Is an agentic AI buildout a capital investment or an operating expense? PE boardrooms are split, and the answer decides how fast a portfolio company moves, how much room it has under its covenants, and what a buyer pays for the result.
Software buying used to sort itself. SaaS subscriptions landed in operating budgets. ERP and infrastructure projects lived in capital conversations. Agentic AI sits between them because it behaves more like labor than like software.
When a portco puts agents into customer support, finance operations, procurement, or field service scheduling, what is it buying? A tool, a workforce multiplier, a transformation program. The answer changes with who is in the room, and each one carries a different line on the P&L.
CFOs want a position because the numbers are material. These are not $200K pilots. They are multi-million dollar programs tied to EBITDA improvement and SG&A reduction, and spreading that cost over its useful life changes what the company can attempt this year.
Operators see the same system differently. Models retrain, workflows adapt, agents accumulate context. The implementation is never finished, which is the strongest argument for treating it as ongoing operational spend.
There is no AI-specific US GAAP. Companies apply ASC 350-40, the internal-use software standard, by analogy. Preliminary work gets expensed. Once management authorizes the project and commits funding and completion is probable, direct development costs capitalize, and training, validating, and optimizing a model during development is generally treated like coding. Once the system is live, maintenance and monitoring return to expense.
FASB's ASU 2025-06 rewrote the front half of that test. It removed the waterfall project stages and replaced them with two conditions: funding is authorized and committed, and completion is probable. Capitalization waits while "significant development uncertainty" remains, which the standard ties to novel technology and undefined performance requirements. It takes effect for annual periods beginning after December 15, 2027, with early adoption permitted.
Hold an agent program against that bar and the split gets clear.
The foundational layer capitalizes cleanly. The agents running on top are where it gets contested, because "undefined performance requirements" describes most agent projects at kickoff, and under ASU 2025-06 that uncertainty defers capitalization rather than supporting it.
Which is the part most AI business cases skip. A software asset gets cheaper to run as it scales. An agent gets more expensive.
Classification determines which budget the program has to win.
Called opex, it competes with this quarter's margin, inside a company carrying an EBITDA-based leverage covenant and a board watching that number monthly. Ambition gets rationed into pilots small enough to pay back within the quarter.
Called capital, it competes with the capital plan and gets judged on a multi-year return, which is the only place enterprise-wide programs survive.
Same technology, very different velocity. Most portcos have landed somewhere on this by default rather than by decision, and their sponsors often do not know where.
Capitalizing lifts reported EBITDA during the hold. Buyers know it.
Quality of earnings work routinely reverses aggressive capitalization and values the business on cash EBITDA. Capitalizing one to three percent of revenue can add 100 to 300 basis points of reported EBITDA margin before amortization ramps, and the QofE team's job is to take it back out. A policy that changed mid-hold gets found. So does a capitalized software balance that grew faster than anything the company shipped.
The treatment that survives diligence is the one the company can document: a real authorization date, evidence the performance criteria were defined and met, a clean line between development that added functionality and maintenance that kept the lights on, and agent run costs tracked separately from build costs from day one.
For software portcos there is one effect no capitalization policy fixes. Customer-triggered inference belongs in COGS, and ICONIQ benchmarks put AI-native companies near 52 percent gross margin against the 70 to 80 percent that built SaaS valuations. Gross margin drives the multiple more directly than EBITDA does, so an agent strategy that moves cost above that line can lower exit value while improving the operating story.
Operating partners are asking something the accounting cannot settle. If agents absorb work people used to do, should they be measured like technology investments at all?
The models being built now look like workforce models: cost-to-serve compression, less outsourced labor, faster cycle times, wider spans of control, higher revenue per employee. That is operating model redesign wearing a software budget.
It explains why these programs left IT. Ownership sits across operations, finance, transformation, and increasingly the deal team, and the question has moved from which model the company uses to how permanently this changes the way it runs.
The firms getting the most out of this instrument agent spend well enough to defend whatever classification they chose, with build and run separated from day one. And they treat the classification as a hold-period decision rather than a verdict on what the investment is worth, because the frameworks are going to trail the operating reality here by years.
.png)
We're offering $50K in funding for PE-backed businesses investing in AI innovation. Here's how it works:
This is the fourth post in a series on AI value creation across the private equity portfolio.
