Building the Business Case for AI, and the Operating Model to Scale It

After two years of experimentation, boards are asking a harder question about AI. Not "can we?" but "what is it actually worth?" The period of funding pilots on faith is ending. An AI business case now has to clear the same bar as any other capital allocation: a credible link to the P&L, an honest account of cost, and an operating model that turns the investment into results.

Most cases fail that test, and not because the technology is weak. They fail because they describe a tool rather than a change in how the business runs. This article sets out what a credible business case for AI contains, why it has to be designed together with the operating model, and how to reach a fundable version in about ninety days.

Why most AI business cases are wrong

Two failure modes dominate.

The first is cost-out theatre: a deck full of efficiency claims with no link to the P&L. "Much faster" means nothing if the freed time does not reduce cost, absorb growth without new hires, or improve something a customer will pay for. Unless the case says where the capacity goes, it has not identified a benefit. It has identified a feeling.

The second is revenue fantasy: speculative upside with no path to capture it. But a revenue line needs an owner, a mechanism and a baseline, and many AI cases have none of the three.

What a credible AI business case contains

A credible AI business case is specific about four things.

• The value lever. Cost reduction, revenue growth, risk reduction or cycle time. Pick one as the primary lever and state it in units the business already reports. A case that claims all four equally has not been thought through.
• The baseline. What the process costs, how long it takes and how often it goes wrong today, measured before anything is built. Without it, every later claim is an argument rather than a measurement.
• The capture mechanism. How the benefit reaches the P&L. Freed capacity is redeployed to named work, a vacancy is not filled, a contract is renegotiated, a loss rate falls. If nobody can describe the mechanism, the benefit will not appear.
• The named owner. A business leader, not the technology team, who signs up to the number and carries it in their budget once the system is live.

Ranges are more honest than point estimates. A case that shows a cautious, expected and stretch outcome, with the assumptions that separate them, earns more trust than a single confident figure. It also makes the later work of measuring AI return from pilot to P&L far easier, because everyone agreed in advance what would be counted.

The full cost, stated honestly

Licences and model usage are the visible part of the bill, and rarely the largest. The costs that sink AI cases are the ones left out of the first version:

• Data work. Finding, cleaning, permissioning and maintaining the data the system depends on.
• Integration. Connecting the model to the systems where work actually happens.
• Evaluation and testing. Building the test sets and evaluation harness that tell you whether the system is good enough, and whether it still is after each change.
• Change and adoption. Training, process redesign and the management time needed to make new ways of working stick.
• Run costs. Monitoring, support, model updates and usage charges that grow with adoption.
• Governance. Risk assessment, documentation, human oversight and audit, which scale with the consequence of the decisions being assisted.

A case that includes these looks less attractive. It also survives contact with reality.

Design the operating model with the business case

AI changes how a company works, not just what it spends. That is why the business case and the operating model have to be designed together.

• Where do decisions move? When AI handles triage, pricing or forecasting, accountability and skills shift with it. Decide who is answerable for an AI-assisted decision before the first one is made.
• Who owns the models in production? Performance drifts as data, customers and upstream models change. Someone must own monitoring, improvement and the budget for both.
• How does work get redesigned? The value is in the redesigned process: fewer handoffs, different roles, exceptions routed to people with the judgement to handle them.
• What is the governance? As decisions become AI-assisted, oversight and audit have to scale with them. An AI governance operating model gives the structure, and regulation such as the EU AI Act makes parts of it mandatory for some uses. Confirm which obligations apply to you with counsel.

The companies winning with AI are not the ones with the best models. They are the ones that redesigned the operating model around them.

None of this happens without people. Budget for organisational change and AI adoption inside the case, not as an afterthought once usage figures disappoint.

A portfolio, not a project

Treat AI as a portfolio: a small number of high-conviction bets that change the economics of the business, plus a managed pipeline of smaller efficiency plays. The big bets deserve executive sponsorship and patience. The small plays should be cheap, quick and built on shared platforms.

Fund in stages. Release money against evidence: a baseline measured, a prototype evaluated on real data, a limited rollout with benefits tracked, then scale. Agree the kill criteria at the start, while nobody is yet emotionally invested. Stop what does not work quickly and without blame, and double down where value is proven. A quarterly value review, with finance in the room, keeps the portfolio honest.

Where to start: a 90-day plan

• Weeks one to three: inventory. List every AI initiative in flight, including the ones business units bought on a credit card. Record owner, spend, value lever and evidence of benefit.
• Weeks three to six: select. Choose three to five candidates where the value lever is clear and the data exists. Measure baselines now.
• Weeks six to nine: build the cases. Full cost, benefit ranges, capture mechanism, named owner and the operating model changes each one needs.
• Weeks nine to twelve: decide. Take the portfolio to the executive team with stage gates and kill criteria. Stop or merge the initiatives that cannot produce a case.
• Throughout: involve finance. A case that finance helped to build is a case that finance will defend.

Common questions

How do you calculate ROI on an AI project?

Start with a measured baseline, then estimate the change in cost, revenue, risk or cycle time as a range rather than a single number. Set that against the full cost of ownership, including data, integration, change and run costs. Then define how the benefit will be captured in a budget, because uncaptured savings do not count as return.

What costs are usually missing from an AI business case?

Data preparation, integration with core systems, evaluation and testing, change management and the ongoing cost of running and monitoring the system. Governance and compliance effort is also commonly left out. Together these can easily outweigh the cost of the model or licence itself.

Who should own the AI business case?

A business leader who will carry the benefit in their own budget, supported by technology and finance. When the technology function owns the case alone, the project tends to deliver a working system and no change in results.

Where Ganexa can help

Ganexa's AI Business Case and Operating Model work helps leaders build the financial case, prioritise the portfolio and design the decisions, accountability and governance that turn AI ambition into measurable results, and our AI strategy and implementation team carries the chosen bets into delivery. If you are preparing an AI investment for board approval, book a free consultation.