Investment and ROI
AI business cases tend to understate cost and overstate certainty. The model is usually the smallest and most visible line; integration, data preparation, evaluation and change management are larger and less visible. Benefits are real but harder to prove than in most IT investments, because they depend on how people and processes change around the system.
Cost lines
| Cost line | What drives it | Capex or opex |
|---|---|---|
| Licences | Per-seat or platform subscriptions for AI tools and assistants. | Usually opex. |
| Model usage | Tokens × price per token × volume of requests. Routing simple requests to a small model and only complex ones to a large model changes this line substantially. | Opex; varies with demand. |
| Infrastructure | Hosting for retrieval indexes, vector stores, self-hosted models, GPUs where used. | Capex if owned, opex if cloud. |
| Integration | Connecting to CRM, billing, product inventory, OSS, knowledge bases and channels. | Often capitalised build. |
| Data preparation | Cleaning knowledge articles, resolving identity across systems, labelling. | Often expensed; frequently underestimated. |
| Evaluation | Test sets, quality measurement, red-teaming, ongoing monitoring of answer quality. | Build and run. |
| Run and support | Platform operations, incident handling, prompt and model updates, vendor management. | Opex. |
| Change management | Training, process redesign, communications, workforce consultation. | Opex; frequently omitted. |
- Licences
- Model usage
- Integration
- Data preparation
- Evaluation set-up
- Infrastructure
- Run and support
- Ongoing evaluation
- Change management
- Training
- Workforce consultation
The capex/opex split matters to finance because AI shifts cost from one-off build to recurring usage. A care assistant that is cheap to launch can carry a run cost that grows with every contact it handles. That is acceptable when benefit grows in step; it becomes a problem when usage grows but measured benefit does not.
Benefit types and their uncertainty
- Avoided cost: less handling effort, fewer escalations, fewer truck rolls. Only real if the freed capacity is redeployed or not replaced; time saved that is absorbed elsewhere does not reduce cost.
- Revenue: better retention, faster B2B quotes, higher conversion. Hard to attribute because pricing, campaigns and competitor moves change at the same time.
- Risk reduction: fewer compliance breaches, earlier fraud detection, less revenue leakage. Valuable but expressed as expected loss avoided, which finance may discount.
Benefits are uncertain for three reasons: quality in a pilot rarely matches quality at scale; adoption depends on whether staff trust the output; and the baseline is often not measured before the change, so the improvement cannot be shown afterwards.
Why not every initiative should be funded
ROI is one input, not the decision. An initiative with modest return may be worth funding because it builds the shared retrieval layer that later initiatives need. One with an attractive return may be declined because the data is not trustworthy, the risk is unacceptable, or the organisation cannot absorb another change in the same team this year. Strategic fit, feasibility, risk and readiness carry equal weight.
Benefit realisation and ownership
The business owner — not the AI team — owns the benefit. Finance agrees the baseline before the pilot, the measure is reviewed at each gate, and post-production tracking continues for at least a full budget cycle. If the benefit is capacity, the owner must show where that capacity went.