AI Pricing Structures and ROI Maths

This is the reference for the money side of the AWS Certified AI Business Strategist (AIB-C01) exam: how AI is priced, how to build an ROI case, and how to spot a case that is fooling you. Every price point is verified against AWS’s published Bedrock pricing as of 15 September 2026, and every sum on this page is recomputed from those figures. The exam does not ask you to recall a per-token rate; it asks you to reason about cost drivers, commitments, ROI and unit economics — so that is what this page trains.

The three pricing structures

AWS names three structures. The exam tests whether you can identify a structure from a scenario and name its failure mode.

  spend ▲
        │ consumption   ╱  (rises with every request — no ceiling)
        │             ╱
        │ instance   ────────  (flat while running, waste when idle)
        │ seat       ════════  (flat per licensed user, value may be zero)
        └──────────────────────────────────────────▶ usage volume
  match the structure to the traffic shape: spiky → consumption risk;
  steady high volume → commitment; people-hours → seats

Commitment-based discounts — and when they destroy value

You can trade flexibility for a lower rate: Savings Plans, Provisioned Throughput commitments (1- or 6-month, bought in model units), and the Reserved Bedrock tier. All follow the same logic — commit to a baseline of usage, pay less per unit.

  a commitment pays off only above a break-even utilisation:

  total cost │  on-demand  ╱
             │           ╱
             │         ╱
   committed │────────●───────  commitment (fixed) + overage
             │       ╱ break-even
             └──────────────────────▶ actual usage
   below the ● you paid for capacity you never used —
   the commitment destroyed value

A commitment destroys value when actual usage sits below the break-even point: you have converted a variable cost you could have avoided into a fixed cost you cannot. The trap on the exam is a stem where usage is uncertain, new or seasonal and an option recommends a 6-month commitment “to save money” — the strategist answer is to stay on-demand until the volume is proven, then commit.

Worked break-even. Suppose an on-demand workload costs about $10,000/month and a 6-month Provisioned Throughput commitment costs $48,000 for the six months (a flat $8,000/month equivalent) but only covers a fixed capacity. Over six months, on-demand at steady volume = 6 × $10,000 = $60,000; the commitment = $48,000 — a $12,000 saving if usage holds. But if a reorganisation cuts usage to a third after month two, you still owe the full $48,000 while your on-demand cost would have fallen to roughly $10,000 + $10,000 + 4 × $3,333 ≈ $33,333. The commitment now costs $14,667 more than staying flexible. Same lever, opposite outcome — driven entirely by the utilisation assumption.

Bedrock price points with worked monthly costs

Every figure below is from AWS’s published Bedrock pricing; the monthly totals are computed for the stated volume. A text unit for Guardrails is up to 1,000 characters, and Guardrails is priced per 1,000 text units.

Three tools: Pricing Calculator vs Cost Explorer vs Marketplace

  BEFORE build        DURING/AFTER run       BUY-vs-BUILD
  ─────────────       ──────────────────     ────────────
  Pricing Calculator  Cost Explorer          Marketplace
  "what will it cost?" "what did it cost?"    "should we buy instead?"

The ROI worksheet

A defensible AI business case has four parts: a baseline, benefits, costs, and a time horizon over which you net them.

1. Baseline (task 2.2.2 — before implementation). You cannot claim improvement you cannot measure against a starting point. Capture the current cost, time, volume and quality before you deploy. No baseline, no ROI claim.

2. Benefit categories.

Time savings and cost reduction are tangible; customer satisfaction and productivity often start intangible — quantify what you can and label the rest honestly.

3. Cost categories — the ones cases forget are the last four:

Worked three-year case — Northwind Support

Northwind Support handles 200,000 support tickets a year. Baseline: each ticket costs $6.00 in agent time (200,000 × $6.00 = $1,200,000/year). A Bedrock-based assistant with Knowledge Bases and Guardrails is expected to deflect 30% of tickets fully and cut handling time on the rest.

Benefits per year (steady state, from year 1 for simplicity of the illustration):

• Deflection: 30% × 200,000 = 60,000 tickets × $6.00 = $360,000 saved.
• Faster handling on the remaining 140,000: 20% time cut × 140,000 × $6.00 = $168,000 saved.
• Annual benefit = 360,000 + 168,000 = $528,000/year.

Costs.

Net benefit per year: Year 1 = 528,000 − 410,000 = $118,000; Year 2 = 528,000 − 250,000 = $278,000; Year 3 = $278,000.

Cumulative net benefit: Year 1 = $118,000; Year 2 = 118,000 + 278,000 = $396,000; Year 3 = 396,000 + 278,000 = $674,000.

Payback period. Cumulative net benefit turns positive during Year 1 (net +$118,000 at year end), so payback is under 12 months. Within the year, monthly net benefit ≈ (528,000 − 410,000)/12 ≈ $9,833, but the $120,000 integration cost is front-loaded; treating benefits as accruing evenly and the one-off cost at the start, the cumulative turns positive at roughly month 9. Three-year ROI = 674,000 / (410,000 + 250,000 + 250,000) = 674,000 / 910,000 ≈ 74%.

Sensitivity on adoption rate. The 30% deflection assumption is the fragile one. Recompute annual benefit at other rates (faster-handling saving held at $168,000):

At 15% adoption the case still clears over three years (+$134,000) but loses money in year 1 (−$62,000) — which tells the strategist to phase the commitment and gate scale-up on hitting the deflection target, not to bet the full spend on the optimistic number.

Unit economics — why unit cost beats total cost

When you are deciding whether to scale, total cost misleads because it moves with volume. Unit cost — cost per resolved ticket, per document processed, per lead qualified — tells you whether scaling improves or worsens the economics.

Worked unit cost. Northwind’s assistant resolves 60,000 tickets/year for a platform+oversight+monitoring cost of about $230,000/year (the recurring portion): $230,000 / 60,000 ≈ $3.83 per resolved ticket, versus the $6.00 baseline. Because unit cost falls below baseline, more volume makes the case stronger — the green light to scale. If unit cost had come out above baseline, total savings from a bigger pilot would be an illusion: you would be scaling a loss.

The four ways AI business cases overstate benefit

Key takeaways

• Three pricing structures: consumption-based (usage-driven, no ceiling), instance-based (pay while running, idle waste), seat-based (per user, value-decoupled). Name the failure mode of the one in the stem.
• Commitments (Savings Plans, Provisioned Throughput, Reserved) lower unit cost only above a break-even utilisation; on uncertain or new volume they destroy value — stay on-demand until volume is proven.
• Verified Bedrock add-on costs sit on top of model cost: Guardrails at $0.35 per 1,000 text units for a three-filter setup = $700/month at 2M units; a 20 GB Knowledge Base with 500k Standard Retrievals = $600/month; routing over 3M requests = $3,000/month; prompt optimisation over 10M tokens = $300/month.
• Pricing Calculator forecasts, Cost Explorer reports actuals, Marketplace evaluates buy/partner — do not swap them.
• A defensible ROI case needs a baseline, honest benefit and cost categories (including oversight and monitoring), and a sensitivity on adoption. The Northwind case nets $674,000 over three years at 30% deflection, ~74% three-year ROI, payback under a year — but goes negative in year 1 at 15% adoption.
• When deciding to scale, judge on unit cost (per ticket, per document, per lead), not total cost.
• AI cases most often overstate benefit through no baseline, hidden costs, optimistic adoption and attribution creep.