single “AI cost” number unless your workload actually uses only one meter. A price in a vendor table is a dated public list price, not a quote and not a guarantee of your account’s final invoice.

The cost boundary should also be consistent. If the add-on gets feature-level observability and the bundle gets none, the comparison is biased. FinOps guidance recommends tracking cost per inference, per user, and per successful outcome as distributions, then choosing the least expensive model that meets the required quality, latency, and reliability.^[5] Apply the same quality bar to both packaging tests.

Worked cost example: find the break-even attach rate

Assume a B2B SaaS product has 10,000 eligible paid accounts. The team is testing a document-drafting AI feature for one month. These are illustrative planning assumptions in USD, not vendor quotes:

• Add-on price: $12 per attached account/month.
• Expected incremental COGS for an attached account: $2.40/month.
• Add-on support and payment cost: $0.60 per attached account/month.
• Bundle price uplift: $6 per retained account/month.
• Expected bundle COGS: $1.56 per account/month, assuming 65% of accounts use the feature and each active account has $2.40 of feature COGS.
• Bundle support and payment cost: $0.39 per retained account/month.
• Bundle retention: 97%, or 9,700 retained accounts.
• Contribution at risk from a bundle-induced lost account: $30 per account/month. This is a business assumption used to price the retention downside, not a market benchmark.

The bundle produces $30,285 of modeled net monthly contribution. The add-on needs 30,285 ÷ $9 = 3,365 attached accounts to match it. Against 10,000 eligible accounts, the break-even attach rate is 33.65%, rounded to 33.7%. Below that rate, the bundle wins under these assumptions. Above it, the add-on wins, provided the attach rate does not come from unusually expensive power users and the add-on does not create a larger churn penalty.

This result is sensitive to the assumptions. At a 20% add-on attach rate, the add-on contributes $18,000 and trails the bundle. At 35%, it contributes $31,500 and slightly exceeds the bundle. At 50%, it contributes $45,000. If bundle retention rises from 97% to 99%, the bundle’s lost-account cost falls by $6,000 in this example, raising its net contribution to $36,285 and raising the add-on break-even attach rate to 40.3%.

The important output is not the single recommendation. It is the threshold your experiment must measure. Instrument impressions, eligibility, checkout starts, paid attachments, first successful AI action, repeat usage, COGS, support contacts, downgrade, and cancellation by cohort. Compare customers exposed to the same feature and quality level. If the package changes entitlement, model the value and cost differences explicitly rather than calling the result an adoption test.

A practical packaging test plan

First, define the feature boundary. Write down which requests, tools, storage, retries, and human-support paths count as incremental. The boundary should match your internal cost ledger and the customer promise. Microsoft’s Copilot documentation illustrates why this matters: it distinguishes fixed licensing from usage measured in Copilot Credits, and documents pre-purchase, pay-as-you-go, prepaid capacity, spending policies, alerts, and consumption monitoring.^[3]

Second, create a control and two treatments. The control keeps current packaging. Treatment A exposes the feature as a clearly labeled add-on. Treatment B includes the same feature in a higher tier or bundle. Avoid changing model quality, limits, onboarding, and support at the same time unless that is the packaging decision itself.

Third, pre-register guardrails. Set an acceptable p90 and p99 COGS per account, a maximum subsidy for non-users, a quality threshold, and a churn review window. A bundle that wins on month-one contribution but creates expensive renewals or support load is not a durable win. Likewise, an add-on with high initial contribution but low repeat use may be a one-time purchase effect.

Fourth, make the meter visible. Show included allowance, expected unit, current usage, and what happens at the limit. Stripe’s documentation identifies threshold monitoring and usage visibility as distinct operational capabilities, while Microsoft documents spending policies and user-level thresholds for Copilot Credits.^{[3] [4]} A customer should not have to reverse-engineer an invoice to understand the package.

Fifth, review the tail. Segment light, median, and heavy users. If the bundle’s p99 user creates negative contribution, add a fair-use policy, route simple requests to a cheaper model, cap burst behavior, or move excess usage to an overage path. OpenAI’s pricing page shows why model and meter mix matter: input, cached input, output, search, image, voice, and container usage have different units and rates.^[1]

When should an AI feature be a paid add-on?

Use a paid add-on when the feature has a distinct use case, a distinct buyer, or a usage distribution that differs sharply from the core product. It is also a useful first package when you need to learn willingness to pay without repricing every customer. Keep the entitlement clear. “AI included” and “advanced AI add-on” should describe materially different capability, volume, or service level, not an arbitrary paywall.

An add-on is weaker when the AI feature is required to complete the product’s advertised job. In that case, a separate charge can lower adoption or create a perception that the core product is incomplete. Test whether customers understand the feature as optional before optimizing the price.

When should AI be bundled into a plan?

Bundle AI when most customers in a plan should experience it, usage is bounded enough to forecast, and the price uplift remains acceptable for non-users. A higher tier is often cleaner than adding a small charge to every existing plan when the AI feature belongs with other premium controls, security, or workflow capacity.

Do not call the feature “free.” The cost is funded by the plan price, cross-subsidy, or margin. State the allowance and fair-use behavior, and watch the p90 and p99 account economics.

FAQ

Is an AI add-on always more profitable than bundling?

No. An add-on has a higher contribution per adopter in many designs, but it also has lower adoption and more purchase friction. Bundling can produce more total contribution when the price uplift is modest, retention is stable, and enough customers use the capability. Calculate both on the same eligible-account denominator and include retention, support, and heavy-user cost.

What attach rate makes an AI add-on worthwhile?

There is no universal attach-rate target. Compute it from the competing bundle: bundle net contribution ÷ eligible accounts ÷ add-on contribution per attached account. In the worked example, the bundle contributes $30,285 and the add-on contributes $9 per attached account, so the break-even attach rate is 33.7%. Replace every assumption with observed cohort data before making a launch decision.

Should AI usage be included in a bundle or metered separately?

A hybrid is often the safer starting point when adoption should be broad but usage is variable. Include a bounded allowance, show consumption, and meter only usage beyond the allowance. Keep the unit understandable and define the failure behavior at the limit. Stripe describes flat-rate subscriptions combined with usage billing, prepaid credits, meters, and monitoring as separate building blocks.^[4]

How should incremental AI COGS be calculated?

Start with provider-reported usage and preserve units: input and output tokens, cached tokens, requests, minutes, images, retrieval, storage, retries, and tools. Add directly attributable infrastructure, observability, support, payment, and platform costs where your accounting boundary requires them. FinOps recommends measuring cost per inference, per user, and per successful outcome, not relying on one blended average.^[5]

What should happen when a bundled user exceeds the allowance?

Choose the behavior before launch. Options include a soft fair-use review, throttling, an upgrade prompt, a paid overage, a credit top-up, or a hard stop. The right option depends on the workflow’s tolerance for interruption and your ability to forecast cost. Whatever you choose, show the allowance, alert before the limit, and record the event so finance can compare actual COGS with the packaging model.

Related calculations

Continue the estimate with AI Agent Cost Calculator | Model Full Monthly Spend, compare the model assumptions with AI API invoice reconciliation calculator, and check realized charges using AI observability ROI calculator.