AI ROI & the Business Case
An AI business case is the honest accounting of what an initiative will cost end-to-end and what it will return in money and reclaimed time. It matters because the headline numbers are seductive and misleading: McKinsey’s analysis of 340 deployments found a median ROI of 210% over three years, yet only about 6% of programs pay back in under 12 months, and integration, data preparation, and change management routinely add 20–40% to the cost nobody budgeted for (McKinsey, State of AI; industry ROI research, 2025). A real business case counts the full cost of ownership, values the return in the units that matter, and, crucially, prices the cost of doing nothing.
- 01McKinsey’s study of 340 deployments found a median ROI of 210% over three years; mature programs report 200–400% over 3–5 years (McKinsey, State of AI).
- 02The license is rarely the real cost: integration, data preparation, and change management add 20–40% on top of the sticker price (industry ROI research, 2025).
- 03Only about 6% of AI programs pay back in under 12 months; 12–18 months is the norm, so the business case must be built on that horizon, not a quarter.
- 04The biggest return is usually capacity, not layoffs: AI lets the same team handle more, employees report roughly 40% higher productivity on suitable tasks, which grows the business rather than shrinking payroll.
- 05The cost of doing nothing is real and compounding: competitors who adopt (the top 12% "vanguard" see 2.5x higher returns) widen the gap every quarter you wait (McKinsey, State of AI).
Total cost of ownership
The true cost of AI is the license plus integration, data preparation, change management, and ongoing maintenance. These "hidden" layers typically add 20–40% to the sticker price, and are the single most common reason a business case that looked profitable turns out not to be.
The subscription fee is the part everyone sees and the part that matters least. The real spend is getting the AI connected to your systems, cleaning the data it needs, and getting your team to actually use it, change management alone can be the largest line. Successful programs plan for this, allocating roughly 70% of effort to people and process, 20% to data and technology, and only 10% to the algorithms themselves (industry ROI research, 2025).
If a vendor quote is only the license fee, it is not a budget, it is a down payment. Add integration, data cleanup, and the hours your team spends changing how they work. That fully-loaded number is the one your business case has to beat.
Build vs buy vs subscribe
Subscribing is fastest and cheapest to start but you rent the capability and control nothing. Buying/building is slowest and most expensive up front but you own an asset that compounds. Most businesses should subscribe to prove value, then build the workflows that become a durable advantage.
The decision is not ideological, it is about where the capability sits on your critical path. Rent commodity capabilities (a general chatbot, transcription) where owning them buys you nothing. Own the workflows that are specific to how you win, because a rented workflow can be shut off, repriced, or copied by every competitor on the same subscription. The mistake is building what you should rent, or renting what should be your moat.
- Subscribe, fastest, lowest upfront cost, zero ownership. Best for proving value and for commodity capabilities you do not compete on.
- Buy / integrate, middle path: assemble owned workflows on top of vendor models. Best for the processes specific to your business.
- Build, highest cost and time, full control and ownership. Justified only where the capability is a genuine, durable competitive advantage.
Labor arbitrage vs capacity
There are two ways AI returns money: cut labor cost (arbitrage) or let the same team do more (capacity). Capacity is almost always the bigger prize, it grows revenue instead of just shrinking a cost line, and it does not carry the morale and knowledge-loss cost of layoffs.
Framing AI purely as headcount reduction caps its value at the payroll you can cut and demoralizes the team you need to adopt it. The larger return is capacity: employees report around 40% higher productivity on suitable tasks, which means the same people handle more volume, respond faster, and take on work you previously turned away (industry productivity research). For most growing businesses, capturing new capacity is worth far more than the salary of the roles AI could replace.
You can only cut a cost line to zero once. But capacity, handling more customers with the same team, compounds as you grow. Build the business case on the revenue that new capacity unlocks, not just the payroll it saves.
Quantifying reclaimed time
Reclaimed time becomes money one of two ways: it is redeployed to revenue-generating work, or it is eliminated as cost. Time that is merely "freed" but reabsorbed by low-value busywork has no ROI, so a real business case names where the reclaimed hours go.
This is where soft business cases fall apart. "Saves 10 hours a week" is meaningless until you say whose hours, valued at what rate, redeployed to what. If those 10 hours let a salesperson make more calls, it is revenue. If they let you avoid a hire, it is cost avoided. If they just evaporate into more email, the ROI is zero regardless of the hours saved. Name the destination of every reclaimed hour, or do not count it.
- Count the hours at the process level, which task, whose time, how many hours before vs after.
- Assign a rate, the loaded cost of that person’s hour, or the revenue that hour can generate if redeployed.
- Name the destination, redeployed to revenue work, or eliminated as cost. If neither, the hours do not count.
The cost of doing nothing
Inaction is not the zero-cost option, it is a compounding cost. Competitors who adopt widen their advantage every quarter, and the top 12% of adopters already see 2.5x higher returns. The business case for AI has to be measured against a moving baseline, not a static one.
Every business case implicitly compares "adopt" against "do nothing", and treats "do nothing" as free. It is not. The top 12% of organizations, the "vanguard" embedding AI across four or more functions, report 2.5x higher returns than their peers, and that gap compounds as they reinvest it (McKinsey, State of AI). Meanwhile 42% of companies abandoned most AI initiatives in 2025, up from 17% the year before, usually from skipping the honest business case this pillar describes (S&P Global, 2025). The risk is not that AI fails; it is doing it without a plan while disciplined competitors pull ahead.
Build vs buy vs subscribe, how the three paths compare
| Dimension | Subscribe | Buy / integrate | Build |
|---|---|---|---|
| Upfront cost | Lowest | Medium | Highest |
| Time to value | Days to weeks | Weeks to months | Months+ |
| Ownership | None, you rent | Partial, owned workflows on vendor models | Full, you own the asset |
| Control & customization | Limited to vendor options | High on your layer | Complete |
| Risk if vendor changes | High, repricing or shutoff | Medium, model swappable | Low, you control it |
| Best for | Proving value; commodity capability | Processes specific to your business | Durable competitive advantage only |
Frequently asked questions.
What is a realistic ROI to expect from AI?
McKinsey’s analysis of 340 deployments found a median 210% ROI over three years, with mature programs reaching 200–400% over 3–5 years. But returns concentrate in programs that scoped full cost and value honestly, the average includes many that lost money.
What costs do businesses forget when budgeting for AI?
Integration, data preparation, and change management, which together add 20–40% on top of the license. Successful programs spend roughly 70% of effort on people and process, only 10% on the algorithms themselves (industry ROI research).
Should I build my own AI or subscribe to a tool?
Subscribe to prove value and for commodity capabilities; own the workflows specific to how your business wins. Build only where the capability is a durable competitive advantage, building what you should rent wastes money and time.
Is doing nothing a safe option?
No. Inaction is a compounding cost. The top 12% of adopters already see 2.5x higher returns, and that gap widens each quarter. The business case must compare adoption against a competitor who is moving, not a static baseline.
Five deep dives in this pillar.
The True Total Cost of Ownership for AI
The license plus integration, data preparation, change management, and ongoing maintenance. These layers typically add 20–40% to the sticker price and are the most common reason an AI business case that looked profitable turns out not to be.
Build vs Buy vs Subscribe
Subscribe to prove value and for commodity capabilities; buy/integrate for workflows specific to your business; build only where the capability is a durable competitive advantage. The mistake is building what you should rent, or renting what should be your moat.
Labor Arbitrage vs Capacity Gains
Both are possible, but capacity almost always wins. Cutting labor caps out at the payroll you can remove; capacity, the same team handling more, grows revenue and compounds as you scale, without the morale and knowledge-loss cost of layoffs.
Quantifying Reclaimed Time
Count the hours at the process level, assign a loaded rate, and name the destination, redeployed to revenue work or eliminated as cost. Time that is freed but reabsorbed by busywork has zero ROI, no matter how many hours it is.
The Cost of Doing Nothing
More than it looks. Inaction is a compounding cost: the top 12% of adopters already see 2.5x higher returns, and that gap widens every quarter. The right comparison for any AI business case is against a moving competitor, not a static baseline.