Courses / AI Integration For Companies / Week 9

Week 9 of 12 · Intermediate

ROI And Running Costs

You learn what AI really costs to run, how token and licence pricing work, and how to measure time saved without fooling yourself. You build a simple, honest cost and benefit case for one use case and set up FinOps habits to control spend.

3 hours17 slides6 min read

Open slides Read study notes

From Pilot To A Number A Finance Team Trusts The Cost Drivers Licence And Token Costs Measuring Time Saved Honestly Benefits Beyond Time Building The Case With Ranges FinOps For AI

You Will Be Able To

List the cost drivers of an AI use case, one time and recurring
Estimate token cost and licence cost for a use case with a simple formula
Measure time saved against a baseline and adjust it for review time
Build a cost and benefit case with a range instead of one number
Set up basic FinOps controls: tagging, budgets, alerts and reviews

From Pilot To A Number A Finance Team Trusts

In week 7 you recorded the minutes per step of a process. In week 8 you planned how people adopt the new way. Now we turn both into money. Return on investment, or ROI, compares what you gain with what you spend. The formula is simple: net benefit divided by cost. The hard part is getting honest inputs for both sides.

A finance team will challenge your numbers, and that is helpful. Prepare by writing every assumption down, giving ranges, and showing how you measured. A modest benefit with solid evidence is worth more than a large claim that nobody can check.

The Cost Drivers

AI cost has one time parts and running parts. Many first budgets include only the licence and miss the rest. List all of them before you decide.

One time costs include discovery and design, data preparation from week 3, integration with existing systems, security and legal review from week 5, evaluation work, and training from week 8. Running costs include licences or usage fees, cloud hosting for the systems around the model, monitoring, support, model and prompt maintenance, and the time of the people who review outputs. People are often the largest running cost and the least visible one.

One time: design, data preparation, integration, review, training
Running: licences or usage fees, hosting, monitoring, support
People: owners, reviewers, champions, maintenance work
Hidden: rework after errors, extra tools, and duplicated subscriptions

One Time And Running Costs

Cost	Type	Example	Often missed
Discovery and design	One time	Process mapping, pilot design	Yes
Data preparation	One time	Cleaning and access rules	Yes
Licences or usage	Running	Seats, tokens	No
Hosting and monitoring	Running	Cloud, dashboards	Sometimes
Review and support	Running	People who check outputs	Yes

A budget that lists only the licence is usually too low.

Licence And Token Costs

There are two common pricing styles. A seat licence charges a fixed fee per user per month, which makes cost easy to predict for assistants used by many staff. Usage pricing charges for what you consume. For language models the unit is the token, a small piece of text, roughly a short word or part of a word. Providers usually charge separately for input tokens, which are what you send, and output tokens, which are what the model writes back. Output tokens often cost more than input tokens, and prices differ by model, so check the current price list of your provider and do not rely on old figures.

A simple estimate: tokens per request, times requests per month, times price per token. Include everything you send, not only the user question. Long documents, retrieved passages, system instructions and chat history all count as input. An agent that makes ten calls to finish one task multiplies the cost by ten. Smaller models often handle simple steps at a fraction of the price, so route easy steps to a small model and hard steps to a large one.

Monthly cost = requests x (input tokens x input price + output tokens x output price)
Check vendor price lists often, because prices change
Caching repeated input can reduce cost on many platforms
Set rate limits and maximum lengths to prevent runaway use

Token Cost Estimate

Count input tokens
Count output tokens
Multiply by prices
Multiply by requests
Add a safety margin

Include system instructions and retrieved text in the input count.

Measuring Time Saved Honestly

Time saved is the most common benefit claim and the easiest one to overstate. Start from the baseline you measured before the pilot. Compare the same task, done by comparable people, with and without the tool. Count the whole task, including the time to write the request, wait, read the output, check it, fix it and handle the cases where the tool failed.

Self reported savings are a start but not proof. In 2025 the research group METR ran a randomised study with experienced open source developers working on their own projects. The developers expected AI tools to speed them up and believed afterwards that they had. The measured result was that tasks took longer with the tools. This does not mean AI never helps. It shows that feelings about speed can differ from measured speed, so measure.

Then ask what happens to the saved time. Ten minutes saved per person per day is real only when it turns into something: more cases handled, faster response, lower overtime, or work not hired for. If the time simply disappears into the day, the benefit exists but cannot be booked as a cash saving. Say which kind of benefit you claim.

Benefits Beyond Time

Time is one benefit among several. Others include quality, such as fewer errors in invoices; speed, such as shorter customer wait; capacity, such as handling a peak without hiring; risk, such as catching contract issues earlier; and revenue, such as faster proposals. Name the metric for each and measure it before and after. Put a money value on a benefit only when you can justify it. Otherwise report it as a non-financial result next to the financial ones.

Be careful to avoid double counting. If the same ten minutes is counted as time saved and also as extra capacity, you have counted it twice. Keep a single list of benefits and tick each only once.

Building The Case With Ranges

Put cost and benefit in one table by year, and show three scenarios: low, expected and high. Include a sensitivity check: which input, if wrong, changes the answer the most? Often it is adoption, because a tool used by half the expected users produces half the benefit at nearly the same cost. Include payback time, which is the number of months until cumulative benefit exceeds cumulative cost.

Compare against the alternative as well. The alternative to the AI solution might be doing nothing, hiring, outsourcing, or using a simpler automation without a model. If a plain rule based automation delivers most of the benefit at lower cost and risk, choose it. This is a good result and not a failure of the AI programme.

Finance teams also like to see what you will do if the result is weaker than planned. Add a short note to the case that lists the trigger, for example adoption below a stated share after training, and the response, such as extra training, a narrower scope or a stop. This shows that the budget is protected and that the programme is managed with discipline. It also makes later reviews easier, because everyone agreed in advance what a weak result would mean.

Net Benefit By Adoption Level, One Use Case

thousand per year (illustrative)

Illustrative. Cost stays almost fixed while benefit rises with use, so adoption drives the result.

FinOps For AI

FinOps is the practice of managing cloud spend through shared responsibility between engineering, finance and business teams. For AI it brings a few habits that pay off from the first month. Tag every AI workload with a team and a use case, so spend can be traced. Set a budget for each use case and alerts at set percentages of that budget. Review the bill monthly with the product owner, not only with finance.

Add technical controls. Set maximum tokens per request, rate limits per user, and a kill switch for runaway loops. Use the smallest model that passes your evaluation. Track unit cost, such as cost per ticket resolved or per invoice processed, because total spend rises with success and unit cost tells you whether you are becoming more efficient. Review every licence quarterly and remove seats that nobody uses.

Key Terms

ROI: Return on investment, the net benefit divided by the cost.
Token: A small piece of text that language models read and write, and the usual unit of usage pricing.
Seat licence: A fixed fee per user per month for access to a tool.
Usage pricing: A charge based on how much of a service you consume, such as tokens.
Total cost of ownership: The full cost of a system over its life, including one time and running costs.
Baseline: The measured performance of a process before the AI change.
Payback time: The number of months until cumulative benefit exceeds cumulative cost.
FinOps: Managing cloud spend through shared work between engineering, finance and the business.
Unit cost: The cost of one outcome, such as one resolved ticket.

Exercise: Cost And Benefit Case For One Use Case

Take the process from week 7 and its minutes per step.
List all one time and running costs, including review time by people.
Estimate monthly token or licence cost with the formula, using a vendor price list.
Measure or estimate time saved per task, subtract review time, and state your confidence.
Build a table of low, expected and high scenarios with payback time.
Name the one assumption that would change your answer the most.