The AI Project Cost-of-Delay Framework

Cost-of-delay (CoD) is the lean-product framework for valuing time itself, and it is more important for AI projects than for any other kind of software because AI features decay structurally over time. The original Don Reinertsen framework defines CoD as the marginal value lost per unit time of delay and combines it with job duration to produce Weighted Shortest Job First (WSJF) prioritization. The translation to AI work has three twists. First, AI features have unusually steep value decay because token prices drop 60 to 80 percent per year and frontier capability commoditizes the marginal-AI-feature niche on a 6 to 12 month cycle. Second, AI competitive timing windows are sharper than traditional software because the first product to ship a working AI feature in a category often captures 60 to 80 percent of the category’s eventual revenue. Third, AI delay cost compounds because a delayed launch misses the model-upgrade cycle and ends up shipping on a generation behind, which costs another 3 to 6 months of catch-up. This piece adapts CoD for AI projects, gives a quantified delay-cost model, and shows where to apply WSJF inside an AI roadmap.

This is a spoke under the AI project economics manifesto. The manifesto names evaluation cost as the unit of account; cost-of-delay is the temporal dimension that names the cost of not shipping the eval-cleared feature on time.

Why AI cost-of-delay is structurally steeper

The traditional cost-of-delay assumption is that a feature delayed by one month loses one month’s worth of revenue. The math is linear, the slope is gentle, and the conclusion is “ship sooner.” For AI features the slope is non-linear and steep because three forces compound during the delay window:

1. Token-price decay. OpenAI, Anthropic, and Google have shipped roughly 60 to 80 percent annual token-price drops since 2023. A six-month delay puts a feature into a market where the per-call cost has fallen 30 to 40 percent; meaning competitors can offer the same feature at lower price, or higher quality at the same price, by virtue of having shipped on the cheaper substrate.

2. Capability commoditization. Frontier model releases ship roughly quarterly. Each release commoditizes a layer of capability that was previously a moat. A feature whose moat is “we built clever prompting around GPT-4” becomes a feature whose moat is “we use GPT-5’s native capability that does this for free” within 6 to 9 months. Delay extends the window during which the moat is dissolving.

3. Compounding lateness. A delayed AI launch typically misses the model upgrade it was designed for. The team then spends 2 to 4 months porting prompts and re-running evals against the next-generation model; and ships against that model right as the next generation deprecates it. The compounding-lateness pattern is what makes 6-month AI delays often produce 9 to 12 months of effective lost time.

The result: AI cost-of-delay is typically 2 to 4x steeper than traditional software cost-of-delay for the same revenue magnitude. This is the structural reason why “ship a B-grade AI product fast” often beats “ship an A-grade AI product slow” on lifetime value.

The three components of AI delay cost

A defensible AI delay cost model has three additive components:

• Direct revenue delay (the traditional CoD line): The monthly revenue the feature will generate, prorated to the delay window. For a feature projected at $200K/month ARR contribution, a three-month delay costs $600K in foregone revenue plus the time-shifted compounding of that revenue across the product lifetime.
• Decay-adjusted opportunity cost: The per-call cost decay during the delay window. A 6-month delay at the historical 50 to 70 percent annual price decay rate means a competitor shipping today gets a 25 to 35 percent unit-cost advantage relative to the team that delays. On a feature with non-trivial inference cost, this is 5 to 15 percent of feature revenue annualized.
• Competitive position cost: The probability-weighted cost of being beaten to market by a competitor, including category-leader premium loss. For categories where the first mover captures 60 to 80 percent of share, this is the largest delay component; typically 30 to 60 percent of feature lifetime value if the delay is long enough that a competitor ships first.

Total AI delay cost on a representative feature is typically 2 to 5x the team’s intuition based on revenue-only thinking. The component that dominates depends on the category; in commoditizing categories the decay-adjusted line dominates; in winner-take-most categories the competitive position line dominates; in steady-demand B2B categories the direct revenue line dominates.

WSJF for AI roadmaps

Weighted Shortest Job First (WSJF) is the Reinertsen prioritization rule: rank work by (cost of delay ÷ job size). Highest ratio first. The rule is correct for traditional software and even more correct for AI software because the cost-of-delay numerator is structurally larger.

A worked WSJF table for a representative AI roadmap:

Feature	Monthly CoD (revenue + decay + competitive)	Job size (eng-months)	WSJF
Voice agent for sales	$180K	4	45
RAG over support docs	$90K	3	30
Eval suite expansion	$30K	1	30
Multi-modal input	$250K	8	31
Compliance audit feature	$40K	2	20

Voice agent ships first by WSJF; high CoD per engineering month. Multi-modal input has the highest absolute CoD but ranks third because its job size dwarfs the others. Compliance audit ranks last not because it’s unimportant but because the CoD per engineering month is the lowest.

The discipline that AI roadmaps usually lack is making the CoD numerator explicit. Most roadmaps rank by intuited importance, which under-weights time-sensitive features and over-weights features that look big in a demo. WSJF forces the question “how much does each delay month cost?”; and the AI version of that question has answers in the $100K to $500K per month range for any meaningful feature.

The pricing-models-by-alignment-with-outcomes piece covers how WSJF interacts with milestone-based contracts on the buyer side.

Token-price decay as a delay multiplier

Per-call token costs have dropped 60 to 80 percent annually since 2023. The decay rate matters for cost-of-delay because most month the team holds a feature back on the high-cost substrate is a month closer to a competitor shipping it on a 30 to 40 percent cheaper substrate.

A worked example. A feature that costs $0.40 per call to operate at GPT-4 prices in 2024 costs $0.12 per call by mid-2026 if the team ships on time and rides the decay. The same feature delayed 12 months arrives in a market where competitors are operating at $0.08 to $0.10 per call. The delayed team’s unit economics are 30 to 50 percent worse than the on-time team’s, even before considering that the on-time team has 12 months of usage data, prompt-tuning, and customer feedback that the delayed team does not.

The token-price-decay effect is asymmetric; the team that ships first gets the decay benefit through the entire lifetime of the feature; the team that ships second gets only the residual decay from their later start point. This is why “ship now and re-tune later” beats “wait for token prices to drop further before shipping” for almost many features. The waiting team forgoes both the time-in-market and the decay-leveraged unit economics.

The competitive timing window

AI categories have unusually sharp first-mover dynamics because:

• Distribution premium. The first product to ship a working version of an AI feature gets organic press, conference invitations, and analyst coverage that compound into customer acquisition advantages worth 6 to 18 months of paid acquisition.
• Data moat compounding. First-mover product accumulates production traffic that improves the eval set, feedback loop, and prompt registry. By the time a competitor ships, the first-mover has 3 to 9 months of real-world data that the competitor must catch up on.
• Talent gravity. First-mover AI products attract the engineers who know how to ship AI products. The competitor’s hiring pool 6 months later is materially smaller and more expensive.
• Buyer-perception lock-in. Enterprise buyers who have evaluated and bought the first product are unlikely to re-evaluate within 18 months. The competitive window for the second product is the residual buyer pool that hasn’t yet bought, not the full market.

The cumulative effect: in many AI categories the first mover captures 60 to 80 percent of the eventual market because the first 12 to 18 months of compounding produce a position that subsequent competitors cannot easily dislodge. The cost of being second is therefore not “we get 50 percent share instead of 60 percent”; it is “we get 20 to 40 percent share instead of 60 to 80 percent.”

This is why the competitive position component of cost-of-delay is often the largest line on AI features. The delay that puts a team behind a competitor’s launch is rarely worth the quality improvement it bought.

Quantifying delay cost on a real project

A worked example. Team is debating shipping a 75 percent-complete AI customer support automation feature versus delaying 90 days for additional eval coverage. The feature is projected at $150K/month ARR contribution at full ramp, the category has one credible competitor 4 months behind, and the inference unit cost is $0.18 per resolved ticket.

• Direct revenue delay (90 days): $450K in foregone revenue.
• Decay-adjusted opportunity cost (90 days): Token prices fall ~12 to 15 percent in 90 days. On 4M projected tickets/year at $0.18 each, the team forfeits roughly $90K to $110K in unit-cost advantage versus a competitor shipping on the cheaper substrate during the delay.
• Competitive position cost: 30 percent probability the competitor ships during the 90-day window. Lifetime value loss conditional on losing first-mover position is roughly $4M to $6M (40-60 percent of category share). Probability-weighted: $1.2M to $1.8M.
• Total delay cost: $1.7M to $2.4M for 90 days.

The team’s instinct is to consider only the $450K direct revenue line and conclude the delay is “worth it for quality.” The full CoD model shows the delay is 4 to 5x more expensive than the team intuited. The decision to delay should require a quality argument worth $1.7M+ in lifetime value; usually the answer is to ship and improve quality post-launch.

The staged-payback gate logic inside the manifesto gives the structure for shipping fast and improving fast; eval-feedback at 90 days, capability at 12 months, compounding at 24 months.

What to ship slow versus what to ship fast

Not most AI feature should be optimized for shipping speed. The high-CoD-ship-fast pattern applies cleanly to features where:

• Direct revenue contribution is high.
• The competitive timing window is open.
• Inference unit cost is meaningful (so token-price decay matters).
• The eval suite already covers 80 percent of the use case.

The opposite; ship-slow patterns; apply to features where the CoD math is different:

• Compliance and safety features. External failure cost (regulatory, brand) is the dominant economic line; CoD is small relative to the cost of getting it wrong.
• Foundational eval and observability work. Job size is small and the work compounds across many subsequent features; CoD-as-prerequisite is the right framing rather than CoD-as-product-feature.
• Highly differentiated quality plays. Features where the buyer values 95 percent quality over 75 percent quality by 5x or more; usually high-stakes verticals like medical or legal where a wrong answer is catastrophic.

The discipline is to make the ship-fast-versus-ship-slow decision explicit and CoD-quantified rather than intuited. Most AI roadmaps default to ship-slow without measuring the delay cost; most teams that adopt CoD discipline find that 60 to 80 percent of their roadmap belongs in the ship-fast bucket.

Frequently asked questions

What is cost-of-delay and why does it matter for AI?

Cost-of-delay (CoD) is the marginal value lost per unit time of delay on a given feature or project. For AI work, CoD is structurally steeper than for traditional software because three forces compound during delay: token-price decay (60 to 80 percent per year), capability commoditization (frontier model releases dissolve moats most 6 to 9 months), and compounding lateness (delayed AI features miss model-upgrade cycles). Total AI delay cost is typically 2 to 4x what teams intuit from revenue-only thinking.

What’s the right way to estimate AI delay cost?

Three additive components: direct revenue delay (foregone monthly revenue prorated to delay window), decay-adjusted opportunity cost (the per-call cost advantage you forfeit to competitors shipping on cheaper substrate during your delay), and competitive position cost (probability-weighted lifetime value loss from being beaten to market). On a representative feature these sum to 2 to 5x the team’s intuition based on revenue alone.

How does WSJF apply to AI roadmaps?

WSJF (Weighted Shortest Job First) ranks work by cost-of-delay divided by job size. Highest ratio ships first. For AI roadmaps the discipline is to make the CoD numerator explicit; most roadmaps rank by intuited importance, which under-weights time-sensitive features and over-weights features that look big in a demo. Worked through a representative roadmap, WSJF typically reorders 30 to 50 percent of the planned sequence.

Why does token-price decay multiply delay cost?

Token prices have dropped 60 to 80 percent annually since 2023. A feature delayed 6 months arrives in a market where competitors are operating at 25 to 35 percent lower per-call cost. The on-time team gets the decay benefit through the entire feature lifetime; the delayed team gets only residual decay from their later start point. The asymmetry compounds over the feature’s life and is why “ship now and re-tune later” beats “wait for token prices to drop” for most features.

How big is the first-mover advantage in AI?

In many AI categories, the first product to ship a working version captures 60 to 80 percent of the eventual market. Distribution premium, data moat compounding, talent gravity, and buyer-perception lock-in compound during the first 12 to 18 months in market. Being second therefore means 20 to 40 percent share rather than 50 percent; a much larger penalty than in traditional software categories.

When should I ship slow rather than fast on AI?

When external failure cost dominates revenue (compliance, safety, high-stakes verticals), when the work compounds across many future features (eval suite, observability), or when buyers genuinely value 95 percent quality over 75 percent quality by 5x or more (medical, legal, financial). For most other AI features, ship fast and improve fast through the staged-payback structure: 90-day eval-feedback gate, 12-month capability gate, 24-month compounding gate.

How does AI cost-of-delay differ from traditional software CoD?

Three structural differences. First, the slope is 2 to 4x steeper because of token-price decay, capability commoditization, and compounding lateness. Second, the competitive position component is larger because AI categories have sharper first-mover dynamics than traditional software categories. Third, delay risk is asymmetric; shipping fast and tuning rarely loses, shipping slow and missing the model-upgrade cycle often loses double the projected delay window.

What’s the relationship between cost-of-delay and cost-of-quality?

Cost-of-delay names the cost of not-yet-shipping. Cost-of-quality names the cost of shipping with insufficient quality. The two frameworks together produce the right tension: ship fast (CoD) but with sufficient prevention spend (CoQ) that external failure does not consume the gains. Healthy AI projects optimize both; they ship faster than competitors and have lower external failure rates because their prevention infrastructure is robust.

Does cost-of-delay apply to internal AI projects without external competitors?

Yes, with one component substituted. Internal projects do not have a competitive position cost line, but they do have an opportunity cost line; the alternative AI feature that the engineering team is not building during the delay. On a constrained AI engineering team, most quarter spent on a delayed project is a quarter not spent on the next-best project. The opportunity-cost line typically replaces 50 to 70 percent of the competitive-position line in magnitude.

How do I make the case for shipping faster to a quality-focused team?

Quantify the delay cost. Most quality-focused teams under-estimate delay cost by 3 to 5x because they consider only the foregone revenue line. Walking through the three-component CoD model on a specific feature usually moves the conversation from “we should delay for quality” to “the delay costs $1.7M and the quality improvement is worth $200K; let’s ship and tune.” The conversation works when the delay cost is concrete and quantified rather than vibey.

Key takeaways

• AI cost-of-delay is structurally 2 to 4x steeper than traditional software CoD because token prices decay 60 to 80 percent per year, frontier capability commoditizes on a 6 to 9 month cycle, and delayed launches miss model-upgrade cycles.
• A defensible AI delay cost model is three components: direct revenue delay, decay-adjusted opportunity cost, competitive position cost. On a representative feature these sum to 2 to 5x what teams intuit from revenue-only thinking.
• WSJF prioritization (CoD ÷ job size) reorders 30 to 50 percent of the typical AI roadmap when the CoD numerator is made explicit. Voice and high-revenue features ship sooner; large-job-size features that ranked high on intuition ship later.
• First-mover advantage in AI categories is sharp; 60 to 80 percent of the eventual market often goes to the first credible product. The competitive position component of CoD is therefore often the largest line.
• Ship fast applies to most AI features. Ship slow applies to compliance, safety, foundational eval/observability, and high-stakes quality verticals where 95 percent quality is worth 5x more than 75 percent.

Cost-of-delay is the framework that turns “we should ship sooner” into “this delay costs $1.7M and the quality improvement is worth $200K.” Most AI projects in 2026 under-account for delay cost by 3 to 5x and ship slower than the economics warrant. The quantified CoD model is how to fix it before the competitor does.