How to Build an AI ROI Business Case That Survives Scrutiny

By the Wynstra team · · 8 min read

Every AI budget conversation in 2026 now comes with a follow-up question: where's the proof? Boards have seen enough pilots and demos. They want numbers that hold up next to any other capital request. The uncomfortable truth is that most organisations still can't produce them. This article sets out a workable method for building an AI ROI business case that survives that scrutiny, using the most recent data on where AI spending actually pays off.

Why so many AI business cases collapse under scrutiny

Start with the sobering context. MIT's NANDA initiative published research in 2025 finding that a new report from MIT has sent shockwaves through the enterprise AI world: 95% of generative AI pilots deliver zero return on investment, based on 300 public deployments, more than 150 executive interviews, and $30 to $40 billion of enterprise spending analysed. Much of that failure wasn't about the models. Researchers pointed to what some commentators call a verification tax: employees end up spending so much time double-checking outputs that the promised efficiencies never materialise. Tellingly, tools bought from specialised vendors fared much better than homegrown builds: AI systems sourced from specialized external vendors show a 67% success rate, more than double the performance of internally built tools.

It's worth treating that 95% figure with some care rather than quoting it as gospel. The study was preliminary, and published as preliminary research, with anonymized organizations and outcomes reported by interviewees, where success meant deployment beyond pilot with measurable KPIs, and ROI was measured six months after the pilot. As one summary put it, that describes one studied group, not a failure rate for AI in general. The lesson isn't that AI doesn't pay off; it's that most businesses aren't measuring or structuring their pilots well enough to find out.

What the 2026 surveys actually tell you to expect

If you're setting expectations for a board or finance committee, recent survey data gives a more grounded picture than vendor marketing does. Across several large 2026 studies, established ROI sits around 7%, and only 37% of companies in McKinsey's 2026 survey see any EBIT impact from AI, and 56% of CEOs told PwC they got neither higher revenue nor lower costs. Payback also takes longer than most technology investments: a typical AI use case takes two to four years to pay back, against 7 to 12 months for other technology, Deloitte found.

Returns are also concentrated, not evenly spread. BCG's 2026 analysis found that the top 20% of companies captured 74% of AI-driven returns. What separates that top group isn't bigger budgets so much as discipline: KPMG's research found organizations where the CEO is accountable for decisions based on AI outputs report established ROI at 14% versus 4%, and those with strong cost visibility 15% versus 3%. In other words, naming an owner and tracking cost properly roughly triples your odds of being able to show a return at all.

A five-step method for a credible business case

1. Baseline before you automate. Measure the current cost, cycle time, error rate and volume of the process you intend to change, for at least four to six weeks. Without this, any "after" number is unfalsifiable.

2. Separate hard savings from soft gains. Headcount hours redeployed, error reduction, and faster cycle times that show up in invoices, claims processed or cases closed are hard. Reported productivity or satisfaction gains are real but soft, and boards increasingly want the former isolated from the latter, since most companies in McKinsey's 2026 survey see any EBIT impact from AI only at a low rate despite much higher self-reported productivity gains.

3. Cost the whole programme, not just the licence. Build in integration, data preparation, change management, ongoing model and output monitoring, and the human review time needed to catch errors — the verification tax described above. Many business cases fail later because this was never in the original model.

4. Model payback with a realistic range and run a sensitivity check. Use the 12–24 month range for narrow, well-scoped workflow automation, and the 2–4 year range Deloitte found typical for broader AI use cases, rather than promising fast wins across the board. Show what happens to payback if adoption is 50% lower than planned, since adoption gaps are a common reason pilots stall.

5. Set a named owner and a review gate. Put one person's name against the target, baseline and budget, with a fixed date — 90 to 180 days after go-live is typical — to review actual numbers against the case and decide whether to scale, fix or stop. This single step is strongly linked to whether organisations can later demonstrate ROI at all.

Picking the first use case to anchor the business case

Not every process is equally good for a first business case. The processes with the clearest, most repeatable evidence share a pattern: a frequent, well-defined workflow with a measurable outcome. In finance and procurement, IBM's Institute for Business Value found a median 8% reduction in total finance cost for robust deployments, rising to 18% when AI is built into end-to-end processes like procure-to-pay and order-to-cash. In customer service, cost reductions are also among the most consistently reported outcomes, with one documented telecom case cutting pre- and post-call administrative work and saving several million dollars a year through automated call summarising and CRM updates. Choosing a first use case with this profile — high volume, clear rules, an existing manual cost you can measure today — makes the business case far easier to defend than a broad, exploratory pilot.

A one-page business case checklist

Before funding any AI automation programme, make sure the case answers each of these on a single page:

  • The specific problem and process, named precisely (not "customer service AI" but "first-response triage for billing queries")
  • The baseline metric, measured before any AI involvement
  • The target metric and the date by which it will be assessed
  • Total cost, including build, integration, licensing, oversight and change management
  • Expected hard savings, kept separate from soft productivity claims
  • A stated payback range, with a sensitivity case for lower adoption
  • A named business owner accountable for the outcome, not just the project manager
  • A fixed review date with a scale, fix or stop decision attached
  • For regulated sectors such as healthcare or finance, a note on the compliance, audit and data-handling implications that affect cost and timeline

Conclusion

The data from 2026 is consistent: AI automation can pay for itself, but most organisations currently can't prove it because they didn't set up the measurement discipline from day one. A credible business case isn't about promising transformation — it's about baselining honestly, costing the whole programme, picking a well-bounded first use case, and reviewing results against a number you agreed on in advance. If you want help building that case or running the pilot that proves it, Wynstra works with teams on exactly this kind of design and measurement before committing to a full rollout.

Sources

  • Ai Roi
  • Business Case
  • Automation Strategy
  • Ai Governance

Have a project in mind?

Tell us what you need. We'll reply within one business day with ideas, a rough timeline and an estimate. No obligation.