Most AI feature has an end. The cost of getting it there honorably runs 12 to 22 percent of original build cost, and roughly two-thirds of enterprise AI budgets do not name a reserve for it. The unbudgeted sunset is one of the more predictable failure modes of 2026 AI engagements: the feature reaches end-of-economic-life, the off-ramp work is real, no one wants to pay for it, and the cost gets dragged out of operating margin over six months instead of running cleanly through a 90-day decommissioning window. Naming the off-ramp budget at contract time avoids the failure entirely.
This is a spoke under the AI project economics manifesto, which argues that evaluation cost has replaced feature cost as the unit of account. Off-ramp is what happens to that unit at end-of-life; the part most lifecycle models leave out.
Why AI feature sunsets cost more than SaaS feature sunsets
A traditional SaaS feature sunset is mechanically simple. The feature flag flips off. The endpoints return a deprecation notice. The associated database tables are archived per retention policy. Users are emailed. The work is real but bounded; typically 4 to 8 percent of the feature’s original build cost.
AI feature sunsets are structurally larger work for three reasons.
Output-data portability obligations. AI features generate output data; extracted entities, generated drafts, classified records, summarized documents; that users have stored downstream in their own systems. When the feature is removed, the user’s downstream system still references that output, often with metadata pointing back to the originating model and version. The portability obligation is real: users need a clean export of historical outputs with provenance metadata, formatted in a way their downstream system can continue to interpret. This is a deliverable, not an afterthought.
Eval sets, training data, and observability stacks have their own retirement cost. Each of these is an asset during the feature’s life. At sunset each must be archived per data-retention policy, unhooked from adjacent features that may share components, and decommissioned without leaving orphan jobs running. Eval-set retirement alone runs $4,000 to $15,000 on most enterprise features and is the single most-missed line item we see.
Agentic dependency cascades. AI features in 2026 are rarely standalone. Feature A’s output feeds Feature B’s tool catalog. Feature C’s classifier conditions Feature D’s routing. Removing Feature A breaks tool calls in adjacent features that nobody mapped explicitly. The off-ramp must include a dependency audit, a fallback strategy for downstream consumers, and a migration window long enough for those consumers to switch off the feature being retired.
These three factors push AI feature sunset cost from the 4-to-8 percent SaaS band into the 12-to-22 percent AI band. Engagements that try to retire AI features on SaaS budgets discover the gap mid-decommission and either over-run or leave the feature partly retired indefinitely.
The off-ramp cost model
The empirical cost lines for sunsetting a customer-facing AI feature in 2026:
| Off-ramp line | Cost band | Drivers |
|---|---|---|
| User notification and migration messaging | 1 to 3 percent | User base size, communication depth, languages |
| Output-data export tooling and provenance | 3 to 6 percent | Volume, format complexity, metadata depth |
| Model and infrastructure decommissioning | 2 to 4 percent | Self-hosted vs managed, observability stack |
| Eval-set retirement and labeling-vendor unwind | 1 to 3 percent | Stratification depth, vendor count |
| Agentic dependency audit and fallback | 2 to 4 percent | Number of consuming features, fallback complexity |
| Contractual unwind with vendors and integrators | 1 to 2 percent | Vendor count, contract notice periods |
| Contingency reserve for unexpected dependencies | 2 to 3 percent | Usually present; ratio scales with complexity |
Total defensible band: 12 to 22 percent of original build cost. A feature that cost $400,000 to build carries an off-ramp reserve of $48,000 to $88,000. A feature that cost $1.2 million carries an off-ramp reserve of $144,000 to $264,000.
The contingency line is not optional. Off-ramps reliably surface dependencies the team did not know existed; usage patterns nobody documented, integrations with internal tools that drifted from spec, third-party consumers reading endpoints that were rarely officially supported. The 2 to 3 percent contingency line is what funds the discovery of these dependencies and the work to handle them gracefully rather than ignoring them.
The 90-day deprecation playbook
Ninety days is the structurally defensible timeline for a customer-facing AI feature off-ramp in 2026. The work is divided into three thirty-day windows.
Days 1 to 30: Notification and export window. The decommissioning announcement goes out on day one with a clear end-date and a documented data-export path. Users get the export tooling immediately. Internal stakeholders; sales, support, product marketing; get talking points and migration support documentation. The eval set transitions to a frozen state; no further training data accumulates. The agentic dependency audit completes by day twenty and is published to the team.
Days 31 to 60: Reduced-availability migration window. New user enrollment closes on day thirty-one. Existing users continue to use the feature but receive in-app notification of the closing window. Downstream features running on this feature’s outputs migrate to fallbacks during this window with engineering support. The model serving infrastructure begins scale-down; the feature is supported but not invested in. Customer support ramps up to handle migration questions.
Days 61 to 90: Decommissioning and unwind. The feature endpoints transition to read-only on day sixty-one. Day seventy-five is final shutdown for active inference. The remaining fifteen days unwind contractual obligations, archive data per retention policy, decommission infrastructure, and close out the eval set and labeling-vendor relationships. Day ninety closes the off-ramp with a structured decommissioning report that documents what was archived where and what dependencies were handled.
This timeline produces a clean off-ramp. Faster timelines work for internal-only features with small user bases; sometimes thirty to sixty days. Slower timelines apply when regulated data retention or contractual notice periods force extended availability; sometimes 120 to 180 days. The 90-day default is the right starting point for most customer-facing features.
The structural reserve: 12 to 22 percent of build cost
Twelve to twenty-two percent of original build cost is the defensible reserve range. The drivers that move a project to the lower or upper end:
Lower end (12 to 14 percent): standalone feature with no agentic dependencies, no regulated data, small user base, single-vendor stack, simple output structure. These are typically internal MVP-to-deprecation tools that ran their economic life and are being retired cleanly.
Mid range (15 to 18 percent): customer-facing feature with moderate agentic integration, standard data classes, mid-size user base, two-to-three-vendor stack. Most enterprise AI features land here. The 15-to-18 percent reserve covers the work without leaving margin to absorb unexpected dependencies.
Upper end (19 to 22 percent): customer-facing feature with deep agentic integration, regulated data subject to retention rules, large user base, multi-vendor stack, complex output structure with metadata obligations, downstream consumers in third-party systems the team does not control. These off-ramps fund what amounts to a small product launch in reverse.
Naming the reserve at contract time is the operationally important step. A buyer that signed a $500,000 build contract with no off-ramp reserve discovers at sunset that the work costs $80,000 and either the buyer or the agency must absorb it. A buyer that signed a $580,000 contract with $80,000 named as a 90-day decommissioning reserve drawable on trigger has converted a finance surprise into a planned spend. The economics manifesto’s posture on this is direct: most cost line should be named, and off-ramp is one of the lines.
The triggers that should fire an off-ramp review
Off-ramp should not be a debated judgment call when the trigger fires. Three structural triggers convert end-of-life into a structured threshold review.
Monthly active users below the contracted floor for two consecutive quarters. Most customer-facing AI feature should ship with an MAU floor in the contract; the minimum user volume that justifies the operating cost. When MAU falls below the floor for two quarters, an off-ramp review fires automatically. Either the feature recovers, the contract amends the floor, or the off-ramp begins.
Eval-set accuracy regression below the contracted threshold for two consecutive quarters with no remediation path inside the project budget. When the feature can no longer hit its acceptance threshold and remediation would require unbudgeted work, the off-ramp is the structurally honest answer. Forcing the team to operate a regression-debt feature on shrinking margin is the alternative, and it is worse for both buyer and agency.
Model-vendor end-of-life on a foundational model the feature cannot economically migrate from. When the underlying model reaches vendor end-of-life and the migration cost to a successor exceeds the residual feature value, the off-ramp fires. This trigger is increasingly common in 2026 as model lifecycles compress and migration costs accumulate. The dynamics of model deprecation are detailed in why your AI project budget should have a model deprecation reserve.
Each trigger should be named in the original contract with the corresponding off-ramp reserve drawable upon firing. This converts end-of-life from a contested decision into a structured event with defined cost and defined timeline.
The most-missed line items
Across off-ramp budgets reviewed in 2026 engagements, four line items are consistently missed.
Eval-set retirement. The eval set was an asset during the feature’s life; labeled examples, stratified test cases, regression baselines, possibly third-party labeling vendor relationships. At sunset it must be archived, dependencies unhooked, and vendor relationships unwound. Empirically this line runs $4,000 to $15,000 and is missed by roughly two-thirds of off-ramp budgets we review.
Provenance metadata in output exports. Users need their historical outputs back with provenance; which model version produced them, when, against which inputs. Building the provenance export tooling is real engineering work, often $8,000 to $25,000, and it is consistently underestimated as “we’ll just dump the database.”
Downstream consumer migration support. Internal teams that built features against the deprecating feature need engineering support to migrate. This is not an external user concern; it is internal engineering time that must be funded explicitly or it gets pulled out of unrelated capacity. Empirically this runs 1 to 3 percent of original build cost.
Decommissioning report and post-mortem. A structured close-out; what was retired, what was archived, what dependencies surfaced, what to do differently next time; is a real deliverable that takes 5 to 10 person-days. Skipping it means the next AI feature off-ramp in the same organization repeats the same mistakes from scratch.
Internal versus customer-facing off-ramp
Internal AI tools carry a smaller off-ramp reserve; typically 6 to 10 percent of original build cost rather than 12 to 22. The user-migration and data-portability components are smaller because the user base is internal and the export obligations are governed by internal policy rather than customer contract.
The eval-set retirement, infrastructure decommissioning, and integration-unhooking work is proportionally similar. Internal tools that ship without an off-ramp reserve tend to leave dead AI tools running indefinitely because no one will fund their removal; the cost is invisible until someone audits the bill and discovers $30,000 a year of inference spend on a tool with three users. The internal off-ramp reserve, even at 6 to 10 percent, prevents this by funding the cleanup work as a planned event rather than a discretionary one.
The interaction with the AI project insurance line is worth naming. Both are lifecycle-aware reserves drawable on defined trigger events. The insurance line covers in-life incident cost; the off-ramp covers end-of-life decommissioning cost. Naming both as named line items rather than rolling them into general contingency converts AI feature lifecycle from a finance surprise into a structured cost curve.
Frequently asked questions
Should the off-ramp reserve be held by the buyer or the agency? By the buyer, drawable to the agency or successor on a defined trigger. Holding the reserve at the agency makes it look like agency profit on a successful project; holding it at the buyer makes it visible as a planned future spend.
Does the off-ramp reserve roll forward unspent if the feature has a longer life than expected? Yes. Each year the feature operates without firing an off-ramp trigger, the reserve refreshes against current build cost. A feature whose value compounded should carry a slightly larger reserve, not a smaller one; because its off-ramp will be more entangled.
How does the off-ramp interact with multi-year master service agreements? Cleanly. The off-ramp reserve becomes a named line in the MSA drawable on triggered events. Each statement of work names the percentage of build cost reserved against that SOW’s specific feature.
What happens to the off-ramp reserve if the buyer migrates the feature to a successor system rather than retiring it? A migration is a different event from a retirement. The migration carries its own line items; data portability, eval-set transfer, observability handoff; but the decommissioning components fire only against the originating feature, not the successor. Most engagements name migration as a separate triggerable event from off-ramp.
Should the off-ramp budget be reviewed quarterly or only at trigger events? Quarterly review is defensible. The reserve should track the feature’s current scope rather than the original SOW’s scope; features that grew during their life carry larger off-ramps than they did at signing.
How does the off-ramp interact with the sunk-cost trap? It defuses it. Engagements that named an off-ramp reserve treat sunset as a planned event rather than a defeat, which removes the emotional incentive to keep underperforming features alive past their economic life. The dynamics are documented in the AI project sunk-cost trap.
What governance change makes the off-ramp budget operational? Two changes. First, most AI feature SOW names an off-ramp percentage and the triggers that fire it. Second, the quarterly business review reads the off-ramp triggers as standing items, not exceptions. Together these convert off-ramp from emergency work into routine lifecycle management.
Does the off-ramp budget apply to AI features that ship as part of a larger product? Yes; proportionally. The off-ramp reserve is calculated against the build cost attributable to the AI feature, not the entire product. Engagements often miscalculate this by treating the AI feature as free-rider on the product budget; the off-ramp work is feature-specific and the reserve must be sized to it.
What is the right way to communicate the off-ramp budget to a board or buyer who has not seen one before? Frame it as the AI-feature equivalent of equipment depreciation. SaaS features have minimal end-of-life cost; AI features have meaningful end-of-life cost. The reserve is the structurally honest accounting for that fact.
How does the off-ramp interact with the AI project economics manifesto’s eval-as-unit-of-account principle? Tightly. The eval set is the asset created during build. Operating cost runs against it. End-of-life retirement cost runs against it. Naming many three converts the eval from an opaque cost center into a defined-lifecycle asset with a beginning, middle, and end.
Key takeaways
- Sunsetting an AI feature costs 12 to 22 percent of original build cost, structurally larger than equivalent SaaS deprecations because of output-data portability, eval-set retirement, and agentic dependency cascades.
- The 90-day decommissioning playbook divides the work into a notification window, a reduced-availability migration window, and a final unwind window; each thirty days.
- The reserve should be named at contract time, drawable on defined trigger events: MAU below floor for two quarters, eval regression below threshold with no remediation path, or model-vendor end-of-life.
- The most-missed line items are eval-set retirement, provenance metadata in exports, downstream consumer migration support, and the decommissioning report itself.
- Internal AI tools carry a smaller 6 to 10 percent reserve but skip it at the cost of dead tools running indefinitely on inference spend nobody audits.
- The off-ramp pairs with the AI project insurance line as the two named lifecycle reserves that convert AI features from open-ended spend into defined-lifecycle assets.
- Naming the off-ramp budget defuses the sunk-cost trap by making sunset a planned event rather than a defeat; buyers and agencies both benefit.
Arthur Wandzel
