EXECUTIVE SUMMARY
From AI adoption to measurable enterprise outcomes.
Many large enterprises in India and the Middle East now use AI. Far fewer can show what it has returned. The AI value gap is the difference between the value an AI initiative was expected to deliver and the value the organization actually captures, net of its fully loaded cost.
Global research points the same way. In McKinsey’s 2025 survey, 88% of respondents said their organizations use AI regularly in at least one function, yet only 39% reported any impact on enterprise-level earnings before interest and taxes (EBIT).[11] The gap is rarely a technology problem. It opens when better work is not turned into a business result, when costs are understated, and when nobody keeps watching after go-live. This paper shows how to measure the gap, close it, and recognize when to stop.
At a glance
- Define value before measuring it: expected, delivered, captured, and net realized value are different numbers, and boards need all four.
- Most of the gap is conversion, not technology: improved work that nobody turns into a business result.
- Count the fully loaded cost: tokens are one line among many. Reviewer time integration, controls, and support often matter more.
- Watch the gap after go-live: leading indicators show it widening long before the annual review.
- Assure the claims: the business owns them, Finance validates them, and Internal Audit provides independent assurance.
WHAT THE AI VALUE GAP IS
An AI value case passes through four stages, and value can leak at each one. Naming the stages lets leaders see exactly where the gap sits.
1
Expected Value
The benefit in the approved value case
Set at funding, with a baseline and an owner.
2
Delivered Value
Verified improvement in the work: time, quality, throughput
Lost value: low adoption, poor quality, or an unchanged process.
3
Captured Value
Improvement converted into an economic or financial result
Latent value: improvement achieved but not yet converted.
4
Net Realized Value
Captured value minus the fully loaded cost of AI
Cost erosion: the full cost of building, running and assuring AI.
Exhibit 1 | The value waterfall. The value gap is the distance between expected value and net realized value.
Not all unrealized value is equal. Latent value is real and still recoverable: hours have been freed, or a process runs faster, but nobody has decided what to do with the gain. Latent value should be tracked and managed until it is captured or consciously written off. Lost value never materialized, usually because users bypassed the tool, quality fell short, or the surrounding process did not change. Lost value calls for a fix or a decision to stop.
| Term | Simple definition | Illustration (see Exhibit 8) |
|---|---|---|
| Expected value | The benefit the value case promised. | Halve the handling time of 6,000 reconciliation exceptions a month. |
| Delivered value | Verified improvement in the work itself. Operational, not yet financial. | 9,600 analyst hours a year freed, confirmed by time study. |
| Captured: economic | A measurable gain that does not change spending, such as capacity redeployed or risk reduced. | One analyst’s time redeployed to control testing. |
| Captured: financial | A gain that shows up in budgets or the ledger. | Three budgeted hires avoided, and a contractor ended. |
| Net realized value | Captured financial value minus the fully loaded cost of AI. | USD 55,000 a year after all AI costs. |
| Latent value | Delivered but not yet captured. Still recoverable. | One analyst’s worth of capacity with no assigned use. |
| Lost value | Expected but never delivered. | Complex exceptions still handled manually because users bypass the tool. |
Exhibit 2 | Value vocabulary. Agreeing these terms with Finance at the start avoids most later disputes about what AI has achieved.
WHY THE GAP OPENS: Six Conditions
Realized value depends on six conditions that reinforce each other.[4] Any one of them, if weak, can become the binding constraint. The leadership task is to find that constraint and fix it before funding the next pilot.
| Condition | The executive question | Typical failure | Typical remedy |
|---|---|---|---|
| Use-case economics | Is there a baselined outcome and a full view of cost? | Impressive demos; unproven return. | A value case with baseline, counterfactual, and loaded cost before funding. |
| Process redesign | Has the end-to-end work changed, not just one task? | Time saved, but approvals and waiting remain. | Redesigned hand-offs and approval chains alongside the AI. |
| Enterprise context | Can AI work reliably with the data and rules it needs? | Low accuracy; heavy checking erodes the gain. | Curated and owned knowledge sources. |
| Enterprise architecture | Do integration and deployment support scale? | Pilots work; roll-outs stall. | Approved integration patterns (see Paper 1). |
| Risk and governance | Are controls proportionate to impact and autonomy? | Risk teams block scale, or exceptions multiply. | Risk-tiered release tracks with clear evidence. |
| Adoption and ownership | Do people use the new process, and is someone accountable for the result? | No owner; staff bypass the system. | A named business owner, redesigned roles, and safe escalation. |
Exhibit 4 | The Uniqus Value Gap Framework: a diagnostic lens, not a calibrated return on investment (ROI) formula.
Why the gap compounds
A reconciliation assistant can be accurate, secure, and widely used, yet deliver little financial benefit if the close calendar, hand-offs, and approval chain stay the same. Activity has sped up; the business result has not.
CONVERTING BETTER WORK INTO BUSINESS VALUE
The critical step is the third one on the bridge below. Without a deliberate capture decision, released capacity is absorbed informally and disappears from view.
1
Capability
Copilot, functional AI or agent
2
Operating Effect
Less effort, rework or delay
3
Value Capture
A named action converts the gain
4
Business Outcome
Margin, cash, revenue, service or risk
Exhibit 5 | The value realization bridge.
For example, 20% less analyst time does not mean 20% lower finance cost. The team might absorb higher volumes, avoid a planned hire, end a contractor, or redeploy experts to controls. Each is a different kind of value, and only some change spending. The capture lever must be chosen, owned, and evidenced.
| Capture lever | Simple definition | What it looks like | Evidence finance should accept |
|---|---|---|---|
| Cost avoided | Spending that would have happened without AI and now will not. Today’s cost base does not fall. | Growth absorbed without hiring; vacancies not backfilled. | An approved hiring or spend plan compared with actual commitments. |
| Cost removed | Spending that existed before and has now stopped. | Contractor, outsourcing or license spend ended. | Contract, purchase-order or ledger evidence. |
| Cash and working capital | Money collected sooner or paid later. Improves cash, not necessarily profit. | Faster collections; early-payment discounts captured. | Cash released, net of any financing cost. |
| Revenue and margin | Extra sales, or better prices and margins, caused by the AI-enabled change. | Higher conversion; less revenue leakage. | A controlled comparison or matched cohort. |
| Risk and compliance | Lower likelihood or impact of losses, penalties or control failures. | Fewer errors, write-offs or incidents. | Observed loss data, or a modeled expected-loss reduction labeled as an estimate. |
Exhibit 6 | Five capture levers. Redeployed capacity is valid economic value, but it is not a saving unless spending or an agreed plan changes. Count each benefit once.
Capture depends on people
Every capture lever changes someone’s work. Professionals need to know who decides, who can override the AI, and who stays accountable. Give them a safe way to escalate and correct results. Where workforce nationalization commitments apply, link redeployment to the entity’s actual commitments and skills plans.
THE FULLY LOADED COST OF AI
Token and license costs are the visible part of AI spending. A credible value case also counts the people, controls, and change effort that AI needs to work, and the costs of things going wrong. Seven cost layers cover the full picture.
| Cost layer | What to include | Often missed |
|---|---|---|
| Build (one-time) | Discovery and design, integration and data engineering, process redesign, test-set creation, security and privacy assessments, vendor implementation fees. | Internal staff time on the project; data clean-up. |
| Run: technology | Licenses and subscriptions; model and application programming interface (API) consumption (tokens, agent runs, embeddings); cloud compute, storage, vector databases and data egress; monitoring and guardrail tools. | Retries, failed runs and long prompts; in-country hosting premiums; AI add-ons bundled into existing contracts. |
| Run: people | Support and agent operations (AgentOps) staff, model and product owners, human review and exception handling. | Reviewer time, which is often the largest hidden cost. |
| Control and assurance | Model validation and re-testing, privacy and legal reviews, regulatory approvals, internal audit coverage. | Re-testing every time a provider updates a model. |
| Change and transition | Training, parallel running, the temporary productivity dip, redeployment and reskilling. | The productivity dip while users adapt. |
| Risk and exit | Expected cost of errors (rework, customer remediation), incident response, vendor switching or exit. | Exit costs for proprietary workflows and data. |
| Shared allocations | A fair share of the AI Center of Excellence (CoE), data platform and shared models. | An allocation method agreed with Finance in advance. |
Exhibit 7 | Seven layers of the fully loaded cost of AI.
How to compute it
Annual loaded cost equals the build cost spread over the solution’s expected life, plus annual run costs for technology and people, control and assurance costs, change costs incurred in the year, the expected cost of errors, and the shared allocation. Divide by the number of successfully completed outcomes to get a unit cost, such as cost per resolved exception. Subtract the total from captured financial value to get net realized value.
Consumption is the most volatile layer. Set budgets and quotas by use case, route work to the cheapest model that passes the task’s evaluation set, cap retries and context length, and alert or throttle when spending passes an approved limit. Tie the spend to unit economics rather than to cost per token.[13] A cheaper model is not cheaper if it doubles retries or human rework.
WORKED EXAMPLE: The USD 300,000 That Was Really USD 55,000
A finance shared-services team clears 6,000 reconciliation exceptions a month. It deploys AI to propose matches and gather the supporting evidence. The plan was to halve handling time; it came down by 40% because complex items are still done by hand. Even so, 9,600 hours a year were freed, about six analysts’ worth of time.
The sponsor’s slide values this at USD 300,000 a year: six analysts at USD 50,000 each. Finance asks a simpler question. What actually changed in the budget?
From headline to net return
USD 300k
The Claim
Six analysts’ time × USD 50,000. Nothing in the budget has changed yet.
USD 210k
Captured
Three budgeted hires avoided (150k), and a contractor ended (60k).
USD 155k
Less Fully Loaded Cost
Build 60k a year over three years, licenses 45k, consumption 12k, support 20k, assurance and CoE 18k.
USD 55k
Net Realized
A 35% return on loaded cost, and one Finance can defend.
Where the six analysts’ time went
3
Financial
Absorbed volume growth, so three budgeted hires were not needed.
1
Financial
Took over a contractor’s work, so the contract was not renewed.
1
Economic
Moved to control testing. Real value, tracked but not booked.
1
Latent
No assigned use yet. The fastest win still on the table.
Exhibit 8 | Illustrative example; all figures are hypothetical. Assumes a loaded labor cost of USD 50,000 per analyst per year.
What the example shows
- The headline overstated the return by more than fivefold: USD 300,000 claimed, USD 55,000 realized. The smaller number is the one that survives an audit.
- The idle analyst is the fastest win: give that capacity a capture action, such as avoiding one more hire, and the net return nearly doubles.
- Labor cost decides the case: at half the labor cost, the same tool loses money. The value case would then have to rest on throughput, quality, or risk.
- Expect a lag: build costs land in year one while benefits ramp up, so this tool pays back only in year three. Show that profile up front so early quarters are not mistaken for failure.
KEEPING THE GAP FROM WIDENING: Value Monitoring
The value gap rarely closes on its own after go-live. More often it widens quietly, as adoption plateaus, models drift, costs creep, and freed capacity goes unused. Each value case therefore needs its own dashboard that tracks expected, delivered, captured, latent, and net realized value, with leading indicators reviewed monthly and outcomes quarterly. International Organization for Standardization/International Electrotechnical Commission (ISO/IEC) 42001 makes monitoring and evaluation of AI performance a formal requirement of an AI management system.[18]
| Early warning | What it tells you | Illustrative trigger | Decision |
|---|---|---|---|
| Adoption stalls | Value is being lost: users are not taking up the new way of working. | Active use below 60% of eligible users after 90 days. | Fix the experience or training, or narrow the scope. |
| Overrides or rework rise | Quality is decaying, from model drift or poor context. | Override rate up by more than a third above its baseline. | Re-evaluate the model and context; pause expansion. |
| Freed capacity unassigned | Latent value is building up. | No capture action 90 days after go-live. | Assign a capture action, or reduce licenses and scope. |
| Unit cost rises | Cost is eroding the return. | Cost per successful outcome 20% above plan. | Route to cheaper models, cap usage, or renegotiate. |
| Benefit behind plan | The value case is not holding. | Run-rate below 70% of plan for two quarters. | Redesign the solution or stop it. |
| Volumes or mix shift | The baseline no longer holds, so attribution is at risk. | Volumes move by more than 15%. | Rebase with Finance before reporting benefit. |
Exhibit 9 | Value-gap early warnings. Triggers are illustrative; set them per use case. Committee of Sponsoring Organizations of the Treadway Commission (COSO) 2026 guidance on generative AI similarly points to monitoring activities, such as override rates, to detect drift early.[14]
Measure each category of AI on its own terms. Holding a copilot to enterprise-level profit tests is as misleading as judging an enterprise agent on usage alone.
| AI category | Leading indicators | Value indicators |
|---|---|---|
| Productivity AI | Active use, task penetration and sampled output quality. | Time saved in sampled task studies, and what the freed time is used for. |
| Functional AI | Straight-through rate, overrides and exception aging. | Cycle time, error rate, and cost to serve against baseline. |
| Enterprise AI | Agent completion and escalation rates; cost per run. | Cash conversion, end-to-end cost, revenue capture, and risk losses. |
Exhibit 10 | Indicators by AI category, using the three AI categories from Paper 1.
Agree stop criteria before funding
The hardest decision is to stop. It is far easier if the conditions for stopping were written into the value case at the start, such as a minimum adoption level, a maximum unit cost, or a benefit run-rate by a set date, and if the early warnings above are reviewed against them.
REGIONAL CONTEXT: India, UAE, KSA and Qatar
The same evidence standard applies everywhere, but what drives the value case differs by market. Each market is set out on its own terms below. The observations are starting points for diligence on a specific entity, not assumptions to apply across the board.
India
In our view, hours saved alone rarely pay back where labor costs are modest relative to AI services priced in US dollars. Build the case on throughput, quality and growth.
What shapes the value case
- Role-level labor costs that vary widely by city and function.
- A large base of shared services and global capability centers, where AI changes the delivery model itself.
- AI, cloud, and model costs are largely priced in US dollars.
- Reserve Bank of India (RBI) and Securities and Exchange Board of India (SEBI) expectations for regulated entities, and Digital Personal Data Protection (DPDP) obligations phasing in through May 2027.[5][6]
Most relevant capture levers
- Growth absorbed without hiring.
- Throughput and service-level gains in shared services.
- Error and rework reduction.
- For service providers, new revenue or better delivery margins.
Watch-outs
- Claiming hours as savings without a capture action.
- Clients asking to share productivity gains through lower prices.
- Currency exposure on consumption costs.
United Arab Emirates
In our view, adoption momentum is strong. The value case has to show conversion, not activity.
What shapes the value case
- Government digital programs, such as Abu Dhabi’s 2025–2027 strategy, and service-level targets.[7]
- Emiratization commitments where they apply.
- Central Bank of the UAE (CBUAE) guidance on AI for licensed financial institutions (February 2026), which adds control effort.[8]
- Separate onshore, Dubai International Financial Centre (DIFC) and Abu Dhabi Global Market (ADGM) regimes that can affect hosting choices and cost.
Most relevant capture levers
- Processing time and service quality.
- Growth absorbed without hiring.
- Redeployment aligned to Emiratization plans.
- Risk reduction in financial services.
Watch-outs
- Counting reductions in roles held by nationals as savings.
- Leaving control, compliance and hosting costs out of the loaded cost.
Kingdom of Saudi Arabia
In our view, growth is the main capture lever. Build workforce nationalization into the value case from the start.
What shapes the value case
- Vision 2030 programs and project growth that create demand.
- Saudization quotas that are specific to sector and role.
- Expanding in-country AI capacity, including Public Investment Fund (PIF)-backed HUMAIN.[9]
- Personal Data Protection Law (PDPL) transfer rules that shape hosting options.
Most relevant capture levers
- Growth absorbed without hiring.
- Cash and working capital in project-heavy operations.
- Throughput and cycle time.
- Capability building that supports nationalization goals.
Watch-outs
- Headcount-reduction cases that conflict with Saudization commitments.
- Unbudgeted in-country hosting premiums.
- Assuming every workload must be hosted onshore.
Qatar
In our view, expect fewer, larger value cases. Regulatory approval timing shapes when benefits start.
What shapes the value case
- Concentrated activity in government, energy and financial services.
- Qatarization commitments where they apply.
- Qatar Central Bank (QCB) approval requirements for high-risk AI at licensed entities.[10]
Most relevant capture levers
- Service quality and processing time.
- Risk and compliance improvements.
- Cost removed from external and outsourced spend.
Watch-outs
- Benefit timelines that ignore approval lead times.
- Fixed costs spread over modest volumes, which raise unit cost.
Exhibit 11 | Regional value-capture factors. Workforce and labor-cost observations are qualitative Uniqus perspectives and must be quantified for each entity. Regulatory references are summarized as of October 2026.
GOVERNANCE AND BENEFIT ASSURANCE
AI value claims should be governed like any other material management assertion: owned by the business, validated by a second line and independently assured by a third. Should Internal Audit do this work? In our view, Internal Audit should not calculate or own the benefit, because that would compromise its independence. Its role is to give the board independent assurance that the benefit-realization process is sound and that material claims hold up. The Institute of Internal Auditors’ (IIA) Three Lines Model, which replaced the earlier “three lines of defense,” sets out this division of roles.[15]
| Role | Responsibility for AI value | Typical activities | Reference point |
|---|---|---|---|
| Board or Audit Committee | Oversight of material AI investments and value claims. | Reviews portfolio value and assurance results at least annually. | IIA Three Lines Model [15] |
| AI Council | Portfolio decisions. | Prioritizes investment; reviews benefit capture at 90 and 180 days; decides to scale, redesign, or stop. | ISO/IEC 42001 management review [18]; Control Objectives for Information and Related Technologies (COBIT) benefits delivery [19] |
| First line: business owner | Owns the outcome and the capture action. | Sets the baseline, executes capture actions, reports adoption and benefit. | COSO control activities [14] |
| Second line: Finance | Validates benefit claims and costs. | Checks benefit type, counterfactual, loaded cost, and ledger evidence; signs off net realized value. | COSO monitoring activities [14]; COBIT [19] |
| Second line: Risk and Compliance | Validates control and risk claims. | Reviews risk-reduction benefits, regulatory approvals, and control evidence. | COSO generative AI guidance [14] |
| Third line: Internal Audit | Independent assurance. | Audits the value-realization process; tests a sample of material claims and the completeness of AI costs and the AI spend inventory; advises on control design without owning it. | IIA Global Internal Audit Standards [16]; IIA AI Auditing Framework [17]; ISO/IEC 42001 internal audit [18] |
Exhibit 12 | Assurance over AI value across the three lines. Where an organization publishes AI value figures, external assurance over those figures is worth considering.
Assurance works best when it is built into the lifecycle rather than added at the end:
- Before funding: agree the outcome, baseline, counterfactual, owner, capture lever, timing, loaded cost, and stop criteria.
- Before production: test on representative work, document the redesigned workflow and confirm the capture action.
- At 90 and 180 days: the business owner reports adoption and capture; Finance verifies delivered, captured, and net realized value; the CoE reports quality and unit cost.
- Annually: Internal Audit includes AI value realization in its risk-based plan once AI spend is material, and reports to the Audit Committee.
ROADMAP: Closing the Gap in 90 Days and Beyond
The roadmap starts with visibility of spend and ends with independent assurance. Each phase ends with a decision, not just a deliverable.
1
Days 0–30
Diagnose
Key actions Build the AI spend inventory; list live use cases and pilots; baseline three to five priority value cases; find the binding constraint among the six conditions.
Outputs AI spend baseline and use-case register.
Lead Finance with the AI CoE
2
Days 31–60
Design and commit
Key actions Name business owners; agree capture levers, counterfactuals and the loaded cost model; set dashboards, early warnings and stop criteria; charter the AI Council.
Outputs Approved value cases with stop criteria.
Lead Business owners, Finance and the AI Council.
3
Days 61–90
Prove and decide
Key actions Run redesigned workflows on real volumes; verify delivered value and loaded cost; decide to scale, redesign, or stop.
Outputs Evidence packs and portfolio decisions.
Lead AI Council.
4
Beyond Day 90
Monitor and assure
Key actions Run 90- and 180-day benefit reviews and monthly gap monitoring; Internal Audit reviews the value-realization process within the first year.
Outputs Net realized value reported to the board.
Lead Finance, the AI Council and Internal Audit.
Exhibit 13 | Value roadmap. Timings are indicative and depend on the size of the existing AI estate.
How to build the AI spend inventory
Few organizations know what they spend on AI, because the costs are scattered across budgets, contracts, and cloud accounts. COSO’s generative AI guidance treats an inventory as the second step after governance.[14] Seven sources usually cover it:
| Source | What it reveals | Watch for |
|---|---|---|
| Procurement and accounts payable | Contracts and invoices with AI vendors, model providers and implementation partners. | AI add-ons bundled into wider software contracts. |
| Cloud billing | Model APIs, graphics processing unit (GPU) compute, vector databases, storage and egress, by account and tag. | Untagged resources and shared accounts. |
| License management and single sign-on | AI tool seats and actual use per seat. | Seats that are paid for but unused. |
| Expense claims and corporate cards | Individually purchased AI subscriptions. | Shadow AI outside approved tools. |
| Network and security logs | Use of unsanctioned AI services. | Enterprise data leaving through unapproved tools. |
| Timesheets and project codes | Internal effort on building, running, and reviewing AI. | Reviewer time not coded to AI. |
| Model and API gateway telemetry | Consumption by application and use case. | Consumption that cannot be traced to a use case. |
Exhibit 14 | Sources for the AI spend inventory.
Tag each cost line with its use case, business owner, AI category, cost layer from Exhibit 7, and whether it is one-time or recurring, and whether it is committed or consumed. Link every line to a value case. Spend that cannot be linked to one is the first place to look for savings.
The Uniqus Point of View
In our view, the enterprises that lead the next phase of AI will not be those that deploy the most assistants or agents. They will be those that repeatedly turn better work into measurable business results, count the full cost of doing so, and can show an independent reviewer where the value was captured.
APPENDIX: The Minimum Evidence Pack for an AI Value Case
| Check | Evidence expected |
|---|---|
| Business owner | A named business executive, plus a Finance sponsor responsible for verification. |
| Baseline and counterfactual | Before-state volumes, unit costs, error rates, cycle time and service levels, and how the outcome would plausibly have changed without AI. |
| Value waterfall | Expected, delivered, captured, latent and net realized value, each with its evidence. |
| Capture lever | The action that converts the gain, its owner and date, with no double counting between levers. |
| Loaded cost | All seven cost layers, split into one-time and recurring, with the shared-cost allocation method. |
| Timing | The expected benefit profile, including ramp-up, payback and possible decay. |
| Monitoring and stop criteria | Dashboard indicators, early-warning triggers and the conditions for stopping. |
| Validation and assurance | Finance sign-off at 90 and 180 days; inclusion in Internal Audit’s plan once material. |
Sources and qualification notes
- Uniqus Value Gap Framework: a conceptual management framework developed by Uniqus, not an empirically calibrated ROI model
- Reserve Bank of India, Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) Committee Report (2025) and consultation on model risk management (2026); recommendations are distinct from binding requirements
- SEBI (Intermediaries) (Amendment) Regulations, 2025
- Abu Dhabi Department of Government Enablement, Government Digital Strategy 2025–2027
- CBUAE Rulebook, Guidance Note on consumer protection and responsible use of AI and machine learning by licensed financial institutions, February 2026
- Public Investment Fund, launch of HUMAIN, May 12, 2025
- Qatar Central Bank, Artificial Intelligence Guideline (2024) (see Section 11 of the Guideline)
- McKinsey, The state of AI in 2025: Agents, innovation, and transformation, November 5, 2025 (global survey; self-reported)
- FinOps Foundation, Unit Economics capability
- COSO, Achieving Effective Internal Control Over Generative AI, released February 23, 2026; cited here through The IIA’s Internal Auditor as a secondary summary. Confirm against COSO’s full publication before relying on specific capability descriptions
- The IIA, Three Lines Model (2020, with a refreshed Statement of Position in 2026)
- The IIA, Global Internal Audit Standards, effective January 9, 2025
- The IIA, Artificial Intelligence Auditing Framework
- ISO/IEC 42001:2023, Artificial intelligence management system, clauses 9.1 (monitoring and evaluation), 9.2 (internal audit) and 9.3 (management review)
- ISACA, COBIT 2019, governance objective EDM02 (Ensured Benefits Delivery)
Published case studies are company or vendor statements, not independently assured benchmarks. The worked example is hypothetical. Regional workforce and economic observations must be verified for each entity, and regulatory references are a summary as of October 2026, not legal advice.
