Most finance teams at growth-stage companies have treated AI spend the same way they treated cloud spend in 2014: a line in opex, probably sitting under software subscriptions, reviewed at renewal and otherwise ignored. That’s about to become a problem.

When Anthropic announced the formation of a dedicated AI services company alongside Blackstone, Hellman & Friedman, and Goldman Sachs in May 2026, the signal wasn’t about AI capability. It was about delivery structure. 

Institutional capital is now backing the vehicle that will bring frontier AI into mid-sized companies, such as community banks, regional health systems, and mid-market manufacturers. The engagements that follow won’t look like software subscriptions. And the finance function that treats them like one will have a difficult conversation in the next data room.

The contract structure is the first problem

A typical engagement with an AI services firm starts with a scoping and discovery phase, moves into custom build, and then into ongoing support and iteration. Applied AI engineers work on-site or embedded with the customer’s team. The cost profile across those phases looks nothing like a SaaS contract.

Discovery and scoping fees are often fixed. Build phases are frequently time-and-materials or milestone-based. Ongoing support can be retainer, consumption-based, or a hybrid tied to usage metrics that didn’t exist when the contract was signed. 

Most finance teams haven’t categorised anything like this before, and the categorisation decision (made quickly, under pressure, when the first invoice arrives) tends to stick.

The question of whether an AI deployment engagement is capex or opex isn’t academic. If the engagement produces a system your company will use and control over time, there’s a reasonable argument for capitalisation. If it’s closer to consulting and configuration, it’s opex. Most engagements are somewhere in between, and that ambiguity is exactly the kind of thing an auditor will ask about and an investor will probe.

The companies that move through this cleanly are the ones that made the categorisation decision deliberately, documented the rationale, and applied it consistently from the first invoice. The ones that didn’t are restating periods, or defending inconsistent treatment across quarters, at the worst possible time.

The balance sheet question most teams miss

If an AI deployment creates an asset, such as a custom model, a proprietary workflow, a system built on top of third-party infrastructure, the next question is what that asset is worth and how it depreciates.

This is where most finance teams have no framework at all. A capitalised software development cost has well-established treatment. An AI-powered system built partly by a third-party services firm, running on a foundation model licensed from a vendor, trained on your proprietary data is a different conversation. The useful life isn’t obvious and the impairment indicators aren’t established. And if the underlying foundation model changes significantly, or the services firm relationship ends, the carrying value may need revisiting faster than the depreciation schedule assumed.

None of this is intractable. But it requires a position to have been taken, not just a number to have been entered. Investors reading a data room will find the capitalised costs quickly. If the treatment isn’t documented and defensible, it raises questions about everything else in the financial statements.

Dependency risk is the new concentration risk

Investor diligence on growth-stage companies has always included customer concentration. In the last few years it’s expanded to include supplier concentration and key-person risk. The next addition is AI vendor dependency and it’s arriving faster than most finance teams have prepared for.

An engagement with a firm like the one Anthropic has built with Blackstone and Goldman creates a specific kind of dependency. If the company’s core operations run on systems built and maintained by a third-party AI services provider, the termination or disruption of that relationship is a material operational risk. Most investor questionnaires don’t have a field for it yet. They will.

The finance function’s job here isn’t to avoid the dependency. It’s to make sure it’s documented, disclosed where appropriate, and that the reporting can answer the question when it arrives:

• What does the engagement cover? 
• What’s the notice period? 
• What happens to the system if the relationship ends? 
• Is there a transition plan? 

These are questions a prepared finance team answers in advance. An unprepared one answers them under diligence pressure, without clean documentation, in front of investors who are already forming a view.

What to do before the contract is signed

The window to get the framework right is before the first engagement starts, not after the first invoice arrives.

That means three specific things: 

1. A categorisation policy for AI deployment costs: covering the capex/opex decision, the criteria for capitalisation, and the depreciation methodology if costs are capitalised. 
2. A vendor dependency disclosure framework: a consistent way of assessing and documenting material AI service relationships, so that when an investor asks, the answer is already in a format that’s usable. 
3. A close process: that can actually capture these costs cleanly, in the right period, against the right categories, without manual reconstruction at month end.

This is the kind of work that often gets skipped when a company is moving quickly and AI spend feels small. It feels different when the engagement is material, the auditor has questions, and the data room is open.

The businesses that navigate these transitions cleanly tend to have someone embedded in the finance function who understands both sides: what the technology is doing commercially, and what an investor will ask about it in a data room. 

PIF Advisory works with growth-stage companies as that embedded partner, bringing accounting rigour and investor readiness together before the questions get asked. If you're approaching a new payment rail, an AI services engagement, or a fundraise, get in touch before the process starts. 

Book a Discovery Call with PIF Advisory Today