How AI Investments Affect IT Budgets

The story of many 3+9 rolling AI budget reviews this month: What is our “AI spend? Or - it’s Why do finance, IT and the business all have different AI numbers? Your 2026 AI budget risk IS NOT OVERSPEND, but INVISIBLE SPEND. I’m seeing the same pattern on repeat.  - cloud line items tagged as “platform,”  - vendors rebadging old SaaS as “AI,”  - internal teams parking time under “innovation,” and  - business units expensing copilots and point tools on credit cards. Recent studies put global AI spend at roughly 1.5 trillion by the end of 2025, but a large chunk of that now hides inside generic IT, SaaS, and “productivity” budgets where nobody can see it cleanly. Shadow AI makes the problem worse, not better. Around 80% of office workers already use AI, yet only about 22% stick solely to employer-approved tools, which means most organisations are paying twice:  - once for official platforms,  - once again for unsanctioned apps and duplicated workflows. Data from security and breach reports shows that incidents involving shadow AI cost on the order of hundreds of thousands more per breach, because nobody budgeted for the clean-up! Copilot rollouts are the current slow-motion car crash. - Gartner expects AI-related SaaS costs, especially copilots, to grow 30–50% annually as licences and usage outpace governance. - Enterprises are discovering that what started as “a few pilot seats” has quietly turned into one of the fastest-growing opex lines once you add usage-based fees, training time, storage growth, and auto-renewals finance never saw. If you want a practical 90-day fix, skip the pretty dashboards and do this sort of cost hygiene:  1. define 3-4 AI cost buckets (cloud AI services, model/API, copilots and AI-SaaS, internal labour),  2. enforce tagging by workflow or product,  3. separate elastic inference from fixed platform and people costs so you can actually see unit economics.  4. Then put one simple guardrail in place: no net-new AI spend without a costed workflow and a named owner on the hook for value, not just “adoption.” Because if you don’t get AI spend to finance-grade visibility this year, procurement and the CFO’s office will eventually “solve” it for you. Likely with blunt cuts that don’t distinguish between flaky experiments and the workflows that are actually moving the needle.

𝐌𝐨𝐬𝐭 𝐭𝐞𝐚𝐦𝐬 𝐮𝐧𝐝𝐞𝐫𝐞𝐬𝐭𝐢𝐦𝐚𝐭𝐞 𝐀𝐈 𝐜𝐨𝐬𝐭𝐬. They budget for models… but forget everything around them. That’s why AI projects often look “cheap” in pilots — and expensive in production. Real AI spend isn’t just inference. 𝐈𝐭’𝐬 𝐬𝐩𝐫𝐞𝐚𝐝 𝐚𝐜𝐫𝐨𝐬𝐬 𝟏𝟐 𝐦𝐚𝐣𝐨𝐫 𝐜𝐨𝐬𝐭 𝐛𝐮𝐜𝐤𝐞𝐭𝐬 𝐞𝐯𝐞𝐫𝐲 𝐂𝐅𝐎 𝐚𝐧𝐝 𝐂𝐓𝐎 𝐬𝐡𝐨𝐮𝐥𝐝 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝 👇 𝟏) 𝐂𝐨𝐦𝐩𝐮𝐭𝐞 (𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 + 𝐅𝐢𝐧𝐞-𝐭𝐮𝐧𝐢𝐧𝐠) GPUs, clusters, distributed runs. Costs rise with experiments, retries, and large models. 𝟐) 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 / 𝐑𝐮𝐧𝐭𝐢𝐦𝐞 (𝐓𝐨𝐤𝐞𝐧𝐬) API usage, token billing, agent tool calls. Driven by query volume and long contexts. 𝟑) 𝐃𝐚𝐭𝐚 𝐒𝐭𝐨𝐫𝐚𝐠𝐞 Warehouses, lakes, vector databases, feature stores. Embeddings, duplicates, and retention drive spend. 𝟒) 𝐃𝐚𝐭𝐚 𝐋𝐚𝐛𝐞𝐥𝐢𝐧𝐠 & 𝐇𝐮𝐦𝐚𝐧 𝐑𝐞𝐯𝐢𝐞𝐰 Annotations, SMEs, RLHF, QA checks. High-quality labeling is slow and expensive. 𝟓) 𝐃𝐚𝐭𝐚 𝐏𝐢𝐩𝐞𝐥𝐢𝐧𝐞𝐬 & 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 Ingestion, ETL/ELT, cleaning, transformations. Messy data creates ongoing maintenance costs. 𝟔) 𝐌𝐨𝐝𝐞𝐥 𝐃𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 (𝐏𝐞𝐨𝐩𝐥𝐞 𝐂𝐨𝐬𝐭) ML engineers, data scientists, prompt engineers. Hiring, retention, and specialist premiums add up. 𝟕) 𝐌𝐋𝐎𝐩𝐬 / 𝐋𝐋𝐌𝐎𝐩𝐬 𝐓𝐨𝐨𝐥𝐢𝐧𝐠 Model registries, prompt versioning, evaluations. Tool sprawl and enterprise licenses increase overhead. 𝟖) 𝐌𝐨𝐧𝐢𝐭𝐨𝐫𝐢𝐧𝐠 & 𝐎𝐛𝐬𝐞𝐫𝐯𝐚𝐛𝐢𝐥𝐢𝐭𝐲 Drift detection, hallucination monitoring, logging. Traces, alerts, and eval pipelines aren’t free. 𝟗) 𝐒𝐞𝐜𝐮𝐫𝐢𝐭𝐲 Access control, secrets, red teaming, threat detection. Prompt injection and data exfiltration risks require investment. 𝟏𝟎) 𝐆𝐨𝐯𝐞𝐫𝐧𝐚𝐧𝐜𝐞 & 𝐂𝐨𝐦𝐩𝐥𝐢𝐚𝐧𝐜𝐞 Documentation, policies, audits, legal reviews. Regulations like GDPR and EU AI Act drive ongoing costs. 𝟏𝟏) 𝐈𝐧𝐭𝐞𝐠𝐫𝐚𝐭𝐢𝐨𝐧 & 𝐂𝐡𝐚𝐧𝐠𝐞 𝐌𝐚𝐧𝐚𝐠𝐞𝐦𝐞𝐧𝐭 Connecting AI to apps and workflows, training users. Adoption takes time and process redesign. 𝟏𝟐) 𝐕𝐞𝐧𝐝𝐨𝐫 & 𝐏𝐥𝐚𝐭𝐟𝐨𝐫𝐦 𝐂𝐨𝐬𝐭𝐬 SaaS tools, orchestration platforms, marketplaces. Watch for hidden add-ons and per-seat pricing. 𝐓𝐡𝐞 𝐭𝐚𝐤𝐞𝐚𝐰𝐚𝐲: AI budgeting isn’t a line item. It’s a system. If you only plan for tokens, you’ll miss most of the spend. If you plan across these 12 buckets, you build AI that scales sustainably. Save this if you’re planning AI investments. Share it with your CFO or CTO. ♻️ Repost this to help your network get started ➕ Follow Prem N. for more

A year ago, AI development tools were a rounding error on a software delivery budget. Today, they’re the second-largest variable cost after people — and most leaders aren’t governing them. I’m a partner leading AI product delivery at a global firm. Over the last six months, I’ve watched AI tooling go from “nice productivity boost” to a line item that competes with cloud infrastructure spend on every engagement I touch. Factory, Copilot, Cursor, Replit, agentic coding platforms — they’ve quietly become a real budget category. And the governance has not caught up. Here’s what I’ve learned the hard way: 1. Token consumption is unpredictable. Without per-user caps and weekly tracking, you don’t have a budget — you have a hope. 2. Tool choice matters more than tool quality. A flat-fee subscription and a consumption-based platform can produce comparable work. Defaulting to the expensive one is the most invisible form of waste. 3. Engineers are the investment, not the variable to optimize. AI tools amplify good engineers. They don’t replace them. Govern the tools, not the people. 4. The “4-5x productivity gain” is real, but conditional. It shows up on isolated work like unit testing. It doesn’t show up reliably on complex integration work. Treating it as a flat assumption is how timelines slip. 5. AI tooling cost belongs in the SOW. Most don’t include token estimates. That gap is where surprises live. The honest truth: this is a new discipline, and the playbook is being written right now. The leaders who treat AI tooling as a governance question — not just a developer convenience — will be the ones who deliver predictable outcomes in the next two years. I’m building this discipline alongside my team and contributing what we learn back to my firm’s broader practice. If you’re navigating the same questions, I’d love to hear how your team is approaching it. #AIDelivery #SoftwareEngineering #AIGovernance #EngineeringLeadership #FinOps

JPMorgan Chase's Jamie Dimon confirmed in early 2026 that the bank has reclassified AI spending from discretionary innovation to core infrastructure, treating it with the same priority as cybersecurity and operational resilience. The 2026 tech budget is approximately $19.8 billion. AI is roughly $2 billion of that. Total operating expenses are projected at $105 billion, a 10% year-over-year increase driven heavily by AI and technology investment. The internal LLM Suite serves more than 60,000 employees as a secure interface to external models for document summarization, drafting, and ideation without exposing data to public AI services. The reclassification is the important part for enterprise CIOs. Most enterprise AI budgets are still innovation line items, which means they get cut first when revenue softens. JPMorgan moved AI to the same accounting category as the data center, the payment rails, and the SOC. That category does not get cut. It gets defended. For enterprise CFOs and CIOs, this is the pattern to watch. If your 2027 AI budget is still on the innovation page, you are one bad quarter away from losing it. If it has moved to core infrastructure, you have permanently changed the cost baseline of the company. Most banks have not made this shift yet. The largest one already has. The competitive gap is going to compound through 2027.

Are my clients rethinking their AI strategy and spending levels in light of this week’s AI meltdown? No. If you know how to calculate ROI, you can justify the business’s AI spending upfront, so this week’s AI correction isn’t affecting your strategy. The ROI from last year’s data and AI initiatives should pay for this year’s spending. This year’s ROI should have already paid for next year’s AI budget. Increased spending must be offset by realized gains in revenue or cost savings. Each year during the budgeting process, we break down the gains from the current year’s initiatives and ask for a small part of that ROI to be reinvested into next year’s AI initiatives. Estimating ROI is always workflow-centric. Start with the current state and define the value it creates as the baseline. Introducing data, analytics, machine learning, or AI into the workflow changes it, defining the future state. We can estimate the value of the future state workflow in terms of cost savings, improved business outcomes, or features that customers are willing to pay for. Typically, this is a range, not a single value. The upfront estimation justifies the spending. We must also track ROI after the initiative is deployed to verify our estimation framework’s accuracy. This tracking helps justify next year’s spending. If we made the business $20, we can ask for $5 of it back to fund next year’s data and AI budgets. If we made the business $20 more in 2025 than we did in 2024, we can ask for an additional $5 to fund more data and AI initiatives in 2026. This is how you build an ROI flywheel. Prove value. Use it to justify doing more the following year. Deliver more value. Accelerate delivery and increase the number of initiatives funded next year. Maintain the cycle. Businesses that implement an AI ROI Flywheel don’t need to worry about hype cycles and market sentiment swings. Focus on delivering and quantifying customer and business outcomes. The rest is noise.

The move to implement #AI isn’t a walk in the park. It’s a Formula 1 race. What is keeping many executives awake at night, and rightly so, is the realization that the "entry ticket" for this race isn't just plugging in a LLM. The issue runs deeper: you first need to get your technical house in order. 30% of your budget won't go to innovation, but rather to paying for the past. Did you know that? In a recent IBM IBV study, executives estimate that between now and 2027, a significant slice of AI investment (between 18% and 29%) won't be spent on futuristic models or features. That money has a less glamorous but critical destination: paying off Technical Debt. What does this mean in practice? Imagine that in the past, to deliver software faster, your company took "shortcuts." You chose band-aids instead of surgery, skipped database updates, or accepted integration "workarounds." Technical Debt is the interest you are paying now on those decisions. Why is AI calling in this debt now? Artificial Intelligence is demanding. It’s no use buying a Ferrari (Generative AI) and trying to drive it on a potholed dirt road (Legacy Infrastructure). For AI to run, it needs smooth asphalt: • Clean, structured #data. • Modern APIs. • Robust security. The lesson for leaders: When planning AI #ROI, be realistic. Reserve budget for "housekeeping." Ignoring technical debt doesn't make it disappear; it just makes your AI more expensive and less efficient. How is the balance between "buying the Ferrari" and "paving the road" in your company today? #AIStrategy #DigitalTransformation

Adopting AI agents will not shrink your software budget. It will redefine what you are willing to pay for outcomes. This is critical to understand when designing the pricing and packaging of an AI agent solution. That happened with cloud. It’s happening again with AI agents. In 2007, as CTO at Finjan, I started moving our on-prem infrastructure and product delivery to AWS. It felt magical. For the first time infrastructure delivered value on demand. No hardware procurement No waiting weeks for servers No capacity planning Just build → deploy → learn I also believed it was cheaper. No capex Fewer IT people Flexible configurations Then reality arrived. Over time we discovered the cloud cost roughly 3x more than our on-prem environment. In 2015 I ran the analysis again. Different company. Different scale. Same conclusion: about 3x. And still there was no going back. Because speed and value changed what “acceptable cost” means. Now I’m seeing the same pattern with AI. Teams replace SaaS subscriptions with AI agents and API workflows. They get faster and more personalized outcomes. Then the bill arrives: Higher compute Higher API and token usage More operational complexity Often again… around 3x. Technology waves follow a simple law: Acceleration of time-to-value + personalization increases willingness to pay more than it reduces cost. We didn’t move to cloud because it was cheaper. We moved because velocity compounds business outcomes. We won’t adopt AI because it saves money. We’ll adopt it because getting the right answer immediately beats getting a generic answer cheaply. Cost optimization comes later. Adoption starts with value. Which means the real question isn’t: How much budget do AI agents save? It is, how do you price and package the new immediate value they create? Cost reduction has a hard floor. Willingness to pay is capped only by the economic value you create. Evolution Equity Partners #ai pricing #packaging #centerofexcellence

Your Copilot license costs $30/user/month. Your Copilot investment costs $90-$150 per user/month. The license is the number on the invoice. The investment is the number that determines whether you get value or get a line item nobody can justify at renewal. Here’s what the $30 doesn’t include: → Integration work to connect Copilot to your actual workflows → Governance framework so data doesn’t route outside your tenant → Change management so people use it instead of ignoring it → Measurement infrastructure so you know if it’s working → Ongoing optimization so it gets better, not stale 𝗧𝗵𝗲 𝘁𝗿𝘂𝗲 𝗮𝗹𝗹-𝗶𝗻 𝗰𝗼𝘀𝘁 𝗼𝗳 𝗔𝗜 𝗮𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗶𝘀 𝟯-𝟱𝘅 𝘁𝗵𝗲 𝗹𝗶𝗰𝗲𝗻𝘀𝗲 𝗳𝗲𝗲. Most organizations budget for the license and call it the AI investment. Then wonder why adoption stalls at 20%. The uncomfortable truth: if you’re only spending $30/user, you’re not investing in AI. You’re buying shelf software with a smarter label. The organizations seeing 9x more value from Copilot aren’t spending less. They’re spending more — on governance, enablement, and workflow integration — and getting returns that make the total cost look small. The CFO conversation shouldn’t be “how do we reduce our AI spend?” It should be “are we spending enough on the parts that make AI actually work?” Share this with whoever is building next year’s AI budget. The line item needs more lines. #AIStrategy #Copilot #CIO #DigitalTransformation

AI’s cost curve is flattening the wrong way. While attention remains fixated on GPUs and token rates, the real budget killers—power, observability, orchestration, evaluation, and egress—are compounding beneath the surface. OpenAI has committed to $300 billion in compute contracts from Oracle, requiring 4.5 gigawatts of continuous power—more than the output of two Hoover Dams. AI infrastructure already consumes more than 4% of U.S. electricity, with data centers responsible for over 2.2% of total U.S. CO₂ emissions as of 2024. A representative enterprise AI stack might spend 20–40% on observability, 10–25% on evaluation, and only 15–35% on actual inference. A team spending $10,000/month on model usage can see total costs scale to $45,000–55,000/month after factoring in orchestration, retries, logs, test data, and governance. AI data centers globally are projected to emit 2.5 billion metric tons of CO₂ between now and 2030—on par with annual emissions from commercial aviation. 61% of U.S. adults report concern over AI’s energy consumption; yet most enterprise budgets track only token counts, not total system costs. The future of AI at scale won’t be determined by the speed of GPUs but by the ability to measure and control the full-stack cost structure. Power is becoming a competitive differentiator. Observability is no longer optional. And routing, governance, and task-based costing are table stakes for ROI.

I was in San Francisco this weekend and the billboards screamed AI from both sides of the highway. My feed is filled with sample prompts, product launches, and funding announcements. But who's talking about the cost? I've been building agents and experimenting with new platforms. Somewhere in the middle of building, I got my first real surprise: a massive bill from Claude. I hadn't been careless. I'd been curious and moving at speed. Curiosity at speed has a price tag. This matters especially for marketers. We're the ones greenlighting the tools, running the experiments, and kicking off the campaigns. We move fast, but we don't always own the budget line that pays for AI. As AI becomes infrastructure, that will have to change. Token spend is a real line item. And right now, most marketing teams have no idea what they're actually spending. The financial reality: traditional SaaS was built on near-zero marginal cost — gross margins of 80–90% were the norm. AI changes that. Every prompt, every agent call, every token hits COGS. Bessemer Venture Partners' State of AI puts AI gross margins at 50–60%. Redpoint maps the upside: revenue up 25–35%, EBITDA up 100–250% in 3–5 years. But you have to manage costs to realize those gains. There's also the cost that doesn't show up on any P&L. A study published in December estimated that AI systems alone could generate 32–80 million tons of CO₂ this year — comparable to the entire carbon footprint of New York City. The planet matters, even if it's not on the balance sheet. The billboards aren't going anywhere. Neither is the bill. Know what you're building on. Know what it costs. On the P&L and on the planet.