What you'll learn: how to estimate an AI project in three parts (build effort, model token cost and running cost), combine them into a three-year total cost of ownership (TCO), and show which assumption moves the answer most. Every price on this page is a placeholder price; you replace each with your provider's current published price and its date.
At a glance
Figure 1. From scope to a TCO table with sensitivity cases
How to read this: three separate estimates meet in one table, and the last box is the one the sponsor needs: how wrong could this be.
- Build cost is people times days times availability times a day rate; model cost is contacts times tokens times price.
- In this project the model bill is about 1 percent of the three-year cost; hosting is the largest line.
- Label every price a placeholder, and record where the real price will come from and on what date.
- Show sensitivity cases so the sponsor sees which assumption matters.
Scenario
Composite scenario, not a real company. The Head of Customer Operations at Lumen Industrial Supply, a fictional distributor, has agreed to a read-only Order and Policy Assistant for 24 customer-service agents. Finance asks: "What does this cost over three years, and how wrong could you be?"
The design is fixed from earlier pages: a thin Python service behind the existing API gateway, a hosted model and keyword search over the policy library. All numbers are illustrative. You need a table where every line has a formula and an input.
What you will be able to do
- Derive build days and build cost from team size, duration, availability and a day rate.
- Compute tokens and model cost per month from stated assumptions.
- Build a three-year TCO table and a sensitivity table with correct arithmetic.
Prerequisites
- The executive architecture briefing, where the cost range is presented.
- AI Architect cost economics for the cost levers (caching, batching, cascades); not repeated here.
- Latency, tokens, cost and failure capture for recording real usage.
Core ideas
Three estimates, one formula each
| Estimate | Question | Typical error |
|---|---|---|
| Effort | How many person-days, and what do they cost? | Availability taken as 100 percent |
| Usage | What does the model cost each month? | A list price with no date |
| Running | What does it cost to keep alive? | Leaving out support |
build days = engineers x working days x availability + architect share x working days
build cost = build days x day rate
assisted contacts = contacts a month x assisted share
model cost a year = 12 x assisted contacts x (in tokens x in price + out tokens x out price) / 1,000,000
run cost a year = hosting a month x 12 + model cost a year + support share x build cost
three-year TCO = build cost + 3 x run cost a yearThese are a method, not a law; the reader can see and change every input. Total cost of ownership (TCO) means everything you pay over the life of the system, not only the build.
Where the price comes from
A price changes by provider, model and date. A provider's pricing page lists a price per million tokens (MTok), with separate columns for input and output (source: platform.claude.com pricing page, accessed 2026-11-03; this lesson quotes no prices from it). Replace every placeholder with the current published price, with its date and URL. The LLM cost estimator and LLM token counter help check the arithmetic.
Never paste a price into a proposal from memory or from a blog. An old price in a proposal is a quote you may have to honour.
In the interview: "Estimate the cost of an AI assistant for 24 agents." A strong answer asks for contacts per month, tokens per request and who supports it, states the formula, gives a range, and names hosting and support as costs the model bill hides.
Worked example
Assumptions: every number is illustrative; money is in EUR; every price and rate is a placeholder price.
- 9,600 contacts a month; the assistant handles 70 percent.
- One request is 3,200 input tokens and 350 output tokens.
- Placeholder price: 1.00 per million input tokens, 5.00 per million output tokens.
- Team: 2 engineers at 75 percent availability for 40 working days (8 weeks), plus an architect at 20 percent. Placeholder day rate: 450.
- Placeholder: hosting 1,200 a month; support is 20 percent of build cost a year.
- No cached tokens and no buffer; the sensitivity cases are the margin. The three-week proof of concept is not included: price it separately.
Step 1. Effort.
- Person-days: 2 x 40 x 0.75 = 60, plus 0.2 x 40 = 8, so 68 over the 8 weeks.
- Build cost: 68 x 450 = 30,600.
Step 2. Usage. Assisted contacts: 9,600 x 0.70 = 6,720 a month. Tokens: 6,720 x 3,200 = 21,504,000 in and 6,720 x 350 = 2,352,000 out. If the median handling time falls from 9 to 6 minutes, 3 x 6,720 = 20,160 minutes, or 336 hours a month, are freed. That is time freed, not cash saved.