AI Business Case: How to Build One Step by Step

An AI business case is the document that justifies an investment in artificial intelligence by comparing expected benefits with full costs, on a bounded process and over a stated time horizon. It is not meant to persuade the committee. It puts assumptions, thresholds and exit conditions in writing, so that the decision can be checked afterwards.

A solid business case has seven components. We go through them in order, with a hypothetical numerical example at the end.

Which problem and which process do you start from?

The starting point is an operational problem, not a tool. "We want to introduce an AI assistant" is not a problem. "First review of standard contracts takes too long and causes rework" is.

The process must be bounded: a clear start and end, repeated volumes, an owner, a known group of users. The tighter the scope, the easier it is to isolate the effect of AI. If there are many candidate use cases, an AI use case register helps you choose where to begin.

Why do you need a baseline before estimating benefits?

Without a measurement of the process before AI there is nothing to compare against, and every claimed benefit becomes an opinion. The baseline collects a few indicators: volumes, average time per unit, error or rework rate, hourly cost of the staff involved.

The measurement must cover a period long enough to include normal variability (peaks, seasonality). If measuring is not possible, the business case says so and uses estimates from the process owner, marked as such. For the before-and-after comparison method, see how to measure AI ROI.

Which benefits can you estimate?

Benefits fall into three families, to be kept separate.

• Time. Net minutes saved per unit, after subtracting the time spent reviewing outputs. Freed time becomes value only if it is reused, so estimate what share really will be.
• Quality. Fewer errors and less rework, with an avoidable cost you can calculate (number of cases avoided times the cost of each case).
• Revenue. Faster response times, greater sales capacity, new services. These are the most uncertain benefits: if included, they go in a separate scenario, not in the base case.

A prudent rule: the base case contains only benefits that can be traced to a baseline indicator.

Which costs should be included?

The full cost is always higher than the license alone. The minimum items are:

• licenses or usage fees (per seat or per token, depending on the contract);
• initial integration and configuration (one-off);
• user training (working hours, not just the course fee);
• governance and maintenance: updating prompts and knowledge bases, controls, support;
• time spent reviewing outputs, already counted in the net time benefit.

For the overall cost structure, see the hidden costs in enterprise AI TCO. Prices must come from contracts, never from generic price lists.

Which time horizon should you choose?

For a bounded use case, 24 or 36 months is a reasonable horizon: long enough to amortize one-off costs, short enough not to depend on distant assumptions. The business case should separate one-off and recurring costs and allow for a ramp-up phase in which adoption grows gradually and benefits are only partial.

How do you run the sensitivity analysis?

Sensitivity analysis shows how much the result depends on the most fragile assumptions. Identify the two or three variables that weigh most (usually the share of time reused, the adoption level, the integration cost) and recalculate the result with worse values. If the project holds up only in the optimistic scenario, the business case must say so.

Which stop criteria should you set?

Stop criteria are measurable thresholds, decided before launch, below which the project is redesigned or stopped. For example: adoption below a given share of the planned users after six months, net saving per unit lower than one third of the expected one, quality worse than the baseline. A business case without stop criteria is not a decision. It is a bet.

Worked example: reviewing standard contracts

All values are hypothetical and for illustration only. A contracts office of 15 people handles 600 files a month, with 30 minutes of average work per file and 8% of files reworked (about 1 hour each). Fully loaded hourly cost of staff: €40.

Expected benefits at steady state

• Time: from 30 to 21 minutes per file, review included, so 9 net minutes. 600 × 9 = 5,400 minutes a month, or 90 hours; 90 × 12 = 1,080 hours a year. Assume only 50% is reused: 540 hours × €40 = €21,600.
• Quality: rework from 8% to 6%. Annual files are 600 × 12 = 7,200; two percentage points are 144 files avoided × 1 hour × €40 = €5,760.
• Revenue: none included in the base case.
• Annual total at steady state: €21,600 + €5,760 = €27,360.

Costs (hypothetical prices)

• Licenses: 15 users × €30 a month × 12 = €5,400 a year.
• Integration: €8,000 one-off.
• Training: 15 people × 4 hours × €40 = €2,400 one-off.
• Governance and maintenance: €3,000 a year.
• Year 1: €5,400 + €8,000 + €2,400 + €3,000 = €18,800. Years 2 and 3: €5,400 + €3,000 = €8,400 each.

Result over 36 months, base case. In the first year benefits are 60% of steady state: €27,360 × 0.6 = €16,416.

• Total benefits: €16,416 + €27,360 + €27,360 = €71,136.
• Total costs: €18,800 + €8,400 + €8,400 = €35,600.
• Net benefit: €71,136 − €35,600 = €35,536.
• ROI at 36 months: €35,536 ÷ €35,600 = 99.8%.
• Payback: after year 1 the cumulative figure is €16,416 − €18,800 = −€2,384. In year 2 the monthly net is (€27,360 − €8,400) ÷ 12 = €1,580. It takes 2,384 ÷ 1,580 ≈ 1.5 months, so payback occurs around month 14.

Sensitivity: only 25% of time reused. Time benefit: 1,080 × 0.25 = 270 hours × €40 = €10,800. Annual total at steady state: €10,800 + €5,760 = €16,560; in the first year €16,560 × 0.6 = €9,936.

• Total benefits: €9,936 + €16,560 + €16,560 = €43,056.
• Net benefit: €43,056 − €35,600 = €7,456. ROI: 7,456 ÷ 35,600 = 20.9%.
• Payback: cumulative after two years = (€9,936 − €18,800) + (€16,560 − €8,400) = −€8,864 + €8,160 = −€704. With €680 of net benefit per month in year 3, it takes about 1.04 months: payback at month 25.

The project stays positive but much less attractive, and the decisive variable is how much time is reused, not the technology. Possible stop criteria: fewer than 11 active users out of 15 after six months, or net saving below 3 minutes per file.

Recurring mistakes in AI business cases

1. Estimating benefits without a baseline.
2. Counting gross time saved and ignoring review.
3. Treating all freed time as realized value.
4. Leaving integration, training and governance out of the costs.
5. Mixing uncertain revenue benefits with time benefits.
6. Setting neither stop criteria nor a review date.

Where Delphis fits

Delphis is a platform, independent of AI vendors, that connects spend, adoption, impact and regulatory obligations. For the business case it reconciles costs with actual usage, simulates the possible saving and compares expected and realized impact, with before-and-after measurements on bounded processes. See the adoption and ROI page.