Why Most AI Pilots Don’t Reach the P&L

A convincing demo, an unchanged P&L

Over the past twenty-four months, almost every mid-to-large company has approved budgets for generative AI copilots, document-extraction engines, predictive maintenance models or customer-service bots. In the boardroom demo, the technology works.

When the finance team closes the books twelve months later, however, the P&L looks remarkably similar to the year before. Across three major global executive studies published in 2025, the gap between AI deployment and audited financial return is consistent:

Why do three out of four AI investments stall between a working pilot and the income statement? When researchers at the RAND Corporation (August 2024) interviewed 65 data scientists and engineers with at least five years of experience building enterprise AI systems, they found that project failures almost never stemmed from the choice of algorithm. They stemmed from five upstream management errors: misunderstanding the operational problem to be solved, lack of clean operational data, focus on using the latest technology for its own sake, inadequate workflow infrastructure, and asking AI to solve problems beyond its boundary conditions.

Four leaks where AI value disappears before the P&L

When we audit stalled digital and AI initiatives, the promised value typically leaks out in four operational places:

1. Automating a step without redesigning the workflow around it

In McKinsey's August 2025 survey, out of 25 management practices tested, fundamentally redesigning workflows had the single strongest statistical correlation with achieving enterprise-level EBIT impact from AI—yet only 21% of organisations had fundamentally redesigned even part of their workflows around AI.

If a six-stage credit-approval or store-opening pipeline takes 45 days because of sequential handoffs and committee sign-offs (Case 01), putting a generative AI assistant inside Stage 2 to draft a memo in ten minutes instead of four hours saves 3 hours and 50 minutes of individual typing time. The file still waits 44 days in the same departmental queues. You have paid for software licences to create scattered pockets of "saved minutes" that never convert into lower headcount, higher throughput or faster cash collection.

2. Automating a record that already exists somewhere else

Before training an optical-character-recognition (OCR) or large-language-model pipeline to extract and reconcile data from messy internal documents, few teams stop to ask a simpler question: Does a clean, verified record of this transaction already exist inside the closed world of our operations?

When you build AI to read, clean and reconcile a redundant administrative step, you end up paying twice—once for the software and cloud inference, and a second time for the human exception-handlers who still have to check the 8% of fields where the model hallucinates or misreads a crumpled receipt.

3. Orphaned ownership between IT and the P&L owner

Most AI pilots are funded out of a central technology or innovation budget and measured on technical metrics (model accuracy, latency, token cost or weekly active users). Because the business-unit P&L head did not have to give up operational budget to build the tool—and is not held accountable for reducing unit cost once the tool goes live—operations managers treat the AI output as an optional "second opinion" alongside their legacy manual process. Running both the old process and the new AI tool in parallel increases cost instead of reducing it.

4. No pre-agreed baseline or unit-economics gate

Teams regularly launch AI proofs of concept without recording the baseline cost per transaction, error rate or cycle time before the pilot starts. Six months later, when the vendor licence comes up for renewal, nobody can prove whether a shift in output came from the AI model, a seasonal volume change or overtime.

What this looks like in practice: The software that was not built

The sharpest test of an AI or digital proposal is whether redesigning the operational rule removes the need for the software altogether.

In The Expense System That Did Not Need to Be Built (Designed), our client's finance operations team wanted to build a new digital expense-claims platform to fix a slow, error-prone travel reimbursement workflow. The company had already digitized its paper forms and integrated Bank of Thailand (BOT) daily foreign-exchange rates. Even so, travelling staff still spent hours keying individual meal and taxi bills into the system, and finance clerks still spent days manually checking OCR mismatches and currency conversions against attached receipts.

Instead of specifying a more sophisticated AI receipt-scanning engine, we looked at the resources already present inside the transaction:

• Every card payment made by an employee already generated a bank statement line item containing the exact merchant name, date, foreign currency amount, converted Thai Baht amount and bank-verified exchange rate.
• The company's external auditors confirmed that an official corporate credit-card statement was already a fully accepted tax and audit record.

By issuing a company travel card for routine travel spend and defining clear, simple exception routes for small cash or non-card vendors, the bank statement became the expense report. The custom software did not need to be built, the manual keying and OCR reconciliation disappeared, and the solution carried near-zero development cost.

When AI is the right lever—for example, in high-volume engineering document verification or multi-workstream scheduling like Store-Development Cycle Compressed Across Six Workstreams (Client-measured)—we apply the exact same sequence: redesign the workflow dependencies and decision gates first, eliminate the redundant steps second, and point AI only at the irreducible cognitive bottleneck that remains.

Where to start on Monday

Before signing your next AI vendor contract or approving a pilot extension, run every use case through these three gates:

1. Run the "Subtraction Test" before writing a single prompt or spec. Ask: "If we changed the purchasing rule, payment instrument or approval threshold, would this document or reconciliation step need to exist at all?" Never use AI to automate a step that can be eliminated by a policy change.
2. Require the P&L owner to co-sign the workflow change and the headcount/throughput commitment. Do not fund a production rollout from a central IT budget unless the business-unit head agrees in writing on which legacy step will be switched off and how the saved capacity will show up in their unit's budget within 90 days.
3. Price the human exception loop into the ROI model on Day 1. No probabilistic AI model is 100% accurate. Ask the engineering team: "If the model is right 92% of the time, who checks the other 8%—or do we have to check 100% of outputs to find the 8%?" If humans still have to verify every output before action is taken, your labour savings will be close to zero.

What this looks like in Thailand

Thailand's enterprise AI landscape mirrors the global paradox almost to the decimal point—rapid adoption paired with a wide measurement vacuum:

• According to NECTEC's Thailand AI Readiness Survey (September 2025) covering 574 organisations across the country, 53.7% of surveyed organisations reported already using AI in their operations (a threefold jump from 17.8% in 2024), and 73.3% had allocated dedicated AI budgets.
• Yet in that exact same NECTEC survey, 86.4% of Thai organisations using AI could not clearly state what results or return it had delivered, and only 13.6% reported measuring clear positive outcomes.
• Similarly, PwC Thailand's 28th Annual Global CEO Survey (March 2025) noted that while Thai chief executives are optimistic about AI's potential to improve efficiency over the next three years, trust in embedding AI into core processes and linking it to audited financial metrics remains low.

In many Thai conglomerates and financial institutions, the pattern is visible on the ground: central digital teams roll out enterprise chat assistants and OCR pilots to hit board-mandated "AI adoption" targets, while frontline relationship managers, factory planners and back-office staff keep their existing Excel trackers and line-approval stamps intact because compliance and audit rules were never updated.