AI concepts
Last mile problem (the final stretch of AI deployment)
A phenomenon in which a demo and a pilot show promising results, but integrating the model with the organisation's actual workflow turns out to be ten to twenty times more expensive than the model itself. The final stage, where AI plugs into the real processes of people and systems, absorbs 70 to 80 percent of the deployment budget. Most AI projects never reach that stage.
Primary source: BCG The Widening AI Value Gap 2025, MIT Sloan Management Review Q4 2024, Gartner Hype Cycle for AI 2025
The term is borrowed from logistics (last mile delivery), where the final leg of a delivery is the largest operating cost despite the short distance. In AI the mechanism is analogous: building a model that works under laboratory conditions is simple, getting it to work under an organisation's production conditions is the hardest part.
What the last mile actually consists of
BCG The Widening AI Value Gap of September 2025 breaks the deployment cost down into six stages. The first three (data preparation, modelling, prototype) typically absorb 20 to 30 percent of the budget. The next three (integration with legacy systems, change management, monitoring) absorb 70 to 80 percent.
Concrete examples of last mile costs: integration with SAP requires building connectors (3 to 6 months of developer work), training end users (1 to 2 months), data governance for a new data source (a compliance review), monitoring and alerting infrastructure, and a runbook for incident response.
Why companies are caught off guard
MIT Sloan Management Review Q4 2024 analyses 73 enterprise deployments. The conclusion: AI vendors present mainly the first three stages (model, data, prototype), because those are their expertise and show good results on slides. The last mile sits on the client's side, so it disappears from the conversation. The client signs a contract for 200 thousand zloty expecting that to be the total cost. After six months they realise it was 25 percent of the real cost.
Mitigations
Three approaches drawn from mature deployments. First, an end-to-end RFP: require a scope covering all six stages, not just the model. Second, a fixed total cost: a contract with a fixed price covering integration, change management and monitoring, not just the model. Third, capability transfer from day one: your people take part in every stage, so the last mile is internal rather than outsourced.
The Polish context
Gartner Hype Cycle for AI 2025, in the Polish context, shows that enterprise companies are currently in the trough of disillusionment for GenAI, precisely because of the last mile. The early 2023 to 2024 pilots did not scale to production. CFOs are more cautious before the next wave of investment. This is a good moment to start with a readiness audit rather than another pilot.
The yesfor.ai AI Readiness Audit includes an estimate of the last mile cost for the recommended use cases, on top of the cost of the model itself.