Financial Modelling for AI: From Beginner Fundamentals to Advanced AI Economics

July 29, 2026 · 42 min read

AI Playbook author

Financial modelling is the process of translating a business idea, investment, product, project, or company into numbers.

A financial model helps decision-makers answer questions such as:

How much will the AI solution cost?
How will the solution generate financial value?
When will the investment break even?
How much cash will be required?
What happens if adoption is slower than expected?
Is it cheaper to build, buy, or partner?
Which AI architecture provides the best combination of cost, quality, latency, and risk?
What is the company or AI product worth?
Should the organisation approve, delay, redesign, or reject the investment?

For an ordinary software project, financial modelling usually connects customers, prices, employees, and infrastructure costs.

For an AI solution, the model must connect several additional variables:

\text{Business Demand} \rightarrow \text{AI Usage} \rightarrow \text{Model Performance} \rightarrow \text{Technical Cost} \rightarrow \text{Business Outcome} \rightarrow \text{Financial Value}

For example:

\text{Customer enquiries} \rightarrow \text{AI conversations} \rightarrow \text{Model calls and tokens} \rightarrow \text{Resolved cases} \rightarrow \text{Reduced contact-centre cost}

The most important principle is:

Do not measure only cost per token, request, model call, or GPU hour. Measure cost and value per successful business outcome.

Examples of meaningful AI financial units include:

Cost per successfully resolved customer enquiry
Cost per approved insurance claim
Cost per qualified sales lead
Cost per completed legal review
Cost per detected fraud case
Cost per accurate document extraction
Cost per software feature delivered
Cost per clinical document summarised
Revenue per AI-assisted customer
Gross profit per AI agent session

Cloud FinOps applies the same fundamental relationship to AI as to other cloud services:

\text{Cost} = \text{Price} \times \text{Quantity}

However, AI introduces unusual quantities such as input tokens, output tokens, model calls, GPU time, retrieval operations, evaluation runs, agent steps, tool calls, and human-review events. FinOps therefore recommends connecting cloud costs to business-unit economics rather than viewing infrastructure spending in isolation.

Part I: Financial Modelling Fundamentals

2. What type of AI financial model are you building?

Before opening Excel, define the decision the model must support.

There are several common types of AI financial models.

2.1 Enterprise AI business-case model

This model determines whether an organisation should invest in an AI solution.

Examples:

A bank considering an AI customer-service assistant
A manufacturer implementing predictive maintenance
A retailer introducing AI recommendations
A professional-services firm deploying an internal knowledge assistant
An insurer automating claims processing

The model normally calculates:

Implementation investment
Operating costs
Cost savings
Productivity benefits
Revenue uplift
Risk reduction
Net present value
Internal rate of return
Payback period
Scenario outcomes

2.2 AI product or SaaS model

This model forecasts the commercial performance of a company selling an AI product.

Examples:

AI customer-service SaaS
AI coding assistant
AI document-processing platform
AI compliance platform
AI agent marketplace

It normally includes:

Customer acquisition
Subscription revenue
Usage revenue
Implementation revenue
Customer churn
Expansion revenue
AI inference cost
Cloud infrastructure
Customer support
Sales and marketing
Research and development
Cash burn
Runway
Company valuation

2.3 AI unit-economics model

This model determines whether each unit of AI activity is economically sustainable.

For example:

\text{Contribution per AI task} = \text{Revenue per task} - \text{Variable cost per task}

Or, for an internal solution:

\text{Net value per task} = \text{Avoided business cost} - \text{AI cost} - \text{Human-review cost}

2.4 Build-versus-buy model

This compares options such as:

Build an internal AI platform
Buy an enterprise AI product
Use a managed foundation-model API
Host an open-weight model
Fine-tune an existing model
Train a model from scratch
Use a systems-integrator or consulting partner
Develop a hybrid solution

2.5 AI infrastructure and capacity model

This is more technically focused. It considers:

Pay-as-you-go inference
Reserved or provisioned throughput
GPU leasing
GPU ownership
Serverless inference
Batch inference
Model routing
Autoscaling
Utilisation
Latency requirements
Data-transfer costs
Disaster-recovery capacity

2.6 AI company valuation model

This estimates what an AI company or product may be worth using:

Discounted cash flow
Comparable-company multiples
Precedent transactions
Venture-capital method
Revenue multiples
Gross-profit multiples
Scenario-weighted valuation
Real-options valuation

A good model can contain several of these perspectives, but it should still have one clearly defined primary decision.

3. Essential financial concepts

3.1 Revenue

Revenue is the amount earned from selling a product or service.

For an AI business:

\text{Revenue} = \text{Subscription Revenue} + \text{Usage Revenue} + \text{Implementation Revenue} + \text{Support Revenue} + \text{Other Revenue}

Example:

An AI platform has:

100 customers
Annual subscription of $50,000
Average annual usage revenue of $12,000 per customer
$10,000 implementation fee for 30 new customers

Then:

\text{Subscription Revenue} = 100 \times \$50{,}000 = \$5{,}000{,}000

\text{Usage Revenue} = 100 \times \$12{,}000 = \$1{,}200{,}000

\text{Implementation Revenue} = 30 \times \$10{,}000 = \$300{,}000

\text{Total Revenue} = \$6{,}500{,}000

3.2 Cost of goods sold

Cost of goods sold, or COGS, includes the direct costs required to deliver the product or service.

For an AI product, COGS may include:

Foundation-model API charges
GPU inference
Embedding generation
Vector-database usage
Reranking
AI search
Speech-to-text and text-to-speech
Guardrail and moderation services
Observability directly attributable to customers
Customer-specific hosting
Human review
Third-party data
Customer onboarding labour
Direct customer-support labour

A basic formula is:

\text{Gross Profit} = \text{Revenue} - \text{COGS}

\text{Gross Margin} = \frac{\text{Gross Profit}}{\text{Revenue}}

Example:

\text{Revenue} = \$6.5\text{ million},\quad \text{COGS} = \$2.0\text{ million}

\text{Gross Profit} = \$6.5\text{m} - \$2.0\text{m} = \$4.5\text{m}

\text{Gross Margin} = \frac{\$4.5\text{m}}{\$6.5\text{m}} = 69.2\%

3.3 Operating expenses

Operating expenses are the costs of running the wider business.

Examples include:

Engineering salaries
Product management
Finance
Human resources
Legal
Sales
Marketing
Research
Corporate cloud environments
Office costs
Insurance
Audit
General administration

These costs are not necessarily directly attributable to one customer transaction.

3.4 EBITDA

EBITDA means earnings before interest, tax, depreciation, and amortisation.

A simplified formula is:

\text{EBITDA} = \text{Revenue} - \text{COGS} - \text{Operating Expenses}

Example:

\text{Gross Profit} = \$4.5\text{m},\quad \text{Operating Expenses} = \$3.2\text{m}

\text{EBITDA} = \$4.5\text{m} - \$3.2\text{m} = \$1.3\text{m}

EBITDA is useful, but it is not the same as cash flow.

An AI company can report positive EBITDA while still consuming cash because of:

Capitalised development expenditure
Hardware purchases
Slow customer collections
Debt repayments
Tax payments
Prepaid infrastructure contracts
Expansion investments

3.5 Capital expenditure and operating expenditure

Capital expenditure

Capital expenditure, or CapEx, is money invested in assets expected to provide benefits over multiple periods.

Examples could include:

Purchased GPU servers
Data-centre equipment
Certain capitalisable software-development costs
Long-term internal platforms
Purchased licences with qualifying characteristics

Operating expenditure

Operating expenditure, or OpEx, is generally recognised as an expense in the period it is consumed.

Examples include:

Monthly API charges
Cloud hosting
Employee salaries
Monitoring subscriptions
External model subscriptions
Data subscriptions
Routine maintenance

Accounting treatment is not simply a management choice. Under IAS 38, research expenditure is expensed, while qualifying development expenditure may be recognised as an intangible asset when the relevant criteria are satisfied.

For US GAAP reporting, FASB issued updated internal-use software guidance in 2025. The amendments use an authorisation-and-probable-to-complete threshold and become effective for annual reporting periods beginning after December 15, 2027, with early adoption permitted.

Your financial model should therefore separate:

Economic investment
Accounting expense
Capitalised expenditure
Cash payment

These may occur at different times.

3.6 Depreciation and amortisation

Depreciation allocates the cost of tangible assets over their useful life.

Amortisation generally allocates the cost of intangible assets.

Example:

A company buys $1.2 million of GPU infrastructure with an estimated three-year useful life and no residual value.

Straight-line depreciation is:

\text{Annual Depreciation} = \frac{\$1.2\text{m}}{3} = \$400{,}000

The $1.2 million purchase appears as a cash outflow when paid, but only $400,000 is charged as annual depreciation under this simplified example.

3.7 Working capital

Working capital reflects the timing difference between accounting revenue, expenses, and cash.

Important working-capital items include:

Accounts receivable
Accounts payable
Deferred revenue
Accrued expenses
Prepayments

Suppose an enterprise AI customer is invoiced $120,000 annually but pays after 60 days.

Revenue may be recognised before cash is received. The unpaid balance becomes accounts receivable.

A simple accounts-receivable formula is:

\text{Accounts Receivable} = \text{Revenue} \times \frac{\text{Days Sales Outstanding}}{365}

If annual revenue is $12 million and days sales outstanding is 60:

\$12\text{m} \times \frac{60}{365} \approx \$1.97\text{m}

Approximately $1.97 million would be tied up in receivables under this simplified assumption.

3.8 Free cash flow

A simplified unlevered free-cash-flow formula is:

\text{Free Cash Flow} = \text{EBIT}\times(1-\text{Tax Rate}) + \text{Depreciation and Amortisation} - \text{Capital Expenditure} - \Delta\text{Working Capital}

For an internal AI business case, a simpler project cash-flow formula may be used:

\text{Net Project Cash Flow} = \text{Realised Benefits} - \text{Implementation Costs} - \text{Operating Costs} - \text{Additional Investment}

3.9 Net present value

A dollar received five years from now is generally worth less than a dollar received today.

Net present value, or NPV, discounts future cash flows into their present value.

\mathrm{NPV} = -I_0 + \sum_{t=1}^{n} \frac{CF_t}{(1+r)^t}

Where:

(I_0) is the initial investment
(CF_t) is the cash flow in period (t)
(r) is the discount rate
(n) is the number of periods

A positive NPV suggests that the project creates value relative to the selected discount rate.

In Excel:

=NPV(Discount_Rate, Year1:Year5) + Initial_Investment

Initial investment should normally be entered as a negative number.

For cash flows occurring on irregular dates, use:

=XNPV(Discount_Rate, Cash_Flows, Dates)

3.10 Internal rate of return

The internal rate of return, or IRR, is the discount rate at which NPV equals zero.

In Excel:

=IRR(Cash_Flow_Range)

For irregular dates:

=XIRR(Cash_Flows, Dates)

IRR is useful but should not be evaluated alone. It can be misleading when:

Projects have different sizes
Cash flows change sign several times
Projects have different durations
The initial investment is very small
Terminal values dominate the result

NPV should generally remain the central valuation measure.

3.11 Payback period

The payback period shows how long it takes to recover the original investment.

Suppose:

Initial investment: $10 million
Year-one net benefit: $4 million
Year-two net benefit: $8 million

At the end of year one, $6 million remains unrecovered.

Assuming year-two benefits are earned evenly:

\text{Additional Payback Time} = \frac{\$6\text{m}}{\$8\text{m}} = 0.75\text{ years}

The payback period is approximately:

1 + 0.75 = 1.75\text{ years}

Payback is easy to understand but ignores benefits received after payback and may ignore the time value of money.

Part II: Building the Financial Model

4. Recommended workbook structure

A professional AI financial model could use the following tabs.

Tab	Purpose
00_Cover	Model title, version, owner and date
01_Instructions	How to use the model
02_Assumptions	Central input register
03_Scenarios	Downside, base and upside assumptions
04_Historical	Actual financial and operational data
05_Demand	Users, customers, transactions and requests
06_AI_Usage	Tokens, model calls, tools, agents and retrieval
07_Revenue	Subscription, usage and implementation revenue
08_AI_COGS	Model, infrastructure, data and review costs
09_Headcount	Employees, salaries, benefits and hiring
10_OpEx	Sales, marketing, legal, governance and overhead
11_CapEx	Hardware and capitalised development
12_P&L	Income statement
13_Balance_Sheet	Assets, liabilities and equity
14_Cash_Flow	Operating, investing and financing cash flow
15_Valuation	DCF, terminal value and returns
16_Unit_Economics	Cost and value per successful outcome
17_Sensitivity	Two-variable sensitivity tables
18_Risk	Risk-adjusted scenarios and expected losses
19_Dashboard	Executive summary
20_Checks	Model integrity and error checks

Not every model requires every tab. A small internal pilot may use eight to ten tabs, while an investor-grade model may need the full structure.

5. Model design principles

5.1 Separate inputs from formulas

Never repeatedly hard-code the same assumption.

Poor formula:

=10000000*65%*50%*55%

Better formula:

=Annual_Contacts*AI_Eligible_Percentage*Adoption_Rate*Containment_Rate

This makes the model:

Easier to audit
Easier to update
Easier to explain
Less likely to contain inconsistent assumptions

5.2 Use one source of truth

A model price should exist in one assumptions location.

Do not type $0.32 in 20 different formulas. Link all formulas to one named assumption.

5.3 Clearly identify units

Every row should specify its unit.

Examples:

$
$000
$ millions
Customers
Users
Requests
Million tokens
GPU hours
Percentage
Days
Full-time equivalents

A large number of modelling errors result from mixing units.

For example:

2{,}000{,}000\text{ tokens} \neq 2{,}000{,}000\text{ million-token units}

If the price is quoted per million tokens:

\text{Token Cost} = \frac{\text{Tokens}}{1{,}000{,}000} \times \text{Price per Million Tokens}

5.4 Use consistent signs

One common convention is:

Revenue and benefits: positive
Expenses and investments: negative
Cash inflows: positive
Cash outflows: negative

Alternatively, financial statements may display expenses as positive values and subtract them through formulas.

Either approach is acceptable if it is applied consistently.

5.5 Build from operational drivers

Do not forecast revenue simply as:

\text{Next-Year Revenue} = \text{Current Revenue}\times 1.30

Build the operational story:

\text{Revenue} = \text{Customers}\times\text{Average Revenue per Customer}

And:

\text{Customers}_{t} = \text{Customers}_{t-1} + \text{New Customers}_{t} - \text{Churned Customers}_{t}

This makes growth explainable.

5.6 Build model checks

Examples include:

Total Assets - Total Liabilities - Equity = 0

Opening Cash + Net Cash Movement - Closing Cash = 0

Customer Opening + New - Churn - Closing = 0

Allocated Cloud Cost - Total Cloud Cost = 0

Scenario Weights Sum = 100%

AI Requests - Successful - Failed - Escalated = 0

A model should have a visible overall status:

MODEL CHECK: OK

or:

MODEL CHECK: ERROR

Part III: Modelling AI Demand and Technical Usage

6. Start with business demand

Begin with the number of business events, not tokens.

Examples:

Customer enquiries
Documents received
Claims submitted
Sales opportunities
Software-development tasks
Payments screened
Products recommended
Manufacturing assets monitored

Suppose a bank expects 10 million customer enquiries.

Not every enquiry is suitable for AI.

\text{Eligible Enquiries} = \text{Total Enquiries} \times \text{AI Eligibility Rate}

If 65% are eligible:

10\text{m}\times 65\% = 6.5\text{m}

Not every eligible user will use the AI channel.

\text{AI-Touched Enquiries} = \text{Eligible Enquiries} \times \text{AI Adoption Rate}

If adoption is 50%:

6.5\text{m}\times 50\% = 3.25\text{m}

This 3.25 million becomes the operational workload for the AI architecture.

7. Translate demand into model usage

A single business task may create several technical events.

One support conversation may involve:

Input moderation
Query classification
Embedding generation
Vector retrieval
Reranking
Main LLM response
Tool call
Second LLM response
Output moderation
Evaluation logging
Human review

Therefore:

\text{Model Calls} = \text{Business Requests} \times \text{Average Calls per Request} \times \text{Retry Factor}

If:

Business requests = 1 million
Average model calls = 3.2
Retry factor = 1.08

Then:

1\text{m}\times 3.2\times 1.08 = 3.456\text{m model calls}

7.1 Retry factor

The retry factor captures:

Technical failures
Timeout retries
Invalid structured output
Guardrail rejections
Tool failures
Model escalation
Agent loops
User regeneration requests

A retry factor of 1.08 means an average 8% additional workload.

8. Input-token modelling

Input tokens may include:

System prompt
User message
Conversation history
Retrieved documents
Tool outputs
Policies
Examples
Structured schemas
Agent state

\text{Input Tokens} = \text{Model Calls} \times \text{Average Input Tokens per Call}

If:

Calls = 3.456 million
Average input tokens = 2,500

Then:

3.456\text{m}\times 2{,}500 = 8.64\text{ billion input tokens}

The input-token cost is:

\text{Input Cost} = \frac{\text{Input Tokens}}{1{,}000{,}000} \times \text{Price per Million Input Tokens}

9. Output-token modelling

\text{Output Tokens} = \text{Model Calls} \times \text{Average Output Tokens per Call}

If output averages 500 tokens:

3.456\text{m}\times 500 = 1.728\text{ billion output tokens}

\text{Output Cost} = \frac{\text{Output Tokens}}{1{,}000{,}000} \times \text{Price per Million Output Tokens}

Input and output prices should be modelled separately because they often differ.

Current cloud AI services also offer different commercial arrangements. Azure provides pay-as-you-go and provisioned-throughput approaches; Amazon Bedrock offers service tiers including reserved, priority, standard, and flex; and Vertex AI publishes model-specific token and compute pricing. Consequently, model pricing should be held in a dated vendor-pricing assumptions table rather than embedded permanently in formulas.

10. Model cascading and routing

A production solution may use several models.

For example:

70% of requests use a low-cost model
20% use a mid-tier model
10% use a high-capability model

Expected model cost is:

\text{Expected Cost} = \sum_{i=1}^{n} \left( \text{Routing Share}_i \times \text{Cost per Request}_i \right)

Example:

Model tier	Routing share	Cost per request
Small	70%	$0.01
Medium	20%	$0.04
Large	10%	$0.15

(70\%\times\$0.01) + (20\%\times\$0.04) + (10\%\times\$0.15) = \$0.030

The expected blended model cost is $0.03 per request.

However, model routing must also satisfy quality requirements.

The financial optimisation problem is therefore:

\min \text{Total AI Cost}

Subject to:

\text{Accuracy}\geq \text{Required Accuracy}

\text{Latency}\leq \text{Maximum Latency}

\text{Safety Risk}\leq \text{Risk Tolerance}

\text{Capacity}\geq \text{Peak Demand}

The cheapest model is not necessarily the most economical if it causes:

More retries
More escalations
Lower conversion
More human review
Customer dissatisfaction
Incorrect decisions
Regulatory incidents

11. Agentic AI cost modelling

Agentic systems may generate a variable number of steps.

Suppose an agent can:

Search a knowledge base
Query customer records
Calculate an account balance
Generate an answer
Request approval
Update a system

The expected number of calls can be modelled as:

E(\text{Calls per Task}) = \sum_{j=1}^{m} P(\text{Path}_j) \times \text{Calls in Path}_j

Example:

Agent path	Probability	Calls
Simple answer	55%	2
Retrieval answer	25%	4
Tool-assisted answer	15%	7
Human escalation	5%	5

E(\text{Calls}) = 55\%\times 2 + 25\%\times 4 + 15\%\times 7 + 5\%\times 5 = 1.10 + 1.00 + 1.05 + 0.25 = 3.40

The average task creates 3.4 model calls.

You should also model a maximum-step control. Without it, a malfunctioning agent loop could create unpredictable cost.

\text{Maximum Daily Exposure} = \text{Maximum Requests} \times \text{Maximum Agent Steps} \times \text{Maximum Cost per Step}

This becomes a useful financial-risk limit.

12. Retrieval-augmented generation costs

A RAG solution may include:

Initial indexing costs

\text{Initial Embedding Cost} = \frac{\text{Documents}\times\text{Chunks per Document}\times\text{Tokens per Chunk}}{1{,}000{,}000} \times \text{Embedding Price}

Ongoing indexing costs

These include:

New documents
Updated documents
Deleted-document processing
Re-embedding
Metadata enrichment
OCR
Document parsing
Image processing

Query costs

\text{RAG Query Cost} = \text{Embedding Cost} + \text{Vector Search Cost} + \text{Reranking Cost} + \text{Additional Context Token Cost}

RAG can reduce unsupported answers, but retrieved context may substantially increase input-token consumption.

Therefore, the model should separately measure:

Cost without retrieval
Cost with retrieval
Quality improvement
Reduction in escalation
Reduction in errors
Reduction in human review

The economically correct decision depends on the net effect.

13. Caching economics

Caching can avoid repeated model generation.

\text{Cache Hit Rate} = \frac{\text{Cache Hits}}{\text{Eligible Requests}}

A simplified post-cache cost is:

\text{Cost After Caching} = \text{Cache Misses}\times\text{Inference Cost} + \text{Cache Operating Cost} + \text{Cache Read/Write Cost}

Example:

One million eligible requests
30% cache-hit rate
Inference cost of $0.08 per request
Cache platform cost of $5,000

Without caching:

1\text{m}\times\$0.08 = \$80{,}000

With caching:

700{,}000\times\$0.08 + \$5{,}000 = \$61{,}000

Savings:

\$80{,}000 - \$61{,}000 = \$19{,}000

However, the model should include risks such as:

Stale answers
User-specific information leakage
Incorrect cache keys
Reduced personalisation
Compliance restrictions

Part IV: AI Revenue Modelling

14. Subscription pricing

\text{Subscription Revenue} = \text{Average Customers} \times \text{Annual Subscription Price}

Average customers can be approximated as:

\text{Average Customers} = \frac{\text{Opening Customers}+\text{Closing Customers}}{2}

A more accurate monthly model should calculate customers and revenue for every month.

15. Usage-based pricing

Common charging units include:

API calls
Tokens
Documents
Minutes of audio
Images generated
Agent sessions
Completed workflows
Automated resolutions
Seats plus usage
Compute time

\text{Usage Revenue} = \text{Billable Units} \times \text{Price per Unit}

Be careful not to equate technical usage directly with billable usage.

For example:

One customer conversation may contain 12 messages
Each message may create three model calls
The contract may charge for only one resolved conversation

The model needs separate rows for:

Business units
Billable units
Technical units

16. Outcome-based pricing

An AI vendor may charge per successful outcome.

Examples:

Per resolved customer case
Per recovered invoice
Per approved claim
Per qualified lead
As a percentage of verified savings

\text{Outcome Revenue} = \text{Successful Outcomes} \times \text{Price per Outcome}

Success should be precisely defined.

For example, a “resolved case” might require:

No human escalation
No customer reopening within seven days
Correct policy application
Satisfactory customer feedback
No compliance breach

Loose definitions create billing disputes and revenue-model uncertainty.

17. Customer forecasting

A basic customer schedule is:

\text{Closing Customers}_t = \text{Opening Customers}_t + \text{New Customers}_t - \text{Churned Customers}_t

\text{Churned Customers} = \text{Opening Customers} \times \text{Customer Churn Rate}

For enterprise sales, new customers should be driven by a sales funnel:

\text{Qualified Opportunities} = \text{Marketing Leads} \times \text{Qualification Rate}

\text{Proposals} = \text{Qualified Opportunities} \times \text{Proposal Rate}

\text{New Customers} = \text{Proposals} \times \text{Win Rate}

A capacity constraint should also be included:

\text{New Customers} = \min \left( \text{Expected Wins}, \text{Implementation Capacity} \right)

Otherwise, the model may forecast more customers than the implementation team can onboard.

18. Annual recurring revenue

\mathrm{ARR} = \text{Annualised Recurring Subscription Revenue}

Do not automatically include:

One-time implementation revenue
Consulting revenue
Hardware resale
Non-recurring customisation

Monthly recurring revenue is:

\mathrm{MRR}=\frac{\mathrm{ARR}}{12}

19. Net revenue retention

Net revenue retention captures expansion, contraction, and churn from existing customers.

\mathrm{NRR} = \frac{ \text{Opening Recurring Revenue} + \text{Expansion} - \text{Contraction} - \text{Churn} }{ \text{Opening Recurring Revenue} }

Example:

Opening recurring revenue: $10 million
Expansion: $2 million
Contraction: $0.5 million
Churn: $0.8 million

\mathrm{NRR} = \frac{10+2-0.5-0.8}{10} = 107\%

Usage-based AI companies should model NRR carefully because customer consumption may fluctuate even without formal churn.

20. Customer acquisition cost

\mathrm{CAC} = \frac{ \text{Sales and Marketing Cost Attributable to Acquisition} }{ \text{New Customers} }

If sales and marketing cost is $3 million and 30 customers are acquired:

\mathrm{CAC} = \frac{\$3\text{m}}{30} = \$100{,}000

For enterprise AI, CAC may include:

Sales salaries
Commissions
Marketing
Solution engineering
Proof-of-concept costs
Security assessment support
Proposal effort
Partner fees
Travel
Pre-sales cloud usage

21. Customer lifetime value

A simplified SaaS LTV formula is:

\mathrm{LTV} = \frac{ \text{Average Annual Revenue per Customer} \times \text{Gross Margin} }{ \text{Annual Revenue Churn Rate} }

Suppose:

Annual revenue per customer = $100,000
Gross margin = 70%
Revenue churn = 10%

\mathrm{LTV} = \frac{\$100{,}000\times 70\%}{10\%} = \$700{,}000

This formula is a simplification. A cohort-based discounted cash-flow approach is better when:

Revenue changes over time
Consumption is volatile
Margins improve with scale
Customer behaviour differs by cohort
Contract terms vary
Expansion is significant

Part V: Modelling Enterprise AI Benefits

22. Cost-reduction benefits

Suppose an AI solution automates part of a manual process.

Theoretical savings are:

\text{Theoretical Savings} = \text{Automated Volume} \times \text{Current Cost per Unit}

But theoretical savings are not necessarily real cash savings.

You need an avoidability factor:

\text{Realised Savings} = \text{Theoretical Savings} \times \text{Avoidable Cost Percentage} \times \text{Benefit Realisation Percentage}

For example:

Automated cases: 1 million
Current cost: $5 per case
Avoidable cost: 70%
Realisation rate: 80%

1\text{m}\times\$5\times 70\%\times 80\% = \$2.8\text{m}

Not $5 million.

The difference may represent:

Fixed management costs
Buildings
Minimum staffing
Existing contracts
Redundancy delays
Retained specialist employees
Temporary dual-running
Demand growth absorbing freed capacity

23. Productivity benefits

Suppose AI saves each employee four hours per week.

\text{Hours Saved} = \text{Employees} \times \text{Hours Saved per Week} \times \text{Working Weeks}

For 500 employees and 46 working weeks:

500\times 4\times 46 = 92{,}000\text{ hours}

If loaded employee cost is $60 per hour:

92{,}000\times\$60 = \$5.52\text{m}

But this does not automatically mean the organisation receives $5.52 million in cash.

The benefit must be classified as one of the following:

Hard savings

Reduced headcount
Avoided recruitment
Reduced contractor expenditure
Reduced overtime
Reduced outsourcing cost

Capacity release

More work completed without hiring
Faster response times
Lower backlog
More customers supported

Revenue enablement

Employees spend additional time on sales
Faster product delivery
Increased customer capacity

Employee-experience value

Less repetitive work
Lower burnout
Improved retention

These categories should not be added together unless they represent genuinely separate economic outcomes.

24. Revenue-uplift benefits

Revenue impact may come from:

Better lead conversion
Improved cross-selling
Personalised recommendations
Faster response
Reduced customer churn
Increased product availability
Faster product launches
Better pricing
Higher sales productivity

A conversion-uplift model is:

\text{Incremental Revenue} = \text{Eligible Opportunities} \times \text{Adoption} \times \text{Conversion Uplift} \times \text{Average Order Value}

Incremental gross profit is:

\text{Incremental Gross Profit} = \text{Incremental Revenue} \times \text{Gross Margin}

Value should normally be based on incremental profit or cash flow, not revenue alone.

25. Risk-reduction benefits

AI may reduce:

Fraud losses
Compliance penalties
Operational errors
Cybersecurity losses
Credit losses
Customer remediation
Legal claims
Downtime

A simplified expected-loss formula is:

\text{Expected Loss} = \text{Probability of Event} \times \text{Financial Impact}

If:

Annual probability = 8%
Impact = $20 million

Then:

8\%\times\$20\text{m} = \$1.6\text{m}

If an AI control reduces the probability to 5%:

5\%\times\$20\text{m} = \$1\text{m}

Expected risk reduction:

\$1.6\text{m}-\$1\text{m} = \$600{,}000

This is expected value, not guaranteed cash.

Risk modelling should also consider severity ranges, correlated risks, model failures, privacy incidents, harmful outputs, intellectual-property issues, and regulatory consequences. NIST’s AI Risk Management Framework organises AI risk work around GOVERN, MAP, MEASURE, and MANAGE, while its Generative AI Profile adds guidance covering governance, content provenance, pre-deployment testing, and incident disclosure.

Part VI: Complete Worked Example

26. Scenario: AI customer-service platform for a bank

A bank receives 10 million customer contacts annually. It is considering an AI customer-service platform.

The platform will:

Answer common questions
Retrieve policy information
Access approved customer data
Complete selected tasks
Assist human agents when it cannot resolve a case
Escalate sensitive cases
Record interactions for audit

The financial model covers five years.

27. Base-case assumptions

Demand assumptions

Assumption	Year 1	Year 2	Year 3	Year 4	Year 5
Customer contacts	10.00m	10.30m	10.61m	10.93m	11.26m
AI-eligible share	65%	65%	65%	65%	65%
AI adoption	50%	65%	75%	82%	87%
Successful containment	55%	62%	68%	72%	75%

Financial assumptions

Assumption	Year 1	Year 2	Year 3	Year 4	Year 5
Human cost per contact	$7.00	$7.21	$7.43	$7.65	$7.88
Avoidable share	70%	72%	75%	78%	80%
Productivity saving on escalated cases	18%	22%	25%	28%	30%
AI variable cost per AI-touched case	$0.32	$0.29	$0.26	$0.24	$0.22

Additional assumptions:

Initial implementation investment: $10 million
Discount rate: 10%
Platform, governance, change, and enhancement costs are modelled separately
Calculations are before tax
The figures are illustrative, not a bank forecast

28. Year-one calculation

Step 1: AI-eligible contacts

10\text{m}\times 65\% = 6.5\text{m}

Step 2: AI-touched contacts

6.5\text{m}\times 50\% = 3.25\text{m}

Step 3: Successfully contained contacts

3.25\text{m}\times 55\% = 1.7875\text{m}

Step 4: Escalated but AI-assisted contacts

3.25\text{m}-1.7875\text{m} = 1.4625\text{m}

Step 5: Containment savings

1.7875\text{m}\times\$7.00\times 70\% = \$8.76\text{m}

Step 6: Human-agent productivity savings

1.4625\text{m}\times\$7.00\times 70\%\times 18\% = \$1.29\text{m}

Step 7: Total gross benefit

\$8.76\text{m}+\$1.29\text{m} = \$10.05\text{m}

Step 8: AI variable cost

3.25\text{m}\times\$0.32 = \$1.04\text{m}

Step 9: Other annual costs

Assume:

Platform fixed cost: $1.80 million
Governance and risk: $0.75 million
Change management: $1.00 million
Enhancements: $0.80 million

\text{Other Costs} = \$4.35\text{m}

Step 10: Total annual cost

\$1.04\text{m}+\$4.35\text{m} = \$5.39\text{m}

Step 11: Net year-one cash benefit

\$10.05\text{m}-\$5.39\text{m} = \$4.66\text{m}

29. Five-year result

Year	Contacts	AI touched	Contained	Gross benefits	Operating and enhancement cost	Net cash benefit	Present value at 10%
1	10.00m	3.25m	1.79m	$10.05m	$5.39m	$4.66m	$4.24m
2	10.30m	4.35m	2.70m	$15.89m	$4.91m	$10.98m	$9.08m
3	10.61m	5.17m	3.52m	$21.89m	$4.79m	$17.10m	$12.85m
4	10.93m	5.82m	4.19m	$27.74m	$4.84m	$22.90m	$15.64m
5	11.26m	6.36m	4.77m	$33.10m	$4.97m	$28.13m	$17.46m

Initial investment:

-\$10\text{m}

Present value of five-year net benefits:

\$59.27\text{m}

Net present value:

\$59.27\text{m}-\$10\text{m} = \$49.27\text{m}

Indicative results:

NPV: approximately $49.27 million
IRR: approximately 96%
Payback: during year two
Five-year undiscounted net value after initial investment: approximately $73.77 million

These results look attractive, but the model is highly dependent on:

Adoption
Containment
Avoidable labour cost
Realisation timing
Quality
AI operating cost
Regulatory approval
Customer-channel behaviour

The base case should never be presented without sensitivities.

Part VII: Scenario and Sensitivity Analysis

30. Downside, base, and upside scenarios

A scenario model should change several logically related assumptions together.

Downside scenario

Slower adoption
Lower containment
More human review
Higher implementation cost
Longer regulatory approval
Higher model cost
Lower avoidable labour share
Delayed benefit realisation

Base scenario

Management’s most supportable assumptions
Evidence from pilots
Normal implementation progress
Expected adoption
Expected quality improvement

Upside scenario

Faster adoption
Better containment
Lower inference cost
Higher staff productivity
Broader use-case expansion
More effective model routing

Example:

Assumption	Downside	Base	Upside
Year-three adoption	55%	75%	88%
Year-three containment	50%	68%	78%
Avoidable human cost	55%	75%	85%
Initial implementation	$14m	$10m	$8m
AI cost per case	$0.45	$0.26	$0.18
Deployment delay	12 months	3 months	None

31. Weighted scenario valuation

Suppose:

Downside NPV = negative $5 million
Base NPV = $49 million
Upside NPV = $90 million

Probabilities:

Downside: 25%
Base: 55%
Upside: 20%

Expected NPV:

25\%\times(-\$5\text{m}) + 55\%\times\$49\text{m} + 20\%\times\$90\text{m} = -\$1.25\text{m}+\$26.95\text{m}+\$18\text{m} = \$43.70\text{m}

Probability weights should be supportable and should not be manipulated merely to produce approval.

32. Sensitivity tables

A two-variable sensitivity table can show NPV across:

Adoption versus containment
AI unit cost versus volume
Subscription price versus customer churn
Win rate versus sales-cycle duration
Human-review rate versus model accuracy
Gross margin versus customer growth
Discount rate versus terminal growth

Example structure:

Adoption / Containment	50%	60%	70%	80%
40% adoption	NPV	NPV	NPV	NPV
55% adoption	NPV	NPV	NPV	NPV
70% adoption	NPV	NPV	NPV	NPV
85% adoption	NPV	NPV	NPV	NPV

This identifies the economic break-even boundary.

33. Break-even analysis

Break-even adoption

Set NPV to zero and solve for adoption.

Conceptually:

0 = -\text{Initial Investment} + \mathrm{PV}(\text{Benefits driven by Adoption}) - \mathrm{PV}(\text{Costs})

Excel Goal Seek can be used:

Set cell: NPV
To value: 0
By changing cell: Adoption_Rate

Break-even containment

Similarly, determine the minimum containment rate required to justify the investment.

This is useful for defining pilot success criteria.

For example:

The solution should not progress to enterprise rollout unless independently measured containment exceeds 58%, customer satisfaction remains above the agreed threshold, and serious compliance incidents remain at zero.

That connects the financial model to the technical evaluation plan.

Part VIII: Advanced AI Financial Modelling

34. Monte Carlo simulation

Traditional scenarios use a small number of fixed cases.

Monte Carlo simulation samples uncertain variables thousands of times.

Possible variables include:

Customer adoption
Churn
Model accuracy
Containment
Inference prices
Volume growth
Implementation delays
Salary inflation
Revenue conversion
Incident losses

Example distributions:

Variable	Possible distribution
Adoption	Beta or triangular
Implementation cost	Lognormal
Deployment delay	Discrete
Token consumption	Lognormal
Churn	Beta
Incident impact	Heavy-tailed distribution
Model price decline	Triangular

The result may show:

Mean NPV
Median NPV
Probability NPV is negative
Fifth-percentile NPV
Ninety-fifth-percentile NPV
Cash required under stress
Probability of breaching budget

A project with an expected NPV of $20 million may still be unattractive if it has a 40% probability of losing more than $15 million.

35. Fine-tuning break-even model

Fine-tuning may involve:

Data preparation
Annotation
Training
Engineering
Evaluation
Security testing
Deployment
Monitoring
Re-training

Suppose:

Fine-tuning investment = $300,000
Existing model cost per task = $0.18
Fine-tuned model cost per task = $0.10
Additional operating cost remains unchanged

Savings per task:

\$0.18-\$0.10 = \$0.08

Break-even volume:

\frac{\$300{,}000}{\$0.08} = 3.75\text{m tasks}

However, the decision must also incorporate:

Quality improvement
Model maintenance
Re-training frequency
Vendor dependence
Evaluation cost
Deployment complexity
Reduced prompt size
Lower human-review rate

A fine-tuned model that reduces human review can produce far greater value than inference savings alone.

36. Open-weight hosting versus managed API

Managed API cost

\text{Managed Cost} = \text{Tokens}\times\text{API Rate} + \text{Other Services}

Advantages may include:

Low initial investment
Rapid deployment
Elastic capacity
Reduced platform management

Self-hosted model cost

\text{Self-Hosted Cost} = \text{GPU Infrastructure} + \text{Platform Engineering} + \text{MLOps} + \text{Support} + \text{Energy} + \text{Networking} + \text{Security} + \text{Redundancy}

Effective cost per utilised GPU hour is:

\text{Effective GPU-Hour Cost} = \frac{ \text{Monthly Fixed Compute Cost} + \text{Monthly Platform Cost} }{ \text{Available GPU Hours} \times \text{Utilisation} }

Suppose:

Monthly fixed cost = $100,000
Available capacity = 5,000 GPU hours
Utilisation = 40%

\frac{\$100{,}000}{5{,}000\times 40\%} = \$50\text{ per utilised GPU hour}

At 80% utilisation:

\frac{\$100{,}000}{5{,}000\times 80\%} = \$25\text{ per utilised GPU hour}

Utilisation can therefore materially change self-hosting economics.

The model should account for:

Peak-demand headroom
Replication
Failover
Development environments
Idle capacity
Maintenance periods
Model loading
Batch versus real-time demand
Data residency
Security personnel
Hardware obsolescence

37. Reserved capacity versus pay-as-you-go

Reserved capacity may be attractive when usage is stable.

Let:

Reserved monthly cost = (R)
Pay-as-you-go rate per unit = (P)
Monthly demand = (Q)

Pay-as-you-go cost:

P\times Q

Reserved capacity is financially favourable when:

R < P\times Q

Therefore, break-even volume is:

Q^{*} = \frac{R}{P}

But the model must also include:

Underutilisation
Commitment term
Demand uncertainty
Capacity limits
Overage pricing
Latency requirements
Service availability
Cancellation terms

Amazon Bedrock’s service tiers illustrate how providers can differentiate capacity by reservation, latency priority, standard access, and flexible non-time-critical processing. AWS also exposes model, token-type, service-tier, and routing information in cost-and-usage reporting, which can be used for model allocation.

38. AI quality-adjusted unit economics

A raw request can be cheap but unsuccessful.

Define:

\text{Task Success Rate} = \frac{\text{Accepted Successful Tasks}}{\text{Attempted Tasks}}

Then:

\text{Cost per Successful Task} = \frac{\text{Total Task Cost}}{\text{Successful Tasks}}

Example:

Model A

Cost per attempt: $0.03
Success rate: 60%

\text{Cost per Success} = \frac{\$0.03}{60\%} = \$0.05

Model B

Cost per attempt: $0.04
Success rate: 90%

\text{Cost per Success} = \frac{\$0.04}{90\%} = \$0.044

Model B is more expensive per attempt but cheaper per successful outcome.

A complete formula is:

\text{Cost per Successful Outcome} = \frac{ \text{Inference} + \text{Retrieval} + \text{Tools} + \text{Evaluation} + \text{Retries} + \text{Human Review} }{ \text{Accepted Outcomes} }

39. Latency-adjusted economics

Reducing cost by using a slower service tier may reduce conversion or customer satisfaction.

Expected contribution per request can be modelled as:

\text{Expected Contribution} = P(\text{Completion}\mid\text{Latency}) \times \text{Value per Completion} - \text{AI Cost}

Example:

Fast architecture

Completion probability: 92%
Value per completion: $2.00
AI cost: $0.12

92\%\times\$2.00-\$0.12 = \$1.72

Slow architecture

Completion probability: 80%
Value per completion: $2.00
AI cost: $0.07

80\%\times\$2.00-\$0.07 = \$1.53

The slower architecture saves $0.05 in AI cost but loses $0.19 of expected contribution.

40. Human-in-the-loop economics

Human review is often one of the largest AI operating costs.

\text{Human Review Cost} = \text{AI Tasks} \times \text{Review Rate} \times \text{Minutes per Review} \times \text{Cost per Minute}

Suppose:

2 million AI tasks
10% reviewed
Five minutes per review
Reviewer cost of $0.75 per minute

2\text{m}\times 10\%\times 5\times\$0.75 = \$750{,}000

Reducing the review rate from 10% to 6% saves:

2\text{m}\times 4\%\times 5\times\$0.75 = \$300{,}000

But review reductions should be approved on a risk basis, not made solely to improve margin.

41. Expected cost of AI failures

\text{Expected Failure Cost} = \sum \left( \text{Failure Probability} \times \text{Failure Impact} \right)

Example:

Failure	Annual probability	Impact	Expected cost
Minor customer remediation	30%	$100,000	$30,000
Material data incident	3%	$5,000,000	$150,000
Major regulatory incident	0.5%	$30,000,000	$150,000

Total simplified expected cost:

\$330{,}000

Expected-value modelling should not replace controls for low-probability catastrophic risks. Some risks may be unacceptable regardless of their average expected cost.

42. Real-options modelling

An AI pilot can be treated as an option rather than merely a small project.

For example:

Invest $500,000 in a pilot
Learn whether containment exceeds 60%
Expand only if the evidence is positive
Abandon or redesign otherwise

The pilot purchases information and limits downside exposure.

A staged investment model could be:

Stage 1: Discovery

Data assessment
Process mapping
Initial architecture
Business-case validation

Stage 2: Prototype

Limited dataset
Offline evaluation
Security design

Stage 3: Pilot

Selected users
Controlled production
Human supervision

Stage 4: Scale

Wider customer rollout
Integration
Automation
Capacity commitment

At each gate, the organisation has the option to:

Continue
Pause
Redesign
Switch vendors
Reduce scope
Stop

The value of this flexibility is lost in a model that assumes the full investment is irreversibly committed on day one.

Part IX: AI Startup Financial Modelling

43. Revenue build

An AI SaaS company might have:

\text{Total Revenue} = \text{Platform Subscriptions} + \text{Usage Revenue} + \text{Implementation} + \text{Premium Support}

Build separate schedules for:

Small customers
Mid-market customers
Enterprise customers
Channel partners
Geographic regions

For each segment, model:

Opening customers
New customers
Churn
Expansion
Contract value
Usage
Gross margin
Sales cycle
Implementation time
Payment terms

44. AI startup COGS schedule

A detailed COGS model might include:

Cost component	Cost driver
Input tokens	Input tokens consumed
Output tokens	Output tokens generated
Embeddings	Documents and queries embedded
Vector database	Stored vectors and queries
Reranking	Reranked requests
Agent tools	Tool calls
Speech	Audio minutes
Search API	Search requests
Observability	Traces or events
Human review	Reviews and review time
Customer support	Tickets and handling time
Hosting	Customer workloads
Data	Licensed records or requests

The model should calculate gross margin by:

Customer
Product
Use case
Model
Geography
Contract type
Customer cohort

A high-growth customer may still destroy value if its usage cost exceeds its contracted price.

45. Headcount planning

For each employee group, model:

\text{Closing Headcount} = \text{Opening Headcount} + \text{Hires} - \text{Leavers}

\text{Salary Cost} = \text{Average Headcount} \times \text{Average Salary} \times (1+\text{Benefits Load})

Functions may include:

AI engineering
Platform engineering
Data engineering
Product
Design
Security
Governance
Sales
Marketing
Customer success
Finance
Legal
People operations

Hiring dates should be modelled monthly. Assuming every new employee works for the full year overstates cost.

46. Cash runway

\text{Net Burn} = \text{Cash Operating Outflows} - \text{Cash Operating Inflows}

\text{Runway} = \frac{\text{Current Cash}}{\text{Average Monthly Net Burn}}

Example:

Cash balance: $6 million
Monthly net burn: $500,000

\text{Runway} = \frac{\$6\text{m}}{\$0.5\text{m}} = 12\text{ months}

A stronger model forecasts cash month by month and identifies the minimum cash balance.

Fundraising should be modelled before cash reaches zero, allowing for:

Investor preparation
Due diligence
Negotiation
Legal completion
Market delays
Downside contingency

Part X: Valuing an AI Business

47. Discounted cash-flow valuation

Forecast:

Revenue
COGS
Operating expenses
EBIT
Tax
Depreciation and amortisation
Capital expenditure
Working capital
Free cash flow

Then calculate:

\text{Enterprise Value} = \mathrm{PV}(\text{Forecast Free Cash Flows}) + \mathrm{PV}(\text{Terminal Value})

A perpetual-growth terminal value is:

\mathrm{TV} = \frac{FCF_{n+1}}{\mathrm{WACC}-g}

Where:

WACC is the discount rate
g is perpetual growth

The condition must hold:

\mathrm{WACC} > g

Equity value is:

\text{Equity Value} = \text{Enterprise Value} + \text{Cash} - \text{Debt} - \text{Debt-like Items}

48. Comparable-company valuation

Possible multiples include:

Enterprise value/revenue
Enterprise value/ARR
Enterprise value/gross profit
Enterprise value/EBITDA

AI companies should not be compared only because they use AI.

Comparability should consider:

Revenue model
Growth
Gross margin
Customer concentration
Churn
Model dependence
Proprietary data
Intellectual property
Contract length
Cash burn
Regulatory exposure
Services intensity
Capital intensity
Market position

A company reselling model access with substantial consulting labour is economically different from a scalable AI platform.

49. Terminal-value discipline

In high-growth AI models, terminal value can dominate total valuation.

Always show:

\text{Terminal Value as Percentage of Enterprise Value}

A very high percentage means the valuation depends mainly on distant assumptions.

Apply sensitivity analysis to:

Discount rate
Terminal growth
Long-run margin
Long-run capital expenditure
Long-run model cost

Part XI: Executive AI Financial Dashboard

50. Financial metrics

The dashboard should include:

Revenue
ARR
Gross profit
Gross margin
EBITDA
Free cash flow
Cash balance
Monthly burn
Runway
NPV
IRR
Payback
Budget variance
Forecast variance

51. Unit-economics metrics

Revenue per successful task
Cost per successful task
Contribution per task
Cost per customer
Gross margin per customer
CAC
LTV
CAC payback
AI cost as percentage of revenue
Human-review cost per outcome
Cloud cost per business transaction

52. Operational metrics

Active users
Requests
Adoption
Eligible volume
Automation
Containment
Completion
Escalation
Human-review rate
Customer retention
Implementation time

53. Technical metrics

Input tokens
Output tokens
Calls per task
Retries
Cache-hit rate
Average latency
Peak throughput
GPU utilisation
Tool-call success
Retrieval success
Model-routing share

54. Risk and quality metrics

Task-success rate
Unsupported-answer rate
Human override
Policy violations
Privacy incidents
Security incidents
Bias-test failures
Customer complaints
Model drift
Evaluation pass rate
High-severity incidents

Metrics should be connected.

For example:

\text{Lower Accuracy} \rightarrow \text{Higher Escalation} \rightarrow \text{Higher Human Cost} \rightarrow \text{Lower Savings} \rightarrow \text{Lower NPV}

Part XII: Common Financial-Modelling Mistakes

55. Modelling tokens before business demand

Tokens are a technical consequence of demand, not the starting point.

Start with:

\text{Customers} \rightarrow \text{Tasks} \rightarrow \text{Calls} \rightarrow \text{Tokens}

56. Counting productivity as immediate cash savings

Saving employee time does not automatically reduce payroll.

Define how the time will be converted into:

Headcount reduction
Hiring avoidance
Overtime reduction
Increased output
Increased revenue
Better service

57. Ignoring failed AI attempts

Your cost model should include:

Retries
Invalid outputs
Abandoned conversations
Escalation
Reprocessing
Human correction

58. Ignoring adoption

A technically excellent solution can produce little financial value if employees or customers do not use it.

Include an adoption curve rather than assuming immediate 100% usage.

59. Ignoring benefit delay

Implementation costs usually begin before benefits.

Model separately:

Build period
Pilot period
Ramp-up period
Full benefit period

60. Assuming every cost declines continuously

Some AI costs are step-fixed.

For example, additional GPU or provisioned capacity may need to be purchased when demand crosses a threshold.

Use:

=CEILING(Required_Capacity/Capacity_Per_Block,1)*Cost_Per_Block

61. Ignoring shared costs

Shared components might include:

AI gateway
Security
Networking
Observability
Platform engineering
Governance
Evaluation infrastructure

Allocate them using an appropriate driver, such as:

Requests
Tokens
Active users
Revenue
GPU hours
Dedicated capacity
Successful outcomes

62. Assuming model cost is the entire AI cost

A low model API bill can coexist with a high total cost of ownership.

Include:

Integration
Data preparation
Evaluation
Monitoring
Support
Security
Governance
Change management
Human review
Incident management
Vendor management

63. Double-counting benefits

Examples of double counting include:

Counting time saved as both payroll savings and revenue uplift
Counting reduced calls and reduced handling time on the same calls
Counting revenue and gross profit as separate benefits
Counting fraud prevention and avoided remediation when they represent the same loss
Counting customer retention in both customer-volume and revenue-uplift assumptions

Maintain a benefits register with:

Benefit name
Owner
Calculation
Source
Timing
Confidence
Dependencies
Potential overlaps

64. Using unsupported precision

Forecasting a 71.37% containment rate five years from now may create false confidence.

Use:

Rounded assumptions
Evidence ranges
Sensitivity analysis
Confidence ratings
Scenario bands

Precision should reflect evidence quality.

Part XIII: Beginner-to-Advanced Learning Roadmap

65. Beginner stage

Learn:

Revenue
Costs
Profit
Cash flow
CapEx and OpEx
Gross margin
EBITDA
Working capital
NPV
IRR
Payback

Build:

A simple revenue and cost model
A five-year project cash-flow forecast
An NPV and payback calculation
A downside/base/upside scenario

66. Intermediate stage

Learn:

Three-statement modelling
Customer cohorts
SaaS metrics
Headcount schedules
Deferred revenue
Cash runway
DCF valuation
Sensitivity tables
AI usage and COGS

Build:

An AI SaaS model
A customer-acquisition funnel
An AI unit-economics model
A build-versus-buy comparison
A monthly cash forecast

67. Advanced stage

Learn:

Monte Carlo simulation
Real-options analysis
Capacity modelling
Risk-adjusted valuation
Model-routing optimisation
Quality-adjusted unit economics
Cloud commitment modelling
Cohort profitability
Driver attribution
Probabilistic cash runway

Build:

A multi-model routing model
A GPU-versus-API break-even model
A probabilistic AI business case
A risk-adjusted AI portfolio model
An enterprise AI investment dashboard

Part XIV: Final Modelling Framework

A complete AI financial model can be summarised in twelve stages.

Stage 1: Define the decision

What exactly should management decide?

Stage 2: Define the baseline

What happens without the AI investment?

Stage 3: Forecast business demand

How many customers, users, cases, documents, or transactions will exist?

Stage 4: Forecast adoption

What percentage will actually use the AI solution?

Stage 5: Model the technical architecture

How many calls, tokens, tools, retrievals, reviews, and GPU hours will be required?

Stage 6: Calculate total cost of ownership

Include implementation, operation, governance, support, people, and risk.

Stage 7: Calculate business outcomes

How many tasks will be successfully completed?

Stage 8: Monetise benefits

Translate outcomes into cost savings, revenue, capacity, or risk reduction.

Stage 9: Build project cash flow

Place costs and benefits in the periods when cash is expected to move.

Stage 10: Calculate investment returns

Calculate NPV, IRR, payback, and break-even points.

Stage 11: Stress-test the model

Test adoption, quality, volume, cost, delay, and risk.

Stage 12: Connect finance to delivery gates

Define the performance required before moving from prototype to pilot and from pilot to scale.

The final relationship is:

\boxed{ \text{AI Financial Value} = \text{Successful Outcomes} \times \text{Value per Outcome} - \text{Total Cost of Ownership} - \text{Risk-Adjusted Losses} }

And:

\boxed{ \text{Successful Outcomes} = \text{Business Demand} \times \text{Eligibility} \times \text{Adoption} \times \text{Technical Completion} \times \text{Quality Acceptance} }

These two equations provide the foundation for almost every serious AI financial model.

Discussion

Comments

Share feedback or questions about this page. No account required.

Loading comments…

Part I: Financial Modelling Fundamentals

2. What type of AI financial model are you building?​

2.1 Enterprise AI business-case model​

2.2 AI product or SaaS model​

2.3 AI unit-economics model​

2.4 Build-versus-buy model​

2.5 AI infrastructure and capacity model​

2.6 AI company valuation model​

3. Essential financial concepts​

3.1 Revenue​

3.2 Cost of goods sold​

3.3 Operating expenses​

3.4 EBITDA​

3.5 Capital expenditure and operating expenditure​

Capital expenditure​

Operating expenditure​

3.6 Depreciation and amortisation​

3.7 Working capital​

3.8 Free cash flow​

3.9 Net present value​

3.10 Internal rate of return​

3.11 Payback period​

Part II: Building the Financial Model

4. Recommended workbook structure​

5. Model design principles​

5.1 Separate inputs from formulas​

5.2 Use one source of truth​

5.3 Clearly identify units​

5.4 Use consistent signs​

5.5 Build from operational drivers​

5.6 Build model checks​

Part III: Modelling AI Demand and Technical Usage

6. Start with business demand​

7. Translate demand into model usage​

7.1 Retry factor​

8. Input-token modelling​

9. Output-token modelling​

10. Model cascading and routing​

11. Agentic AI cost modelling​

12. Retrieval-augmented generation costs​

Initial indexing costs​

Ongoing indexing costs​

Query costs​

13. Caching economics​

Part IV: AI Revenue Modelling

14. Subscription pricing​

15. Usage-based pricing​

16. Outcome-based pricing​

17. Customer forecasting​

18. Annual recurring revenue​

19. Net revenue retention​

20. Customer acquisition cost​

21. Customer lifetime value​

Part V: Modelling Enterprise AI Benefits

22. Cost-reduction benefits​

23. Productivity benefits​

Hard savings​

Capacity release​

Revenue enablement​

Employee-experience value​

24. Revenue-uplift benefits​

25. Risk-reduction benefits​

Part VI: Complete Worked Example

26. Scenario: AI customer-service platform for a bank​

27. Base-case assumptions​

Demand assumptions​

Financial assumptions​

28. Year-one calculation​

Step 1: AI-eligible contacts​

Step 2: AI-touched contacts​

Step 3: Successfully contained contacts​

Step 4: Escalated but AI-assisted contacts​

Step 5: Containment savings​

Step 6: Human-agent productivity savings​

Step 7: Total gross benefit​

Step 8: AI variable cost​

Step 9: Other annual costs​

Step 10: Total annual cost​

Step 11: Net year-one cash benefit​

29. Five-year result​

2. What type of AI financial model are you building?

2.1 Enterprise AI business-case model

2.2 AI product or SaaS model

2.3 AI unit-economics model

2.4 Build-versus-buy model

2.5 AI infrastructure and capacity model

2.6 AI company valuation model

3. Essential financial concepts

3.1 Revenue

3.2 Cost of goods sold

3.3 Operating expenses

3.4 EBITDA

3.5 Capital expenditure and operating expenditure

Capital expenditure

Operating expenditure

3.6 Depreciation and amortisation

3.7 Working capital

3.8 Free cash flow

3.9 Net present value

3.10 Internal rate of return

3.11 Payback period

4. Recommended workbook structure

5. Model design principles

5.1 Separate inputs from formulas

5.2 Use one source of truth

5.3 Clearly identify units

5.4 Use consistent signs

5.5 Build from operational drivers

5.6 Build model checks

6. Start with business demand

7. Translate demand into model usage

7.1 Retry factor

8. Input-token modelling

9. Output-token modelling

10. Model cascading and routing

11. Agentic AI cost modelling

12. Retrieval-augmented generation costs

Initial indexing costs

Ongoing indexing costs

Query costs

13. Caching economics

14. Subscription pricing

15. Usage-based pricing

16. Outcome-based pricing

17. Customer forecasting

18. Annual recurring revenue

19. Net revenue retention

20. Customer acquisition cost

21. Customer lifetime value

22. Cost-reduction benefits

23. Productivity benefits

Hard savings

Capacity release

Revenue enablement

Employee-experience value

24. Revenue-uplift benefits

25. Risk-reduction benefits

26. Scenario: AI customer-service platform for a bank

27. Base-case assumptions

Demand assumptions

Financial assumptions

28. Year-one calculation

Step 1: AI-eligible contacts

Step 2: AI-touched contacts

Step 3: Successfully contained contacts

Step 4: Escalated but AI-assisted contacts

Step 5: Containment savings

Step 6: Human-agent productivity savings

Step 7: Total gross benefit

Step 8: AI variable cost

Step 9: Other annual costs

Step 10: Total annual cost

Step 11: Net year-one cash benefit

29. Five-year result

30. Downside, base, and upside scenarios

Downside scenario

Base scenario

Upside scenario

31. Weighted scenario valuation

32. Sensitivity tables