The AI-to-P&L Note · No. 3

Why AI pilots stall before they reach the P&L

Most AI pilots do not fail. They succeed at the wrong thing, produce a good demo, and then wait for a decision nobody owns. The cost is not the pilot budget. It is the year lost while the organisation believes it is making progress.

BCG’s 2025 research found that 60% of companies report little or no value from AI, and only 5% achieve value at scale. Behind that figure is a familiar pattern in financial services: a portfolio of pilots, each technically promising, none of them moving the income statement. Six patterns explain most of it.

1. The pilot measures the model, not the process

Pilots are judged on accuracy, precision or user satisfaction. The P&L cares about cost per case, cycle time, loss rates, recoveries and revenue. A model can score well and change none of those, because the surrounding process absorbs the improvement.

2. Nobody owns the number

Technology sponsors the pilot, the business tolerates it, and finance has never seen a baseline. When the pilot ends there is no executive whose target depends on scaling it, so it competes for funding against everything else, and usually loses.

3. The process was never redesigned

The model is bolted onto the existing workflow. Every old check still runs, every handoff remains. The pilot shows that the AI works and that the organisation does not change around it.

4. Governance arrives late

Risk, compliance and privacy see the solution after the demo. Questions about automated decisions under POPIA section 71, explainability and model risk are legitimate, and answering them after the build often means rebuilding.

5. Data and integration are “phase two”

Pilots run on curated extracts and manual uploads. Production requires integration with core systems, access controls and data quality the pilot never had to face. The real cost and timeline appear only when the pilot is declared a success.

6. There is no decision gate

A pilot without written scale-or-stop criteria never ends. It drifts into a permanent experiment, consuming attention and making the next pilot harder to justify.

A pilot is not an experiment in whether AI works. It is a rehearsal for running the business differently.

What a pilot built for the P&L looks like

  • A business case before the build, in the firm’s own numbers, with a baseline for the measure that will move.
  • A named business owner whose target depends on the outcome.
  • A redesigned workflow on paper, showing which steps change and what work stops.
  • A governance pack in parallel: decision inventory, section 71 position, explainability, monitoring and accountability.
  • A defined production path: integration, security, support and run cost, estimated up front.
  • Scale-or-stop criteria with a date, agreed by the executive committee before the pilot starts.
Test your current pilots

Take your three most advanced AI pilots and ask, for each: what is the baseline, who owns it, what changes in the workflow, and what is the decision date? If any answer is missing, that is where the pilot will stall.

The point

Pilots stall in the gap between a working model and a changed organisation. Closing that gap is operating-model work: ownership, process, governance and decision discipline. It is the work the Operating Model Diagnostic is designed to start.

Open the AI-to-P&L Calculator

Sources

  1. Boston Consulting Group, The Widening AI Value Gap, 2025: 60% of companies report little or no value from AI; 5% achieve value at scale.
  2. Protection of Personal Information Act 4 of 2013 (South Africa), section 71.
Next step

Apply this to your organisation.

The Operating Model Diagnostic gives your executive team a prioritised, governed view of where AI can move the P&L. Pricing on request.

Request an Operating Model Diagnostic