Top 3 Things to Know
- Bain's April survey of 951 companies, reported by Bloomberg this month, found roughly 40% achieved less than 10% cost savings from AI against targets of 11-20%. Only about a third hit their targets. Just 7% run fully autonomous agents in production.
- This is the third major research house to land on the same shape of finding after MIT's 95% pilot-failure number and McKinsey's data showing most organizations see no EBIT impact. The technology consistently performs; the operating model consistently does not.
- The companies capturing value share three habits: they redesign the process instead of adding AI to it, they assign an owner to every workflow, and they bank the savings explicitly instead of letting them evaporate into "capacity."
Every few months, one of the big research houses publishes a number that makes the AI budget conversation awkward. Last summer it was MIT: 95% of corporate AI pilots deliver no measurable P&L impact. Then McKinsey: only 39% of organizations report any EBIT effect from AI at all. Now Bain has run the most pointed version of the experiment yet, surveying 951 companies in April about the savings they targeted and the savings they got. Bloomberg summarized the result: companies aimed for 11-20% cost reductions and around 40% of them captured less than 10%. Only about a third hit their targets.
The line Forbes used to summarize Bain's finding deserves to be printed and taped to every AI steering committee's door: the technology worked, but the value did not arrive.
Three research houses, one diagnosis
When MIT published its 95% number, defenders of the status quo could argue methodology. When McKinsey's data landed, they could argue timing: the investments were young. But the findings have now replicated across three independent research operations, different samples, different questions, different years, and the shape is identical. Models perform in evaluation. Pilots impress in demos. And then the value fails to appear on a financial statement.
That replication rules out the usual suspects. The models improved dramatically over the same period the failure numbers stayed flat. Budgets roughly doubled this year, per BCG. Nine in ten CEOs say AI will redefine their industry, so conviction is not the constraint either. What is left is the layer between the technology and the P&L: the operating model.
Where the value actually leaks
Leak one: AI added to a process instead of a process redesigned around AI
The most common pattern in the missing-value cohort: a team gets an AI tool, uses it to do the old process slightly faster, and the process itself, its handoffs, approvals, and waiting time, stays intact. If a contract review involves six handoffs and AI accelerates two of them by 80%, the cycle time barely moves, because the constraint was never those two steps. Savings targets are set against the whole process; tools accelerate fragments of it. The gap between those two numbers is Bain's finding in miniature.
Leak two: nobody owns the workflow
Ask a company who owns their AI strategy and you get a committee. Ask who owns the invoice-processing workflow, its error rate, its cost per invoice, and its improvement roadmap, and you usually get silence. Value capture is a workflow-level activity, and workflows without owners regress. This is the same finding we keep hitting from the adoption side: usage sticks when a named person is responsible for making it stick.
Leak three: savings that evaporate into "capacity"
The subtlest leak. AI genuinely saves four hours per person per week, and then the hours dissolve into slightly longer meetings and slightly slower afternoons. Unless leadership explicitly redeploys the freed capacity, toward more volume, faster cycle times, or actual cost reduction, the savings are real at the task level and fictional at the company level. Bain's respondents did not fail to save time. They failed to collect it.
What the value-capturing third does differently
- They pick the constraint, not the demo. They map the end-to-end process first and aim AI at the step that actually gates throughput, the discipline behind choosing the right workflow.
- They write the value case in the CFO's units. Not "productivity gains" but cost per processed item, cycle time, error rate, with a before-measurement taken seriously enough to survive an audit. Our ROI calculation guide exists because this step is skipped almost everywhere.
- They restructure the work when the numbers prove out. The uncomfortable step. If AI removes 30% of a team's workload, the capturing companies change the team's scope, targets, or size. The non-capturing companies congratulate the team and change nothing.
- They stage autonomy. The 7% running fully autonomous agents did not start autonomous. They ran assisted, then supervised, then autonomous, expanding permissions as error rates earned it. Companies that jump straight to autonomy join Gartner's canceled-project statistics instead.
The bullish read nobody printed
There is a bullish reading of Bain's data that most coverage missed. If roughly a third of companies are hitting double-digit savings targets with the same technology everyone else has, then the differentiator is entirely in execution, and execution is learnable. The gap between the capturing third and everyone else is not a moat of talent or budget. It is a set of operating habits that any disciplined mid-sized company can adopt faster than a Fortune 500 can turn its committee structure.
The technology already works. The value is sitting in your processes, waiting for someone to be made responsible for collecting it.
Is your AI investment leaking value?
Book a free AI Workflow Audit. We will find where your savings are evaporating, put owners and numbers on your highest-value workflows, and build the capture plan your CFO has been asking for.
Book Your Free Audit