From wildfire to workflow: Turning AI experimentation into measurable value

It all starts with a memo. 

Leadership announces an AI strategy at the next town hall while rolling out ChatGPT Enterprise licences. Within weeks, managers in finance, marketing, and operations begin building their own automations. A junior analyst chains prompts to draft variance commentary. A logistics manager wires a custom GPT to a shared inbox to triage supplier queries. 

By the time the steering committee convenes to discuss governance, the organisation already has dozens, sometimes hundreds, of unsanctioned custom GPTs and Projects running. This is the wildfire stage of AI adoption, and most mid-market organisations are now in it. 

However, last year, it became clear that around 95% of generative AI pilots had no measurable impact on the income statement. This aligns with what our previous HLB data series concluded: 70-80% of AI projects fail because the data underlying them is not fit for that purpose. 

Why most AI activity still isn't showing up in the P&L

There is a meaningful gap between productivity and value. A Copilot licence that saves an analyst three hours a week is not the same thing as a process redesign that removes those three hours from the overall cost.

P&L measures something different. It measures revenue uplift, cost-out, working capital release, or a reduction in error rates that an auditor can verify. On the other hand, time saved using AI often dissipates into other low-priority work, without leaving a dent in the income statement.

MIT research points to a misallocation that compounds the problem: more than half of generative AI budgets sit in sales and marketing, where returns are visible but shallow. But back-office automation, which can offer the most ROI, remains underfunded.

The deeper issue, however, is ownership. Most AI activity in mid-market businesses today lacks an executive sponsor with formal accountability for outcomes. 

The CFO as AI owner: a strategic-plan framework

Basically, there is an impressive fire burning within every mid-market organisation. But there is no one responsible for what burns or what grows.

Advisors are increasingly pointing to the CFO for ownership. Few other roles connect data, performance, and governance under the purview of a single office. Plus, CFOs are already paid to evaluate ROI, model risk, and phase investment — the exact disciplines an AI portfolio needs.

Colin Nierenberg's team at GHJ (HLB USA) has developed a CFO-led framework for mid-market finance teams to build AI capabilities. The framework opens with three honest questions to ask before any spending is approved: 

• What opportunity are we solving for? 

• Is it worth solving? 

• How can AI help us get there? 

These obvious questions open up three aspects; laying the foundation, scaling predictive intelligence, and enabling autonomous finance.

Three lenses for prioritising AI use cases

A useful triage is to sort candidate use cases into three buckets: 

• Automate high-volume, rules-based work like bank reconciliations, invoice coding, and regulatory returns. 

• Augment work where AI can help make decisions faster, like variance commentary, contract review, and FP&A scenario building. 

• Reinvent workflows that can be made more efficient end-to-end, often the largest prize and the longest project. 

Most people start with reinvention. This is the trap. Reinventing the wheel takes time and leads to more failures. The discipline is to start where automation can be proven quickly, building the credibility and the data foundations  which can then lead to bigger changes.

Case study: 60 hours to one minute

A long-tenured GHJ (HLB USA) client filed monthly reports to its state Department of Mental Health. This work fell to a single employee who, over 45 years, took around 60 hours to compile the return. 

When she retired, GHJ (HLB USA) rebuilt the workflow using a combination of automation tooling and structured prompts to bring cycle time down from 60 hours to roughly one minute. The cost savings are obvious, but the real win is knowing that 45 years of institutional knowledge required to compile the return now lives in a documented, auditable process. 

As Colin Nierenberg observes, the question every CFO should be asking is "Where is undocumented expertise sitting one resignation away from becoming a problem?", rather than "Where is AI saving us money?". 

Case study: a 73-entity consolidation, built once and run forever

A more ambitious GHJ (HLB USA) engagement shows what enterprise-grade work looks like once foundations are in place. 

The client, an international technology group comprising 73 entities, was undertaking its first consolidated audit. To meet new reporting requirements, GHJ (HLB USA) developed an Alteryx-powered consolidation process that aggregated reporting packages from multiple jurisdictions, calculated foreign currency adjustments, eliminated intercompany balances and profit, and produced consolidated financial statements and disclosures.

Beyond consolidating 73 entities in a single run, the solution transformed the audit process itself. Automated validation checks enabled data issues and intercompany mismatches to be identified earlier, allowing teams to investigate and resolve discrepancies before they became larger reporting challenges. The result was a more accurate and efficient audit process, greater visibility across the group, and a repeatable framework built on consistent, governed financial data.

These kinds of results only become possible once an organisation has done the unglamorous data work the 2025 HLB series argued for: clean masters, governed flows, single sources of truth. 

AI sits on top of that foundation; it does not replace it.

Quick wins your team can build this quarter

Most readers will not be commissioning an engine that can consolidate a 73-entity financial report this quarter. The good news, observes Ondřej Hlaváček from HLB PROXY (HLB Czech Republic), is that for many practical use cases the gap between professional and amateur AI builders has narrowed sharply. 

A finance lead with no coding background can now build a custom GPT to draft variance commentary in the firm's house style, summarise board packs, or pre-screen supplier contracts against a checklist. Meanwhile, AI capabilities embedded in Excel and Google Sheets have improved to the point where competent users can run analyses that two years ago needed a data-science team.

This democratisation is the most underrated trend of 2026. Mid-market businesses no longer have to choose between expensive bespoke builds and shelfware enterprise tools. They can ship governed, useful automations in days.

Conclusion: From wildfire to workflow

The wildfire of bottom-up AI experimentation is an asset, not a problem. It signals curiosity, capability, and intent — three things money cannot buy. 

The job of leadership is to channel it; to give it an owner (increasingly, the CFO), a framework (a small number of well-prioritised use cases across automate, augment, and reinvent), and a measurement discipline that distinguishes time saved from value created. That is how experimentation becomes a portfolio, and a portfolio becomes a P&L line.

In Part 2, we will turn to the question that follows naturally from ownership: trust. As AI moves from individual productivity tool to embedded business process, and into your supply chain, the audit conversation is changing.

Ready to move past experimentation?

If your business has moved past AI experimentation but is struggling to demonstrate value, HLB's Advisory andData Analytics & Business Intelligence teams can help you build a CFO-led AI strategy and a portfolio of high-return use cases.