Written by: Akshay Shrimanker, CPA, and Rakesh Shah
AI is not just a product feature. It is a cost structure shift.
For years, many tech companies could explain their unit economics with a familiar formula: cloud hosting plus a support layer, then Opex dominated by headcount and go-to-market. That model still works for some businesses, but it breaks down fast once AI becomes embedded in the product and in internal workflows.
The shift shows up in two places: COGS (infrastructure) and Opex (especially the software stack and specialized help). If you are building budgets or forecasts the same way you did a couple of years ago, you risk underestimating how quickly these lines can move.
COGS: When “infrastructure” becomes usage-based AI spend
In many companies, infrastructure used to be mostly cloud hosting. It was not perfectly fixed, but it was predictable enough. AI changes that because inference costs behave like a usage meter.
As our Fractional CFO Rakesh Shah explains it, once a product relies on model calls, “every time users do something, that runs up tokens, and those tokens cost money.” Those costs are often small per action, which is why they sneak up. At scale, they can become one of the most volatile components of COGS.
Akshay Shrimanker, CPA, describes the financial surprise this creates in plain terms: “What the old subscription cost covers is the margin to pay for cloud and customer success. But it doesn’t cover the compute costs.” That gap is where gross margin compression comes from, especially for companies that add AI features without rethinking pricing, limits, or packaging.
A practical way to forecast AI COGS
You do not need a perfect telemetry setup to start modeling AI costs. You need a driver and a unit.
- Pick a usage driver you can measure or reasonably estimate:
- AI feature users
- AI-enabled workflows
- Prompts per active customer
- AI-assisted tasks completed
- Estimate a cost per unit from recent actuals.
Rakesh puts it simply: “I just tend to base it on a per-user average cost.” - Add a buffer for outliers.
Power users and edge cases can create situations where a few customers consume a disproportionate share of usage. If your forecast only uses averages, the first time you land a heavy user cohort, your model will miss. Treat this as a real line item, not a rounding error. - Track “token creep” as a trend, not a one-time estimate.
Rakesh’s warning is short and useful: “Pay attention to token creep.” In practice, this means monitoring whether the cost per user is drifting up because features are getting more model-heavy, customers are using AI more frequently, or your architecture is generating more calls per workflow.
The core idea is to stop treating AI costs as part of a single “infrastructure” bucket you only review quarterly. If it is material, separate it and give it an owner.
Opex: AI pushes spend into vendors, tooling, and specialized support
AI also changes Opex, even for companies that are not building AI as their core product. The reason is simple: tools and workflows across the company are becoming AI-enabled, and that tends to come with new pricing tiers and add-ons.
Rakesh notes a pattern many finance teams are seeing: “Every typical SaaS product… now has an upgraded version where you can incorporate AI, and if you do that, it costs more.” That cost increase does not always show up as a clean “AI expense.” It often looks like a subscription re-tiering, a new per-seat add-on, or a higher minimum plan.
This matters for forecasting because the software stack can inflate even if headcount stays flat:
- Pricing changes at renewal
- Seat creep as more teams adopt the tool
- Add-ons that start as optional and become standard
The simplest way to keep control is to treat your vendor stack like you treat payroll: model it with drivers and timing, not as a single annual number.
A practical way to forecast Opex under AI-driven pricing pressure
- Build a renewal calendar with expected uplift
- Separate baseline subscription costs from add-ons where possible
- Tie seat growth assumptions to headcount or department growth
- Add a “new tooling” reserve for the year, especially for security, data, and AI workflow tools
Opex can also rise through people costs tied to AI adoption. Even when a company intends to “use AI to be more efficient,” the implementation period often requires specialized help, whether that is consultants, contractors, or niche hires. If your forecast assumes immediate savings without transition costs, it will miss.
The budgeting takeaway: drivers first, categories second
The biggest budgeting mistake in this new environment is treating COGS and Opex as static categories rather than living systems driven by usage, packaging, and adoption.
Rakesh’s rule applies here: “Budgeting shouldn’t be just numbers on a spreadsheet. Each number should be explained by the drivers.”
For AI-era planning, those drivers usually include:
- AI feature adoption and intensity of use
- Cost per unit of AI usage, plus an outlier buffer
- Vendor repricing and seat growth
- Temporary implementation support and specialized expertise
If you build budgets and forecasts around those drivers, you can react faster. You will spot margin pressure earlier, understand whether a price increase is justified, and avoid being surprised by a COGS line that suddenly behaves like a variable meter.
And that is the point. AI can absolutely be a growth lever, but only if your financial model reflects the way it truly behaves: dynamic, usage-based, and increasingly tied to both COGS and Opex.
How ShayCPA can help
ShayCPA helps early-stage and growth tech companies build models that hold up in the real world, especially when costs are shifting quickly due to AI and vendor repricing. We can help you:
- Set up a driver-based forecast that ties COGS and Opex to usage, headcount, and renewals
- Separate and track AI-related costs (including token-based inference spend) so gross margin stays visible
- Build a renewal calendar and a vendor spend model to reduce surprise stack inflation
- Pressure-test pricing and packaging assumptions based on unit economics and cash runway
- Establish a monthly forecasting cadence with clear owners, assumptions, and variance review
If you want a second set of eyes on your model or help rebuilding it around the drivers that matter, we are happy to help.
Disclaimer:
The content provided on this blog is for general informational purposes only and does not constitute professional accounting, tax, or legal advice. Reading or accessing this material does not create a CPA-client relationship, nor should it be construed as a substitute for individualized guidance from a qualified professional. While we strive for accuracy, Shay CPA PC makes no warranties—express or implied—about the completeness, reliability, or timeliness of the information, and we expressly disclaim liability for any errors or omissions. You should not act or refrain from acting based on any blog content without seeking the advice of a qualified CPA or other professional who can address your specific circumstances. Links to external resources are provided for convenience only and do not imply endorsement. Shay CPA PC is under no obligation to update this content and disclaims responsibility for decisions made in reliance on it.