Free tool · no signup
Build vs buy an AI system: what it really costs over time
Put in a quote for the product you could buy and an estimate for building your own. The calculator runs 1,500 scenarios across your ranges, shows where the lines cross, and names the input that decides the answer. Every formula is on the page.
Use ranges (low, expected, high)
Build it
Engineering hours to first release
Low
Expected
High
Design, integration, evaluation and deployment.
Blended hourly cost
Your team cost or a quote. Used for build and integration hours.
Months until it is live
Low
Expected
High
Yearly maintenance
Low
Expected
High
Share of the build cost spent each year on fixes, upgrades and model changes. A planning range, not a measured figure.
Hosting and monitoring per month
Model usage per month
Low
Expected
High
API tokens once live. The AI model leaderboard lists current prices per million tokens.
Buy it
Subscription per month
Low
Expected
High
Use a real quote, including seats.
Usage fees per month
Low
Expected
High
Per resolution, per minute or overage charges on top of the plan.
Setup fee
Your rollout hours
Yearly price increase
Low
Expected
High
Months until it is live
Horizon and value
Compare over
36 months
What it is worth per month once live
Optional. When set, the option that goes live later carries the value it missed. Leave at 0 to compare cost only.
Over 36 months, on expected values
Building costs $29,640 less.
Building stays cheaper from month 11 onward. Across 1,500 runs sampled from your ranges, building came out cheaper in 100% of them, with a median breakeven at month 12.
Build total
$30,508
Buy total
$60,148
Breakeven
Month 11
Build cheaper
100% of runs
Share of simulated runs where building is cheaper
Cumulative cost, month by month
What decides the answer
Each bar shows how far build minus buy moves at the 36 month mark when one input swings from its low to its high value. Right of the line favours buying.
The part cost does not capture
Six questions that decide more build vs buy calls than the spreadsheet does.
Is this workflow part of how you win customers?
Does an existing product fit the process without workarounds?
Must the data stay in your own cloud account?
Will someone own the system after launch?
How soon must it be live?
Will the process change a lot in the next year?
No black box
Exactly how the numbers are worked out
Build upfront = build hours x hourly cost, spread over the months until live
Build per month = upfront x yearly maintenance % / 12 + hosting + model usage
Buy at month 0 = setup fee + rollout hours x hourly cost
Buy per month = (subscription + usage fees) x (1 + yearly increase) ^ completed years
Time to value = monthly value x months not yet live, added to whichever option is slower
Breakeven month = first month after which cumulative build cost stays at or below buy
Simulation = 1,500 runs, each input drawn from a triangular distribution
over your low, expected and high values (fixed seed, so results repeat)The starting values are examples, not market data. The build hours reflect the roughly three week scope of the agents on my agents page; the hourly cost is the BLS median developer wage with employer costs (agency rates are in my hourly rate guide). Replace every number with your own quote before you decide anything.
Questions
The assumptions, explained
Is 20% a year a fair maintenance figure for software I build?+
It is a planning assumption, not a measured fact. The famous "maintenance is 60 to 80% of lifetime cost" figure traces to Lientz, Swanson and Tompkins (1978), whose own abstract calls 75 to 80% a rough estimate, and when Schach and colleagues (2003) measured real code changes, the maintenance mix in that same survey turned out to be far off, so self reported survey figures make poor point estimates. That is why the calculator takes maintenance as a range and shows how much the answer moves with it.
Does an AI system need more upkeep than normal software?+
Usually, because the models underneath it retire. Across 21 retired or scheduled Anthropic and Google models, the median time from release to retirement was about 14 months, so a build should budget for at least one model migration and a fresh evaluation a year.
Where does the default hourly cost come from?+
The Bureau of Labor Statistics median wage for software developers ($135,980, May 2025) multiplied by its June 2026 employer cost load for professional occupations, which works out to about $95 per hour worked. Agencies and senior specialists charge more; put your own number in.
How much should I assume a subscription will rise each year?+
Sources disagree. Big vendors announced list price rises of roughly 5 to 17% on their main business suites in 2025 and 2026, and up to 43% on some smaller Microsoft plans, while the BLS producer price index for software publishers was flat in 2025 and slightly down in 2026. The default of 7% sits between them, and the range lets you test both ends.
Should cost decide build or buy?+
Rarely on its own. Time to value, fit to your process, data control and who will own the system after launch often matter more, which is why the page has the six non cost questions. Large IT projects also overrun: in a study of 5,392 projects, actual cost averaged 1.8 times the estimate even though the median project landed on budget, so give the build side an honest high value.
Want a second opinion on your numbers?
Send me the share link. I will tell you which inputs look optimistic and whether a build is worth it at all, even when the honest answer is to buy.
Book Meeting