The Agent ROI Floor.
Most enterprise agent business cases plan against the ceiling. The number that decides whether the program is real is the floor.
Every agent program has a business case. Most business cases are fiction — not because the team lied, but because they planned against the ceiling instead of the floor.
The ceiling is what you sell to the board. The floor is what determines whether the business case should have been brought to the board in the first place.
The number that decides the program
The number that decides whether an agent program is real is not the projected ROI. It is the floor — the minimum value the program must clear after the costs that did not make it into the business case, multiplied by the probability the program actually ships.
A board-ready business case starts with the ceiling: projected annual value, three-year horizon, an attractive multiple over development cost. The floor calculation is the opposite. Start with the same projected value. Then subtract everything the business case underweighted: orchestration build, ongoing operations, change management, ship-probability.
If the floor is meaningfully positive, the program is real. If not, it is a projection.
The numbers everyone is quoting
MIT Project NANDA’s State of AI in Business 2025 reported that just 5% of integrated enterprise AI pilots were extracting millions in value, while the vast majority showed no measurable P&L impact. The methodology is transparent enough to use as a directional outside view: more than 300 disclosed initiatives reviewed, 52 structured interviews, and 153 senior-leader survey responses.
BCG’s The Widening Value Gap in AI (September 2025, n=1,250 senior executives) sharpens the picture: 75% of organisations prioritise AI, but only 25% report capturing significant value — and just 5% are “future-built,” with the remaining 60% classified as laggards.
Pair these and the implication for any business case is direct. The base rate is failure. Most agent programs do not earn back the development cost, let alone the ongoing cost of ownership. If your business case does not survive a ship-probability multiplier, you do not have a business case.
What the wrong business case actually costs
The cost of a missed business case is not the wasted spend on the failed pilot. It is the next program.
If the first business case comes in well below its projection, finance learns to discount the next one. Future AI proposals get measured against that earlier shortfall, and real programs get cut at the budget line because the previous one set the wrong baseline. That response is reasonable. It is what any finance leader does after being burned: trust the next forecast less.
The result is a compounding cost. Each unrealistic business case raises the bar for the next, and the company falls behind companies whose first numbers held up under operational pressure.
The four-part floor calculation
The floor is a four-part subtraction from the projected value. Each component is simple to estimate. The discipline is doing the math before the program starts, not after the budget overruns.
Build-to-production cost. The line in the business case is almost always the development cost. The line that should be there is the build-to-production cost, which includes the orchestration layer, the governance work, the observability infrastructure, and the integration with existing systems. At most enterprises today, the build-to-production cost is two to three times the development line in the business case.
Ongoing operations. Model and inference costs at scale. On-call rotations. Compliance reviews. Regulatory re-certification when rules change. For a first-pass floor model, reserve a named annual operating line instead of treating maintenance as a rounding error. Over three years, that compounds.
Change management and adoption cost. Often missed entirely. The recent IBM Institute for Business Value research notes that 83% of CEOs say AI success depends more on people adoption than technology. If the people side of the program is not budgeted — training, role redesign, incentive alignment — the projected value never materializes, no matter how well the system works.
Ship-probability discount. The most important multiplier and the one most business cases skip. Public benchmarks give you a low outside view: BCG’s 2025 Widening Value Gap found only 25% of organisations capturing significant value from AI, and just 5% classified as “future-built.” MIT NANDA found only 5% of integrated pilots producing large measurable value. If you do not have specific reasons to believe your program will outperform that range, start with a conservative multiplier and argue up from there with evidence.
If the residual value is positive and meaningful after the four subtractions, the program is real. If not, kill it before the spend starts.
A starter worksheet
Use conservative ranges first. Let the sponsor argue the number up only with named evidence.
| Line | Conservative default | What improves it |
|---|---|---|
| Projected annual value | Sponsor forecast discounted by 25% | A measured baseline and named adoption owner. |
| Build-to-production cost | 2.5× development estimate | Existing orchestration, governance, and integration patterns. |
| Three-year operations | 30% of build cost per year | A funded operating team and clear on-call ownership. |
| Change management | 15% of build cost | Role redesign, training budget, and business-unit incentives already approved. |
| Ship probability | 0.25-0.30 | Prior production deployments with the same team, risk profile, and governance path. |
This worksheet is intentionally plain. Its value is not precision; its value is forcing the optimistic case to name the assumptions it depends on.
The decision rule
Before the program is funded, write down five numbers and one paragraph.
- Projected annual value at steady state.
- Build-to-production cost, computed as 2.5× the development line.
- Three-year operations cost, at 30% of build cost per year.
- Change-management cost, at 15% of build cost, paid in the first eighteen months.
- An honest probability the program reaches production. Default to 0.25-0.30; argue up from there with specifics.
Compute the floor at the three-year mark. If the floor is positive and meaningful, fund the program. If not, do not start.
The discipline is simple. The pushback you will get is not — because the people building the business case are usually compensated on the ceiling and have nothing at stake on the floor. That is the real reason most business cases are fiction.
When to talk to us
Take the floor calculation to the board before the board takes it to you. The conversation is easier when the number was computed deliberately, not extracted in autopsy. The build-to-production cost line is the same one build-vs-buy gets wrong, and the reliability discount is the same one SLOs are designed to surface.
Computing the floor is half the work. The other half is committing to the answer when it comes back negative.
References
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (2025). mlq.ai PDF
- BCG, The Widening Value Gap in AI (September 2025). bcg.com
- IBM Institute for Business Value, The Rise and ROI of the Chief AI Officer (2026). ibm.com
About the author
Kevin Dickerson is a co-founder of Loom. His machine learning research predates the LLM era, and he has worked at the frontier of production AI across cloud platforms, semiconductor companies, and enterprise programs.
Continue reading
Build vs. Buy in the Agent Era: A Decision Framework.
The build-vs-buy decision is not about engineering capability. It is about whether you can sustain the thing you build.
Enterprise TransformationWhat Cloud Transformation Taught Me About Agentic Programs.
The hardest lessons of enterprise cloud transformation are about to be relearned in AI. Most of them are organisational, not technical.
CAIOThe Fractional CAIO Playbook.
Most full-time Chief AI Officer hires at this stage will fail. The work is real. The shape of the hire is wrong.