AI Value Attribution Framework

The AI Value Attribution Framework is Crescent Capital Advisors' five-level model for verifying what AI investment produced: Cost, Adoption, Productivity, Business Outcomes, Enterprise Value. Most portfolio companies can evidence the first two levels and report as though they had evidenced the last two. The framework identifies that gap and sets out four questions to answer before the next initiative is funded, so the same gap does not appear on it.

What a Dashboard Leaves Out

A dashboard answers the questions someone thought to instrument. Most AI dashboards were specified during the pilot phase, when the interesting questions were about usage: who logged in, how many prompts ran, which agents fired. Those panels keep reporting the same things after the budget discussion has turned to what the spending produced.

The tooling is not the cause. A company can have every number its dashboard produces and still connect none of them to a line in the P&L. That happens when the baseline was never captured, when no one owned the result, or when the outcome was defined after the fact to fit whatever changed.

The framework therefore sets out five levels in a fixed order. Each level has its own question, its own metrics, and a plain statement of where it stops. A leadership team can work through the levels in a morning and find the highest one its evidence supports. That level is the claim it can defend in a board meeting. Anything claimed above it is unsupported narrative.

The Five Levels of AI Measurement

Each level answers a different question and satisfies a different audience. The status column is the pattern CCA sees across mid-market portfolio companies: how many can produce evidence at that level today without building new measurement.

Level	Name	Executive question	Status across most companies
1	Cost	How much are we spending?	Most can measure
2	Adoption	Are people using it?	Most can measure
3	Productivity	Are teams working faster?	Some can measure
4	Business Outcomes	Did business results improve?	Very few can measure
5	Enterprise Value	Did the economics of the business change?	Almost none can measure consistently

Levels 1 and 2 describe activity. Levels 4 and 5 describe economics. Level 3 sits between them, and most AI value stories stop there: velocity improved, everyone agrees it improved, and no one translated the improvement into a financial statement.

Compare the status column with the executive question in each row. Boards, lenders, and buyers ask Level 4 and Level 5 questions. The measurement most companies built answers Levels 1 and 2. The attribution gap is the difference between what is asked and what is measured.

Level 1: Cost

Question: How much are we spending?

Typical metrics: AI software licenses, API spend, model costs, infrastructure, token consumption.

What it tells you: whether spending is rising, and how fast.

Where it stops: cost says nothing about value, and it is routinely understated. See the AI Cost Optimization crosswalk below for why the stated number is usually the wrong denominator.

Level 2: Adoption

Question: Are people using it?

Typical metrics: active users, prompt volume, agent executions, session frequency, seats assigned versus seats used.

What it tells you: whether the deployment reached the people it was bought for.

Where it stops: usage counts the input, not the result. A team can run 40,000 prompts a quarter and close the same number of tickets it closed before.

Level 3: Productivity

Question: Are teams working faster?

Typical metrics: development cycle time, support handling time, content production velocity, time to completion, rework rate.

What it tells you: whether AI changed how the work gets done. This is the first level with a real signal in it, and most programs make their claims at this level.

Where it stops: hours saved become money only when the hours get redeployed or removed. A 30% handling-time improvement with flat headcount and flat volume is a claim about capacity. Capacity has economic value only when someone decides what to do with it.

Level 4: Business Outcomes

Question: Did business results improve?

Typical metrics: customer onboarding time, support resolution rates, renewal and gross retention, customer satisfaction, sales productivity, lead conversion.

What it tells you: whether AI influenced the operating results a CEO already reports monthly.

Where it stops: attribution gets hard here, because other things in the business changed during the same period. Pricing changed, a competitor lost share, the team hired. A Level 4 number is credible only when a baseline was captured before the deployment and the confounders are named.

Level 5: Enterprise Value

Question: Did the economics of the business change?

Typical metrics: EBITDA impact, revenue growth, gross margin, net revenue retention, operating leverage, and the AI narrative a buyer will price.

What it tells you: whether AI changed the figures the sponsor is underwriting.

Where it stops: Level 5 inherits every weakness below it. A Level 5 claim built on an unverified Level 4 outcome and an understated Level 1 cost is a calculation based on two unchecked estimates.

Why the Gap Exists

Most AI programs start as experiments, and experiments are measured on learning and usage. That's the correct instrumentation for the first two quarters. Then spending grows, the CFO starts asking about run rate, and the question changes from what AI can do to what AI did.

The measurement system that would answer the second question had to be built before the deployment, because it needs a baseline. By the time anyone asks, the pre-AI number is gone.

Three Failure Modes

Usage reported as value. Leadership reports user counts, prompt volume, and agent executions. Nobody can state the business impact. Activity has been accepted as a proxy for value, and that works until a board member asks what the activity produced.

Productivity gains with no financial link. Teams did get faster. Cycle time fell, the engineers notice the difference, and the survey scores agree. No one connected the improvement to a financial line, so the gain appears in a presentation and not in the P&L. When the budget is built, the finance team discounts it to zero.

Costs tracked, benefits not tracked. Finance can account for every dollar spent on AI. Nobody can account for a dollar earned, saved, or protected. This failure mode is the one that gets programs cancelled. A cost with no matching benefit is the easiest line to cut in a tight quarter, and the cut removes the initiatives that were working along with the ones that were not.

The Four Pre-Approval Questions

Every AI initiative answers these before it gets funded. Answering the four questions takes one meeting. If they are skipped, the baseline is never captured, and a year of evidence cannot be rebuilt later.

1. What business outcome are we trying to improve? Named at Level 4 or Level 5, in the language the company already uses on its monthly reporting.
2. How will we measure it? The specific metric, the system it comes from, and the reporting cadence.
3. What baseline are we comparing against? Captured before deployment, with the known confounders written down while everyone still remembers them.
4. Who owns the result? One accountable executive, named in the approval. A vendor cannot own this one. An AI working group would own it collectively, and in practice that means nobody does.

The fourth question fails most often, and the failure is rarely noticed at the time. When the metric does not move, an initiative with a shared owner has nobody accountable for explaining why.

Self-Assessment Rubric

Attribution maturity is cumulative. Your defensible level is the highest level where you have evidence and every level beneath it also has evidence. A gap at any lower level limits the claim above it. That is why most companies that believe they are operating at Level 4 can defend only Level 2.

Level	You can claim it when	What caps the claim
1	Total AI cost includes labor, waste, and risk, and rolls up to a named feature owner	An invoice-only figure that counts vendor spend and stops there
2	Usage split by team and workflow, with seats assigned measured against seats used	A license count reported as an adoption number
3	Cycle time or handling time measured against a pre-deployment baseline from the same system	A survey asking people how much time they think they saved
4	An operating metric moved, the baseline predates the deployment, and the confounders in the period are documented	A result claimed in a quarter when pricing, staffing, or product also changed
5	The Level 4 outcome is priced into EBITDA or margin, against a Level 1 cost that includes labor and waste	An EBITDA number derived from hours saved that were never redeployed or removed

Apply the levels to each initiative rather than to the company as a whole. A single portfolio company will have a Level 4 claim on one workflow and a Level 2 claim on five others. The aggregate number it reports to the board averages them into a figure no one can defend.

An Illustrative Attribution Walk

The figures below are modeled to show how the levels interact. They aren't results from a named CCA engagement.

A portfolio company deploys AI across customer support. Finance reports AI spend of $270K a year, covering licenses, API, and infrastructure. Support handling time falls from 14 minutes to 9.8 minutes across 22,000 tickets a quarter. The team reports a 30% productivity gain and, separately, credits AI with two points of gross retention on $14M of ARR. Combined, that's a $592K benefit on $270K of spend, and the deck says 2.2x.

Checked level by level, the reported figures change.

Level 1. The $270K covers the first three cost layers, which account for 45% of true AI spend under the AI Cost Optimization model. Labor at 40%, waste at 10%, and risk at 5% never entered the figure. True annual cost is closer to $600K.

Level 3. The 4.2 minutes saved per ticket across 88,000 tickets a year is 6,160 hours, which is real and measured against a baseline the ticketing system already held. At 1,700 productive hours per agent, that's 3.6 FTE of capacity.

Level 4. Volume grew 19% and the team didn't hire. Four agents that the old ratio would have required were never posted, worth $312K fully loaded. That claim is supported, because the baseline predates the deployment and the ratio is documented. The retention claim is not supported. A pricing change shipped in the same quarter, and no one recorded which accounts were affected by which change. It fails question 3.

Level 5. Verified benefit of $312K against true cost of $600K puts the program at roughly 52 cents on the dollar in year one, against a reported 2.2x.

The finding gives the operating partner two actions. The first is to run the Cost Optimization pass on the $600K, which returns 20 to 34% in year one. The second is to narrow the deployment to the workflows that produce the verified benefit. Doing both reverses the position within a quarter. Without those actions, the program would have been cut on the basis of a number nobody trusted.

Board Questions a Leadership Team Should Be Able to Answer

Seven questions. A team that can answer all seven with evidence is operating at Level 4 or better.

• Where are we spending on AI, including labor and waste?
• Which initiatives are creating measurable value?
• Which ones aren't?
• Which investments should be expanded?
• Which should be stopped?
• What economic outcomes have been realized to date?
• How is AI affecting operating leverage?

The two hardest questions on that list are the third and the fifth. A program with no answer to them has no mechanism for stopping anything. Its cost keeps growing, and the board will ask about it more often.

Crosswalk: The Three Economic Lenses

CCA runs three AI economics frameworks. Each answers a different executive question and uses its own verb.

Question the executive is asking	Framework that owns it	Verb
Where should the next AI dollar go?	AI Value Creation	Allocate
Which dollars are already being wasted?	AI Cost Optimization	Recover
Did the investment produce anything?	AI Value Attribution	Verify

AI Value Creation identifies where AI capital should be deployed across the six AI Investment Categories. This framework verifies whether the deployment produced anything. Run in sequence, capital allocation is followed by outcome verification, and the verification result feeds the next allocation cycle.

AI Cost Optimization works on the same Level 1 number from the other side. Attribution treats cost as the denominator it measures value against. Cost Optimization treats that denominator as the thing to reduce. Its 6 Cost Layers explain why Level 1 is under-measured almost everywhere. People account for 40% of AI spend, waste for 10% and risk for 5%, and none of it appears in a stated AI cost figure built from vendor invoices. A company that measures Level 1 badly will overstate Levels 4 and 5 by the size of the missing 55%.

CLEAR™ runs attribution as a workstream inside Leverage and Accelerate. Leverage deploys AI against EBITDA, and attribution verifies that the EBITDA gain occurred and names the initiatives that produced it. Accelerate applies the same test to revenue-side deployments. There the confounders are heavier and baseline discipline matters more.

PRISM™ gates the whole exercise through the Strategic Data Assets dimension, scored in the AI Readiness Index. Levels 4 and 5 need outcome data that's joined, historical, and trusted. A company scoring low on data foundations cannot measure past Level 2, whatever it intends. For that company the first step is to fix the data. An attribution study commissioned before that would return noise.

Frequently Asked Questions

What is AI Value Attribution? The practice of connecting AI spending to measurable business outcomes. It runs on a five-level hierarchy (Cost, Adoption, Productivity, Business Outcomes, Enterprise Value) and its output is a defensible statement of which AI initiatives produced economic results and which didn't.

How is this different from measuring AI ROI? An ROI figure is a single output. This framework is the method that makes the output defensible: it names the level your evidence reaches, forces a baseline before deployment, and requires an owner on every claim. Most stated AI ROI numbers are Level 3 productivity gains presented in Level 5 language.

Why do most organizations stall at Level 2? Adoption metrics arrive free with the tooling, and outcome metrics have to be designed before the deployment. By the time leadership asks the outcome question, the pre-deployment baseline no longer exists.

Can we start attribution on a program that's already running? Yes, with a smaller claim. Historical baselines can sometimes be reconstructed from systems that were logging anyway, such as ticketing, CRM, or version control. Anything that depended on a survey or a manual process usually cannot be recovered. The practical step is to apply the four questions to everything not yet approved, then rebuild whatever can be reconstructed for the programs already running.

Who owns AI value attribution inside a portfolio company? The CFO owns the measurement standard and the reporting; a named business owner owns each initiative's result. When attribution sits inside an AI center of excellence, that team ends up assessing its own results.

What does this cost in effort? The four pre-approval questions add a meeting per initiative. Building the measurement for a first cohort of initiatives is a few weeks of finance and data engineering time, most of it spent defining metrics rather than building pipelines.

How does this connect to exit value? A buyer discounts an AI story it can't verify, and a diligence team will ask for the baseline. Attribution evidence converts an AI narrative into an underwritable line, which is what the CLEAR™ Realize phase is protecting.

Does this apply to a company still running AI pilots? That's the cheapest time to apply it. Baselines captured while the estate is small cost almost nothing, and they're the asset that makes every later claim defensible.

Next Step

Bring the AI spend, the initiative list, and whatever measurement exists today. CCA takes each initiative through the five levels and records the highest level its evidence supports. We return the defensible number, with the four questions applied to the next initiatives in the queue. Book a call to run it against a specific portfolio company.

v1.0 · Updated August 2026