AI Is Making Your Inefficiencies Run Faster

Boards across the region are approving AI budgets faster than their finance functions can absorb them. The results are not keeping pace.

Bain & Company's 2026 survey of 951 companies found that nearly 40% of firms that measured AI cost savings landed below 10%, well short of the 11% to 20% many had targeted. The technology worked. The value did not arrive. And 90% of those same companies are increasing their budgets again.

Bain's diagnosis is blunt: the problem is not the technology. It is the operations underneath it. The culprit is workflow debt.

Define the debt

Workflow debt is the accumulated friction inside a process — the redundant approvals, the manual handoffs, the workaround that became permanent, the rule nobody wrote down. Each is small. Together they turn an hour of work into a week.

In payments, the debt is everywhere. Reconciliation across cards, bank transfers, wallets, and a growing list of national QR schemes, each with its own file format and cut-off. Disputes chased through email. Regulatory reporting rebuilt market by market because every regulator wants something slightly different.

Today, people absorb all of it. Staff know who to call when a file does not balance, which step to skip when a system is down, how to read the fuzzy rule.

An AI agent cannot do this. It needs clear rules, stable handoffs, and a named owner for every decision. Drop one into a messy process and it does one of two things: push the work back to a human, saving nothing, or make a confident and expensive mistake, costing trust. AI does not absorb complexity. It scales it.

Avoid the parallel work trap

This is where the money leaks.

A payments business builds a model to forecast transaction volumes and revenue. But the process was never redesigned and the team does not yet trust the output, so the old manual forecast keeps running alongside it — and hours are spent reconciling the two.

Two processes now run where one should. Neither is fully trusted. The technology was deployed, the work was not redesigned, and the promised benefits — faster closes, fewer hours, sharper numbers — evaporate. Cost was added, not removed.

Understand why returns fall short

Bain's data points to three causes.

• The business case assumes full automation; the reality is manual. Only 7% of companies run fully autonomous agents in production. Most still require human approval or intervention. The CFO approved one set of economics. The organisation is living with another.
• Automating a broken process bakes the mess into code. Bain calls this the single most costly mistake in AI deployment, because inefficiency written into software becomes far harder and more expensive to unwind.
• The next wave is funded on optimistic math. 44% of companies plan to pay for their next AI programme from savings generated by the last one — savings that, for many, never fully materialised.

There is a regional wrinkle. Data access and integration is the number one barrier to AI progress, cited by 41% of companies. In markets where money moves across wallets, banks, borders, and currencies, data is fragmented by design. If AI cannot reach clean data, it cannot deliver.

Pay down the debt first

A group of companies is beating the pattern. They did not buy better technology. They fixed the operations first. Three moves:

1. 1Start with a clean sheet. Stop tweaking the current process. Ask instead: if this were designed from scratch today, what would it look like? Settle the destination before selecting the tool.
2. 2Simplify before automating. Strip out low-value steps, duplicate approvals, and just-in-case exceptions. An inconsistent workflow will only have its inconsistency multiplied, at speed.
3. 3Define the hybrid team. Be explicit about three buckets: what only a human does, what a human does with AI assistance, and what runs autonomously. Vague ownership is what breaks agents.

Done well, the returns are real. Amazon's finance technology team used AI to track tax rule changes across markets, cutting a task from 26 minutes per update to 2 — a 92% reduction — with 80% of AI-generated summaries accepted without modification. It worked because the workflow was bounded and the data was already accessible. Not a moonshot; a well-chosen, well-designed task.

The bottom line

AI is not a shortcut past operational discipline. It is a multiplier. Applied to a clean process, it compounds advantage. Applied to a messy one, it compounds the mess — faster, and at greater cost.

Fix the plumbing first.

Five things to remember

• AI does not fix a broken process. It speeds it up and locks it in.
• People are hiding the workflow debt today. Agents will not.
• Running AI alongside the old manual process costs more, not less.
• Scattered data means scattered returns. 41% of firms name data access as their top blocker.
• Redesign first, automate second. Automating a broken process is the most expensive mistake available.