Grund InstituteAI has moved from experiment to everyday operations. Organizations run enterprise platforms, build AI into workflows, deploy agents and train their people. The technology does useful work: some tasks get faster, some activities are automated, and some capabilities that were too expensive or difficult before are now within reach.
Showing that AI works is not the same as showing the investment pays off. A company reports thousands of active users without knowing how much their work improved. A department estimates hundreds of hours saved without knowing whether those hours became capacity, lower cost or better outcomes. An agent completes a high volume of tasks while adding supervision, maintenance and exception-handling cost. And sometimes genuine improvements never appear in the financial reporting at all.
The problem is not always that AI is failing. Sometimes value exists and is poorly measured. Sometimes benefits are real but stay inside individual tasks. Sometimes expected gains disappear into rework, workflow friction, weak implementation or costs left out of the original business case.
This is the final briefing in the series. It draws together the threads of the previous eleven, because nearly every pattern they describe ends up here, as value that cannot be proven.
AI activity
Is AI being used?
Licences, active users, interactions, automated tasks, self-reported time saved.
Evidence of implementation
Operational improvement
Is the work better?
Less total effort, shorter cycle times, fewer errors, more throughput, across the whole workflow.
Evidence of effect
Financial return
Is the business better off?
Lower cost, more revenue, redeployed capacity, reduced risk, net of the full cost of AI and attributable to it.
Evidence of value
The core challenge is telling AI activity, operational improvement and financial return apart.
§The value gap is real, and it is changing
Research shows AI can produce measurable gains under real working conditions.
Brynjolfsson, Li & Raymond · Quarterly Journal of Economics, 2025
Results from one task in one setting. They show that AI can create measurable value, not that every implementation will.
Boston Consulting Group · 2025 survey of 1,250 senior executives and AI decision makers
BCG's September 2026 update reported that nearly half of companies now see significant value from AI: real progress, with wide differences in outcome remaining.
None of this says AI adoption is failing across the board. It says organizations are getting better at extracting value, and that outcomes still vary widely between companies using similar technology.
The question is no longer whether AI can create value. It is why some organizations can prove it and others cannot.
§Five reasons ROI is hard to prove
Activity is measured instead of outcomes
Dashboards report licences, active users, training completion, automated tasks and usage. Useful for tracking implementation; silent on business performance. Usage can rise with no change in profit, customer satisfaction, capacity or quality. Adoption is an input to value, not proof of it.
- See also
- Briefing 06, The AI Adoption Trap
Time saved is mistaken for money saved
Reported time savings may be genuine, but they do not reduce cost by themselves. If someone saves five hours a week while salary and total output stay the same, the organization has not banked a saving. The time may become real capacity, higher-value work, better quality, less overtime or more customers served. Each of those needs evidence.
- See also
- Briefing 02, The AI Productivity Illusion
Becomes value
If it can be evidenced
- More output
- Higher-value work
- Better quality
- Less overtime
- Hiring avoided
Counts toward ROI
Disappears
If no one tracks it
- Absorbed by rework
- Spent waiting on approvals
- Lost to coordination
- Unallocated slack
Reported, not realized
Benefits leak across the workflow
AI speeds up one step and adds work elsewhere. Generated reports need review. Recommendations need validation. Agents need supervision. Faster output builds approval queues. People correct errors and reconcile information across systems. Measuring only the step where AI enters overstates the net gain; the financial benefit depends on the whole workflow.
- See also
- Briefings 01, 03, 05 and 11
The true cost of AI is underestimated
AI costs more than its subscription fees. Some costs are direct spend; others are internal labor or the opportunity cost of people and budget diverted from other work. When a business case counts the licence and leaves out the rest, its ROI is overstated before it starts.
| Cost layer | What it includes |
|---|---|
| Direct spend | Licences, infrastructure, integrations |
| Implementation | Security review, governance, configuration, training, change management |
| Ongoing operation | Maintenance, monitoring, verification, correction, exception handling |
| Opportunity cost | Internal time and budget diverted from other priorities |
Effects are not attributed
Performance improves after AI arrives and AI gets the credit. Process changes, staffing, demand, learning effects and market conditions move performance too. Without a baseline and a credible way to separate AI's contribution, value is overstated. The reverse also happens: real AI benefits go unrecognized because the chosen measures do not capture them.
§From adoption statistics to defensible value
A defensible AI ROI does not need perfect precision. It needs a clear chain of evidence from investment to outcome. We recommend building it in six steps.
- Define the outcome before the investmentName the business result the AI is meant to improve: cost, cycle time, quality, capacity, revenue, risk.
- Record a baselineMeasure the work before AI: total human effort, cycle time, error rate, throughput, the outcome itself. Without a baseline, any claim of improvement is untestable.
- Count the full costDirect spend, implementation, ongoing operation and opportunity cost, including the human effort of verification and supervision.
- Measure the whole workflowNet effect after downstream review, correction, coordination and approval, not the gain at the step where AI enters.
- Convert with stated assumptionsTranslate operational change into financial terms openly: how freed capacity is used, what it is worth, and what was assumed.
- Attribute honestlyUse pilots, comparison groups or historical baselines to separate AI's effect from other changes. Report a range where certainty is limited.
Each term in that formula is where business cases usually go wrong: value that is claimed but not attributable, and cost that is real but not counted.
§Two versions of the same business case
Take a hypothetical finance department that deploys AI to prepare monthly management reporting. Six months later it presents the result.
| Element | As presented | As evidenced |
|---|---|---|
| Value claimed | Hours saved on report drafting, multiplied by hourly cost | Drafting time fell; review and reconciliation time rose |
| Capacity | Assumed to be a saving | No change in headcount, overtime or output; time not redeployed |
| Cost | Licence fees | Licences, integration, data preparation, controller review time |
| Attribution | Faster close credited to AI | A process change made in the same quarter also shortened the close |
| Conclusion | Strong ROI | Real but smaller gain, not yet realized financially |
Illustrative example, not an observed result.
The AI is useful. The first business case is not defensible; the second is, and it points to what to change: cut the reconciliation work at source and decide deliberately how the freed time will be used.
§The GrundMind view
Every briefing in this series describes a way AI value can be lost or hidden: workslop that moves effort downstream, time savings that never become capacity, verification that eats the gain, perception gaps that hide friction, broken workflows that absorb speed, adoption metrics that stand in for results, poor cognitive fit, miscalibrated trust, eroded expertise, ungoverned tools and unsupervised agents. The missing ROI is where they all show up on the balance sheet.
Three diagnostics carry this briefing. AI ROI investigates why an organization cannot show a measurable return and builds the evidence chain from investment to outcome. AI Value Leakage locates where expected benefits are lost in the workflow. AI Net Efficiency establishes whether AI reduces total human effort once verification, correction and supervision are counted. Together they separate value that exists but is unmeasured, value that is real but stuck at task level, and value that was never there.
The result is not a single ROI figure for "AI". It is an evidence-based view of which use cases pay, which leak, which need redesign and which should stop.
§Five questions for the leadership team
- What outcome was each AI investment meant to improve, and has it?If the outcome was never defined, the return cannot be shown.
- Do we have baselines for the work AI now supports?Without them, improvement claims cannot be tested.
- Does our business case include the full cost?Implementation, operation, verification and supervision as well as licences.
- Where has saved time actually gone?Capacity that is not redeployed is not a financial return.
- Can we separate AI's effect from everything else that changed?Attribution is what turns an observed improvement into a defensible one.
AI value has to be evidenced, not assumed.
AI can create real, measurable value. Many organizations cannot prove it, because they measure activity instead of outcomes, count hours as money, look at one step instead of the workflow, leave costs out of the case and credit AI for changes it did not cause.
The organizations that can show a return define outcomes first, record baselines, count the full cost, measure across the workflow and attribute honestly. They invest further where the evidence supports it, and stop where it does not.
The question is not whether AI is being used. It is whether the organization can show what that use is worth.
Book an AI adoption diagnostic pilot.
Find out which of your AI investments are creating value, where that value is leaking, and how to build a business case that holds up.
About this briefing
The GrundMind Adoption Briefings are short, leadership-facing notes from the Grund Institute on the measurement and design of AI adoption in organizations. The series accompanies the longer-form Grund Institute white papers on cognitive fit, the Second Mind framework, and the GrundMind diagnostic.
Sources
- Brynjolfsson, E., Li, D. & Raymond, L., Generative AI at Work, Quarterly Journal of Economics, 2025.
- Boston Consulting Group, 2025 survey of 1,250 senior executives and AI decision makers.
- Boston Consulting Group, updated research on AI value, September 2026.
- Grund Institute, Bridging the Cognitive Gap in AI Adoption, 2026.