Governing AI Investment: A Model for Leadership

Governing AI investment means deciding how much to invest in artificial intelligence, where and with which goals, and regularly checking that spend produces adoption and return. It is a management process, not a technology project: it defines who decides, on what information and how often.

The need comes from a recurring pattern: AI investment grows faster than the ability to govern it. Licenses, usage and projects are bought by different functions on separate budgets, and leadership struggles to answer simple questions: how much do we spend, on what, with what result. This article proposes a three-phase model, the roles that support it and the minimum information that makes it work.

Why does AI investment need specific governance?

AI has features that set it apart from other technology investments.

• It is distributed. Any function can enable a tool, often with a small purchase.
• It is variable. A growing share of spend is usage-based and changes month by month.
• Its value is uncertain. Return depends on actual usage and must be measured process by process.
• It carries obligations. GDPR, NIS2 and the EU AI Act apply to tools in different ways.
• It changes quickly. Vendors update models, plans and terms frequently.

An annual budget cycle alone cannot keep up with this pace. You need a process that combines annual planning with a monthly reading of the data.

What is the model: plan, measure, decide?

The model has three repeating phases.

Plan

Set the AI budget by function, process and technology, the adoption and return targets for the main use cases and the purchasing rules (who can enable a tool, with which approval). Details are in enterprise AI budget: how to plan it.

Measure

Every month, compare spend, usage and impact against budget and expectations. Measurement covers three levels kept separate: real adoption, process impact with a before and after baseline, and impact on people in aggregate form. The method is described in how to measure AI ROI.

Decide

Use the data to reallocate resources: reduce or consolidate tools that do not deliver, invest more where results are proven, start training and compliance work where needed. Record each decision together with its rationale, so it can be checked in the next cycle.

Who decides, and with which roles?

The model needs clear roles, otherwise decisions stay inside individual functions.

• AI committee. CEO or general manager, CFO, Chief AI Officer or CIO, heads of the main functions. It approves budget and priorities and reviews results every quarter.
• Operational owner. Usually the Chief AI Officer. Manages the use case portfolio, prepares information for the committee and proposes decisions.
• Function contacts. Accountable for AI spend and results in their area.
• DPO and CISO. Assess the obligations and risks of each tool before adoption and during use.
• HR. Oversees training, AI literacy under article 4 of the EU AI Act and impact on people.

The operational owner's tasks are covered in Chief AI Officer and AI team.

What information does leadership need?

The committee does not need extensive dashboards. It needs six pieces of information, always the same, updated regularly.

1. Total investment: how much has been invested in AI and in what (licenses, usage, projects, training).
2. Budget versus spend: actuals by item and function, with variances explained.
3. Processes and technologies: which processes AI has been introduced into, with which tools and which overlaps.
4. Real adoption: usage by department, role and use case, including usage on personal accounts.
5. Measured impact: return of use cases with a baseline, distinguishing measurements from estimates.
6. Priorities for action and investment: what to reduce, what to consolidate, where to invest more.

This information requires spend, usage, use cases and obligations to be linked. Collected from separate sources, it reaches the committee late and is not comparable. The AI use case register is the starting point to keep it together.

At what cadence?

The cadence follows how quickly each piece of information changes.

• Monthly: spend, variances, adoption. These are the most variable data.
• Quarterly: use case impact, priorities, compliance status. The committee meets on these topics.
• Yearly: next year's budget, contract renewals, review of the AI policy.

A good rule: every committee meeting ends with recorded decisions and an owner for each. A meeting that only reads the data governs nothing.

Which mistakes make governance ineffective?

• Governing only risk. Governance limited to compliance says nothing about cost and return.
• Governing only cost. Cutting licenses without reading impact can remove the use cases that deliver.
• Non-comparable data. Each function measures its own way and the numbers do not add up.
• Assessing individuals. Usage data must always be aggregated by role or department, as explained in monitoring AI use and the Italian Workers' Statute.
• No decisions. Periodic reports without a moment of decision only produce documentation.

How do you launch the model in the first 90 days?

The model does not require a reorganization. It can be launched in three months with small, verifiable steps.

1. Weeks 1 and 2: appoint the roles. AI committee, operational owner and function contacts, with a written note stating who decides on budget, purchases and priorities.
2. Weeks 3 to 6: build the data foundation. Inventory tools, contracts, usage and use cases; map spend to functions and processes; measure real usage by department and role.
3. Weeks 7 and 8: set budget and targets. For the main items, a budget for the rest of the fiscal year and an adoption and return target, with the baseline to start from.
4. Weeks 9 and 10: define the rules. Approval procedure for new tools, criteria to close a use case, rules on people data agreed with the DPO and HR.
5. Weeks 11 and 12: first committee meeting. Review of the six pieces of information, first recorded decisions, calendar of the next meetings.

The first cycle is mainly about making spend visible and finding out where data is missing. It is normal for some information, such as measured impact, to be available only for a few use cases: what matters is to state it and plan the missing measurements. From the second quarter the committee works on comparable data and decisions become faster.

A practical way to choose where to start is economic weight: the items that together make up most of the spend, usually the most widespread licenses and usage-based spend, are those where the first cycle produces the most useful decisions.

Where Delphis fits

Delphis is a platform to govern AI investment, independent of vendors. It brings together in a single view the six pieces of information leadership needs (investment, budget and spend, processes and technologies, adoption, impact and priorities) and supports the plan, measure, decide cycle with data updated every month. See the platform overview.