# AI Pricing Structures and ROI Maths

The three AI pricing structures AIB-C01 names, commitment-based discounts and when they destroy value, verified Bedrock price points with worked monthly costs, Pricing Calculator vs Cost Explorer vs Marketplace, a full ROI worksheet with a three-year worked case, unit economics, and the four ways AI business cases overstate benefit.

import { Tabs, TabItem } from '@prosefly/astro-components';

This is the reference for the money side of the [AWS Certified AI Business Strategist (AIB-C01)](/aws/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.

| Structure | Cost driver | Predictability | Failure mode | The question to ask |
| --- | --- | --- | --- | --- |
| **Consumption-based** | Usage: per token, per request, per page, per image | Low — scales with usage, no floor | A viral spike or a runaway agent produces a runaway bill | "What is the cost per unit of work, and what caps runaway usage?" |
| **Instance-based** | Hours of provisioned compute | Medium — predictable while running | You pay for **idle** capacity; endpoints left running burn money | "What is the utilisation, and do we shut it down when idle?" |
| **Seat-based** | Users licensed (per user per month) | High — flat and predictable | You pay for seats that no one uses; cost is **decoupled from value** | "How many seats deliver real value, and how do we measure per-seat value?" |

```text
  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
```

:::tip[Exam signal]
"Unpredictable spikes" / "per request" → consumption. "Pay while it runs" / "idle" → instance-based. "Per user per month" / "licences" → seat-based. The correct answer usually names the *failure mode* of the structure in the stem, not a cheaper vendor.
:::

## 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.

```text
  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**.

<Tabs>
<TabItem label="Guardrails">

| Component | Price (per 1,000 text units) |
| --- | --- |
| Content filters | $0.15 |
| Denied topics | $0.15 |
| Sensitive-information filter (PII) | $0.10 |
| Contextual grounding check | $0.10 |
| Automated Reasoning check | $0.17 per policy |
| Regex and word filters | Free |
| Image content | $0.00075 per image |

AWS states Guardrails blocks up to **88% of harmful content**.

**Worked monthly cost.** A support assistant processes **2,000,000 text units/month** and applies content filters + sensitive-information filter + contextual grounding:

- Per-1,000-unit rate = $0.15 + $0.10 + $0.10 = **$0.35**
- Units, in thousands = 2,000,000 / 1,000 = **2,000**
- Monthly cost = 2,000 × $0.35 = **$700/month**

Add word filters and it stays $700 (word filters are free). Add one Automated Reasoning policy over the same volume: 2,000 × $0.17 = **$340**, so **$1,040/month** total.

</TabItem>
<TabItem label="Knowledge Bases">

| Component | Price |
| --- | --- |
| Index storage | $5.00 per GB of raw data per month |
| Standard Retrieval | $1.00 per 1,000 API calls |
| Agentic Retrieval | $4.00 per 1,000 calls **plus** $1.00 per 1,000 underlying retrieve calls |

Managed parsing, embeddings and re-ranking are included at no extra charge.

**Worked monthly cost.** A knowledge base holds **20 GB** of raw data and serves **500,000 Standard Retrieval calls/month**:

- Storage = 20 × $5.00 = **$100**
- Retrieval = (500,000 / 1,000) × $1.00 = 500 × $1.00 = **$500**
- Monthly cost = **$600/month**

If those 500,000 calls were **Agentic Retrieval** with one underlying retrieve call each: (500 × $4.00) + (500 × $1.00) = $2,000 + $500 = **$2,500** + $100 storage = **$2,600/month**. Agentic retrieval is materially more expensive — a strategist matches it to genuinely agentic use, not routine lookups.

</TabItem>
<TabItem label="Evaluation and routing">

| Service | Price |
| --- | --- |
| Model Evaluation — algorithmic scores | No extra charge |
| Model Evaluation — human evaluation | $0.21 per completed human task (plus inference) |
| Model Evaluation — LLM-as-a-judge / RAG eval | Billed as token usage |
| Intelligent Prompt Routing | $1 per 1,000 requests (up to 30% cost reduction claimed) |
| Prompt Optimization | $0.03 per 1,000 tokens (simple optimizer) |

**Worked costs.**

- **Human evaluation** of 1,200 completed tasks = 1,200 × $0.21 = **$252** (plus the inference to generate the responses being judged).
- **Intelligent Prompt Routing** over 3,000,000 requests/month = (3,000,000 / 1,000) × $1 = 3,000 × $1 = **$3,000/month** in routing fees — worth it only if the routing saves more than $3,000 in model cost. AWS's up-to-30% claim on a $20,000/month model bill would save ~$6,000, netting ~$3,000.
- **Prompt Optimization** over 10,000,000 tokens/month = (10,000,000 / 1,000) × $0.03 = 10,000 × $0.03 = **$300/month**.

</TabItem>
</Tabs>

:::note[Guardrails, Knowledge Bases and routing are *add-ons* to model cost]
Every figure above is *on top of* the inference cost of the model itself. A business case that budgets only the model tokens and forgets the guardrail, retrieval and evaluation layers understates cost — one of the four ways cases go wrong (below).
:::

## Three tools: Pricing Calculator vs Cost Explorer vs Marketplace

| Tool | What it is | Good for | Not good for |
| --- | --- | --- | --- |
| **AWS Pricing Calculator** | Forward estimator (`calculator.aws`) | *Before* you commit: modelling a forecast, comparing scenarios, sizing a budget | Telling you what you actually spent |
| **AWS Cost Explorer** | Actuals, trends and anomaly investigation | *After* you run: tracking real spend, spotting a runaway, attributing cost | Estimating a workload you have not run yet |
| **AWS Marketplace** | Catalogue of third-party software and models | Evaluating **buy/partner** options against building yourself | Estimating your own consumption cost |

```text
  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?"
```

:::tip[Exam signal]
"Forecast the budget" → Pricing Calculator. "Investigate why the bill jumped" / "track actual spend" → Cost Explorer. "Evaluate a vendor / buy vs build" → Marketplace. Swapping these is a common distractor: Cost Explorer cannot forecast a workload you have not run.
:::

## 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.**

| Benefit | What it means | How to quantify |
| --- | --- | --- |
| Time savings | Hours returned to staff | hours saved × loaded hourly cost |
| Cost reduction | Direct spend removed | old cost − new cost |
| Revenue growth | New or retained revenue | incremental revenue attributable to AI |
| Productivity gains | More output per person | extra output × value per unit |

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:

| Cost | Typical size | Why it is missed |
| --- | --- | --- |
| Platform | The model/service bill | Everyone counts this |
| Integration | Connecting AI to existing systems | Under-scoped |
| Change management | Training, comms, process redesign | Treated as free |
| Oversight | Human-in-the-loop review time | Forgotten — but see governance by design |
| Ongoing monitoring | Drift/bias monitoring, evaluation | Assumed one-off, actually recurring |

### 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.**

| Cost | Year 1 | Year 2 | Year 3 |
| --- | --- | --- | --- |
| Platform (inference + Guardrails + Knowledge Bases) | $150,000 | $150,000 | $150,000 |
| Integration (one-off) | $120,000 | $0 | $0 |
| Change management | $60,000 | $20,000 | $20,000 |
| Oversight (human review) | $50,000 | $50,000 | $50,000 |
| Ongoing monitoring | $30,000 | $30,000 | $30,000 |
| **Total cost** | **$410,000** | **$250,000** | **$250,000** |

**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):

| Deflection rate | Deflection saving | Annual benefit | Year-1 net | 3-yr cumulative net |
| --- | --- | --- | --- | --- |
| 15% | 30,000 × $6 = $180,000 | $348,000 | −$62,000 | 348,000×3 − 910,000 = **$134,000** |
| 30% (base) | $360,000 | $528,000 | +$118,000 | **$674,000** |
| 45% | 90,000 × $6 = $540,000 | $708,000 | +$298,000 | 708,000×3 − 910,000 = **$1,214,000** |

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.

| Metric | Baseline | AI (at pilot volume) | Reading |
| --- | --- | --- | --- |
| Cost per resolved ticket | $6.00 | $3.90 | 35% cheaper per unit — scaling *helps* |
| Cost per document extracted | $2.50 | $2.80 | *More* expensive per unit — do **not** scale yet |
| Cost per qualified lead | $40 | $18 | Less than half — strong case to scale |

**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.

:::tip[Exam signal]
"Should we scale the pilot?" → look at **unit cost**, not total spend. A stem that shows total cost rising with volume is a trap; the discriminator is whether *cost per unit of work* is below the baseline.
:::

## The four ways AI business cases overstate benefit

| Overstatement | What it looks like | The correction |
| --- | --- | --- |
| **No baseline** | "It saves 40%" — of what? | Measure the before-state first (task 2.2.2); a percentage with no baseline is a guess |
| **Hidden costs** | Only the model bill is counted | Add integration, change management, oversight and ongoing monitoring — the recurring ones especially |
| **Optimistic adoption** | 100% of staff use it from day one | Run a sensitivity on adoption rate; gate scale-up on hitting the assumed rate |
| **Attribution creep** | Every good outcome credited to AI | Isolate the AI-attributable share; other initiatives and market moves also move the numbers |

## 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**.