Most AI budgets are built backwards. A team prices licences and inference, gets a number that looks manageable, then discovers nine months later that the software was the cheapest line on the invoice.
What is actually changing in the market
Model capability is no longer the constraint. Access to capable models is close to commoditised, and the gap between vendors narrows with each release. What has not commoditised is the ability to connect a model to messy internal data, define who is accountable when it is wrong, and change how people work around it. That is where the cost sits, and it is also why capability alone has stopped being a competitive advantage.
Where the money actually goes
- Data preparation — cleaning, structuring and permissioning the information the system needs. Routinely the largest line, and almost never the one budgeted for.
- Integration — connecting to the CRM, document store, ticketing system or ERP where the work actually happens.
- Governance — review workflows, access control, audit trails and escalation paths. Non-negotiable in regulated environments.
- Change management — the training and process redesign that determines whether anyone uses it.
- Model and infrastructure — real, ongoing, and usually the smallest of the five.
Why a conventional approach struggles here
A software vendor scopes the licence. A systems integrator scopes the build. A management consultancy scopes the strategy. Each is competent within its boundary, and the failures happen between them — a model chosen before anyone checked the data, a workflow automated before anyone decided who signs off on its output.
This is the specific reason we assess readiness before recommending spend. Our executive guide to AI in business scores use cases on value, feasibility and risk, and names the ones we recommend avoiding. That output is frequently worth more than the build it precedes, because a use case that should not proceed is the cheapest possible finding.
What sensible sequencing looks like
Start where volume is high, rules are bounded and errors are cheap to correct. That produces a measurable return within a quarter and builds the internal confidence harder use cases require. Leading with a flagship customer-facing system is the most common route to a stalled programme, because the risk arrives before the experience does.
Budget in phases rather than as one number. Phase one is assessment and a contained build. Phase two extends into adjacent processes using what phase one proved. Phase three is where governance and scale spending become justified — by evidence rather than by optimism.
What this means for your business
If you are being quoted a single figure for AI transformation before anyone has examined your data or asked who reviews the output, you are being quoted for software rather than for outcomes. Ask which of the five cost categories above the number covers.