Buying or building AI: the complete cost calculation

Buying an AI system gives an organisation access to a capability and changes a workflow. Building it adds responsibilities for design and operation. For me, the comparison therefore starts with the same question: what does a professionally usable result cost under the conditions in which the organisation actually intends to work?

The price of an individual model request answers only part of that question. Integration, output review, corrections, support and subsequent changes belong in the same calculation. This article uses an entirely fictional example. Every price, duration, volume and discount rate is an assumption. They describe no offer from aiomics, observed market price or reported customer result.

Define a comparable service first

A business group wants to convert 60,000 internal supplier documents each year into structured, checked summaries. It compares a purchased application with an internally developed application using an available model. Internal development here involves neither training a foundation model nor complete independence from external components. Both routes must achieve the same professional quality, permitted data scope and operational availability.

Documents containing patient information and medical decisions are excluded. The example assumes that every case is eventually completed to an acceptable standard after the planned review and any correction. A real pilot would separately record abandoned attempts, complete manual reprocessing and missing outputs. Otherwise, a low price per generated draft can conceal the most expensive cases.

Sculley and colleagues’ 2015 paper describes maintenance risks arising from dependencies and additional integration code in machine-learning systems. It provides no current price comparisons. I use it as a reason to include the surrounding software explicitly in a cost assessment. [1]

Separate initial and recurring commitments

For purchasing, I assume one-time costs of €32,000 for integration, €18,000 for initial evaluation and €10,000 for introduction and training: €60,000 in total. For internal development, the assumptions are €180,000 for development including integration, €36,000 for evaluation and €12,000 for introduction and training, totalling €228,000.

Annual fixed costs for purchasing are €48,000 for the application, €18,000 for integration maintenance, €18,000 for ongoing quality assurance and €24,000 for organisational support. This totals €108,000. For internal development, the assumptions are €24,000 for fixed technical infrastructure, €24,000 for integration maintenance, €30,000 for quality assurance and €24,000 for support: €102,000 altogether.

In this model, the fixed technical amounts exclude model usage that varies with the number of cases. Quality assurance covers system-level checks, such as assessments after a change. Individual output review is charged per case. This separation is intended to prevent double counting. Internal hours must likewise be assigned to a single cost item. Existing employees consume available working time even when no additional invoice arrives.

Calculate review and rework per case

For both options, I assume three minutes of professional review per case. An hour of work is valued at an assumed €60, including the employment overhead allocated in this example. Three minutes therefore cost €3. An assumed 20% of cases need another five minutes of correction. Averaged across all cases, that adds one minute and €1.

Technical usage is assumed to cost €0.45 per case for the purchased application and €0.12 for the internally developed one. Including review and correction, the totals become €4.45 and €4.12 per completed case. The €0.33 difference remains relevant. Its importance depends on the fixed and initial commitments it needs to offset.

At 60,000 cases, annual variable costs are €267,000 for purchasing and €247,200 for internal development. Adding fixed costs produces €375,000 and €349,200 per operating year. Including initial work, the first year costs €435,000 for purchasing and €577,200 for internal development.

Continue the comparison through the transition

The UK’s 2026 Green Book addresses appraisal over the full lifetime, alternative uses of resources and sensitivity analysis. It guides UK public decisions. The company comparison below borrows general questions from it and establishes no German accounting or procurement requirement. [2]

For a deliberately bounded comparison over three operating years, I also assume €15,000 to transition away from the purchased solution and €25,000 for the internally developed application. These amounts might cover data export, handover and decommissioning. Any residual value of the internally developed software is initially set to zero. This is a conservative modelling assumption that a team can change with a reasoned justification.

Without discounting, purchasing costs €60,000 + 3 × €375,000 + €15,000 = €1,200,000. Internal development costs €228,000 + 3 × €349,200 + €25,000 = €1,300,600. Across 180,000 completed cases, these totals are approximately €6.67 and €7.23 per result. Internal development costs €100,600 more in this example.

The timing of costs can also affect the ranking. Using a purely illustrative annual discount rate of 8%, initial costs occur at time zero, operating costs at each year’s end and transition costs at the end of year three. Present costs are then approximately €1,038,319 for purchasing and €1,147,768 for internal development. The 8% rate is purely illustrative; an organisation needs its own justified rate.

VAT, income taxes, price increases and financing arrangements are excluded here. This simplification keeps the arithmetic inspectable. Relevant cash consequences must be added for an actual decision. A longer useful life, different introduction dates or an evidenced residual value also requires a fresh comparison over a common time horizon.

Find the assumption that changes the decision

With quality and every other assumption unchanged, internal development becomes cheaper in the simple, undiscounted three-year comparison at about 161,617 cases a year. The calculation starts with €178,000 of additional initial and transition costs and subtracts €18,000 of lower fixed costs over three years. The remaining €160,000 is divided by three years and the €0.33 advantage per case.

This threshold applies only while costs remain linear and capacity is sufficient. More volume may require additional support or infrastructure; discounts may change the purchased option. With the assumed discounting, equal present costs occur at about 188,697 cases a year. The threshold is therefore a question to examine in planning. It establishes no general rule about the market.

Review time deserves at least as much attention. If one option requires an additional minute of review on average, it costs €60,000 more each year at our volume. Over three years, the undiscounted difference is €180,000. A small difference in practical usability can therefore outweigh the assumed price advantage. The next pilot should measure this uncertain quantity directly.

Distinguish released time from available cash

Suppose the existing process additionally takes an assumed twelve minutes per case. The average four minutes for review and correction would mathematically release 8,000 gross hours of case-handling work each year, before deducting time spent on shared support and quality assurance. This begins as a capacity assumption. Whether it produces shorter waits, more completed cases or less paid overtime depends on a specific organisational change.

Time must appear in the calculation only once. If the analysis compares existing resource costs with the new total, it cannot then deduct the same released hours again as an additional benefit. For cash planning, finance should separately identify which payments will actually disappear or arise. Salary expenditure does not automatically fall when an individual task becomes faster.

I would end the decision brief with three open questions: which quality differences could change review time, which commitments arise when changing or exiting the system, and which other work needs the same development capacity? The lower-cost calculation is an important finding. A sound decision also requires the organisation to sustain the assumptions and commitments behind it.