Enterprise AI, week 9
ROI And Running Costs
- Week 9 of 12: AI integration for companies
- Promise: a cost and benefit case your finance team can trust
Outcomes
What You Can Do By The End
- List all cost drivers
- Estimate token and licence cost
- Measure time saved against a baseline
- Build a case with ranges
- Set up FinOps controls
Start here
ROI Is A Simple Formula
- ROI = net benefit divided by cost
- The hard part is honest inputs
- Write every assumption down
- Show ranges, not one number
The full picture
Cost Drivers: One Time And Running
- One time: design, data, integration, review, training
- Running: licences or usage, hosting, monitoring, support
- People: owners, reviewers, champions
- Hidden: rework and duplicated subscriptions
| Cost | Type | Often missed |
|---|
| Design and discovery | One time | Yes |
| Data preparation | One time | Yes |
| Licences or usage | Running | No |
| Hosting and monitoring | Running | Sometimes |
| Review and support | Running | Yes |
Licences and usage
Two Pricing Styles
- Seat licence: fixed fee per user per month
- Usage pricing: pay for what you consume
- Seats suit assistants used by many people
- Usage suits workflows and agents
The unit of usage
What Is A Token
- A token is a small piece of text
- Input tokens are what you send
- Output tokens are what the model writes
- Output often costs more than input
- Prices differ by model and change over time
A worked formula
Estimate Monthly Token Cost
- Count everything you send, not only the question
- Instructions, documents and history count as input
- Agents multiply calls per task
- Use a small model for simple steps
requests_per_month = 20000
input_tokens = 3000 # question + policy text
output_tokens = 400
# prices are placeholders: read your vendor list
in_price = 0.000003 # per input token
out_price = 0.000015 # per output token
cost = requests_per_month * (input_tokens*in_price + output_tokens*out_price)
print(round(cost, 2))
Be honest
Measure Time Saved Against A Baseline
- Same task, comparable people, with and without
- Count the whole task including checking and fixing
- Include cases where the tool failed
- Self reports are a start, not proof
METR, 2025
Case: Faster In The Mind, Slower On The Clock
- Randomised study with experienced developers
- They expected a speedup and believed they got one
- Measured tasks took longer with the tools
- Feelings about speed can differ from measured speed
Book the benefit
Saved Time Must Turn Into Something
- More cases handled
- Faster response to customers
- Less overtime or contractor spend
- Work not hired for
- Say which kind of benefit you claim
Name the metric
Benefits Beyond Time
- Quality: fewer errors
- Speed: shorter waits
- Capacity: handle peaks
- Risk: earlier detection
- Value only what you can justify
| Benefit | Metric | Before | After |
|---|
| Quality | Invoice error rate | measure | measure |
| Speed | Reply time | measure | measure |
| Capacity | Cases per person | measure | measure |
Low, expected, high
Build The Case With Three Scenarios
- One table: costs and benefits by year
- Show payback time in months
- Run a sensitivity check
- Adoption is often the biggest lever
Shared responsibility
FinOps For AI
- Tag every workload by team and use case
- Budget and alerts for each use case
- Monthly review with the product owner
- Quarterly licence review, remove unused seats
Efficiency signal
Track Unit Cost
- Total spend rises when a tool succeeds
- Unit cost: cost per ticket or per invoice
- Falling unit cost means rising efficiency
- Use the smallest model that passes evaluation
What to remember
Recap
- Count one time, running and people costs
- Estimate tokens with a formula and real prices
- Measure the whole task against a baseline
- Present ranges and payback time
- Tag, budget, alert and review
Week 10
Next Week: The Operating Model
- AI council and center of excellence
- Roles, policies and vendor management
Your turn
Questions And Exercise
- Build a three scenario case for one use case
- Name your most sensitive assumption
- Bring the table to week 10