Every fall the same document shows up. Last year’s technology budget, rolled forward, with a percentage on top. For most of the past twenty years that was a defensible act. Technology cost behaved like rent: a few large fixed commitments, a maintenance tail, a project or two, and a total a CFO could hold in mind. Roll it forward, argue about the percentage, move on.
This year that document is the tell. The turn came sometime in the last twelve months, and it did not announce itself as a budget event. It arrived as a clause in a software renewal, a hardware quote that came back higher than the one before it, a board member asking who keeps the list of AI systems in use, and a bankruptcy auction in which a search company paid real money for another company’s email. Each looked like news. Together they are a change in what technology costs, how it is priced, who is accountable for it, and what it is worth. A budget with the same rows as last year and a larger number on the AI line has recorded none of that. The road turned. The budget kept going straight.
What turned
Not the size of the spend. Spend was always going to rise; Gartner now expects worldwide IT spending to grow more than fourteen percent in 2026, and warns in the same release that the money is being stretched by inflation, scarce components, and costlier hardware and memory. Both halves of that sentence matter. Budgets are going up and buying less of the old things. gartner
What turned is the shape. Five things happened at once, and each one moves a row on your sheet. AI stopped being a pilot you bought and became a feature inside the systems you already own, charging by the action rather than the seat. Regulators stopped drafting rules and started dating them, then re-dating them. Boards moved AI from a briefing slide to a fiduciary duty. Your operating data acquired a market price, and your vendors noticed before you did. And the hardware underneath all of it became scarce for the first time in a decade, because the AI build-out is consuming the memory supply.
The sections below take those in the order a CFO will feel them. Every dollar figure is illustrative. The point is the shape of each row, not the number in it.
Governance became a row, not a memo
Two years ago AI governance was a policy document and an hour of training. This year it is a funded function with a name on it, because pressure arrived from three directions at once. From above: more than three in five directors now reserve time on the full board’s agenda for AI, and a board that discusses something expects someone to own it. From Brussels: the Digital Omnibus took effect on July 27, pushing the AI Act’s high-risk deadlines to December 2027 and August 2028 while leaving the transparency rules where they were; those rules, which oblige you to tell a person when they are dealing with an AI and to label synthetic content, arrived on August 2 as planned, and they turn on what a system does rather than its risk tier, so everyday generative tools fall under them. From the states: there is still no comprehensive federal AI statute and no enacted preemption of state laws, so the state laws stand. Colorado rewrote its law in May and reset the effective date to January 1, 2027, trading the original risk-program and impact-assessment duties for four narrower ones: telling people an automated system is in use, disclosing when it produced an adverse result, letting them correct the data behind it, and keeping a meaningful human review. California’s automated-decision rules and training-data transparency duty took effect in January, and Texas required its attorney general’s complaint mechanism to be live by September 1. Navigating AI Adoption and Cybersecurity Oversight | Directors & Boards +6
Read all of that as a budget officer and the lesson is not any one deadline. It is that the deadlines move and the direction does not. So the thing you fund is not compliance with a rule. It is the capacity to comply with whichever rule arrives: an inventory of every AI system in use, including the ones your employees built without asking; a risk class on each; an owner; a decision log; and a review cadence that fits on one calendar. That capacity costs mostly leadership time, which is why it never made it into a budget that only counted software. For an illustrative $60 million company with a $2.4 million technology budget, a governance row of $75,000 to $120,000 is realistic, and most of it is a fraction of one technology leader’s year plus a half day a quarter from the executive team. One more finding belongs here: nearly three in four enterprises have no plan to say who owns technology costs and the tools that create them. The ownership question and the governance question are the same question. A freeway has guardrails so the traffic can move faster; that is what this row buys. Gartner
Your data has a price, and a contract
In August, Google won a bankruptcy auction for the corporate records of Spirit Airlines. The $10 million purchase covered data from the airline’s finance, operations and revenue-management systems, its pricing models and booking curves, roughly 100 million emails, and its software code. The backup bidder was an AI data company at $7.5 million. Read that as a mid-market CEO rather than a traveler. A company’s operating history, the exhaust of running the business for twenty years, cleared a market at a price, and two buyers wanted it. skiftbusinesstraveller
Above the bankruptcy floor, the market is orderly and growing. Wikimedia now sells fast, bulk access built for AI systems as a paid product, with Amazon, Meta, Microsoft, Perplexity and Mistral signed on, and established marketing-data marketplaces have expanded so customers can license data for AI training and for third-party models and applications. The old third-party pipes are constricting at the same time. Since August, California’s Delete Act has let residents scrub themselves from every registered broker through one state system, and more states keep adding a duty to honor browser-based opt-out signals. AI Data Licensing: The Shift to Real-Time Access | Pebblous +2
That puts two rows on your sheet where there used to be a line item for lists. As a buyer, the external data you feed your AI systems for enrichment, benchmarking and grounding now costs more, carries provenance conditions, and needs a record of where it came from. As an owner, your data is an asset with a rights question attached to it. Every renewal above an illustrative $25,000 should answer three questions before it is signed: does the vendor train on our data, may they resell what they derive from it, and what do we keep if we leave. The AI dividend goes to whoever holds those rights. And data readiness, meaning lineage, quality and access control, is the precondition for any production AI you will fund, so it becomes the first draw on that fund rather than a separate wish. If your industry sits in a Rollup shape, where acquirers win by unifying data across the companies they buy, this row is the entire investment thesis.
Software stopped being a fixed cost
The seat is giving way to the meter. Vendors are folding AI into platforms as included credits with overage charges, and the examples are now mundane: Atlassian gives each user a small monthly bundle of AI credits and charges per conversation past it; HubSpot meters AI credits above what a tier includes; Zendesk charges for each conversation its AI agent resolves; and Salesforce changed how it priced the same product three times in about a year and a half. The result is what you would expect. About half of CIOs in Gartner’s 2027 survey expect AI to raise total cost of ownership, even as a larger share are being pushed to show AI-driven savings, and a large majority of IT leaders have already been surprised by a bill from usage-based or AI pricing. Info-Tech put the longer arc plainly: today’s AI prices are artificially low, and consolidation will hand vendors the pricing power to correct that with a shock. 2026 SaaS Pricing Trends Driving Up Enterprise Costs +3
We describe this mechanism as the pricing-unit switch. The productivity gain from AI is real, and the vendor reprices it before it reaches your P&L, by changing what you are charged for. A budget that carries software as one fixed number cannot see this happening. Carry it as two rows instead: committed and variable. Model the variable row as a range with a cap and a named owner who can turn credits off, not a point estimate. For an illustrative 300-person company spending $900,000 on software, if forty percent of that now carries a variable component, a planning figure of $860,000 to $1.05 million with stated triggers is honest and $945,000 flat is not. Then put the renewal calendar on the budget page itself, because vendors are still testing prices and giving ground they will stop giving once the market settles: ratchet clauses that lower seat commitments as agents absorb work, agent seats priced apart from human seats, and consumption caps written into the contract rather than hoped for. AI Magicx
Hardware got scarce again
For a decade the hardware row was the boring one. Not this year. Memory makers moved wafer capacity to the high-bandwidth and DDR5 parts that AI data centers consume, and ordinary servers, laptops and network gear got what was left. Spring contract prices for conventional DRAM were projected to jump by roughly three-fifths in a single quarter, then keep climbing into the third quarter at a slower pace, and SK Hynix’s chief executive said in July that 2027 will be the industry’s worst supply year. Gartner has suggested enterprise PC budgets may need to rise eight to twelve percent just to hold unit counts, or companies will have to run what they own for longer. AI Memory Shortage 2026: What IT Leaders Need to Know +4
The refresh cycle you inherited does not price the same anymore, and the mistake is to drift rather than decide. Anything you will need inside a year, buy now; nothing in the supply picture points to cheaper memory in that window. Anything that can run two more years on maintenance, keep running, and put the money where it returns sooner. Cloud only delays the cost, since providers buy memory on long contracts and pass the increase along later. For an illustrative 250-laptop refresh that priced at $300,000 last fall and now comes in at $340,000 to $360,000, the question is not whether to pay it. It is which 100 machines can wait a year, and what the maintenance line looks like while they do. AzterionAzterion
The people row changed shape
It did not get smaller. It got rearranged. Roles built on throughput shrink; roles built on oversight, judgment and integration grow; and the one thing regulators, insurers and boards all ask for is a named person accountable for AI. Forrester’s 2027 planning guides find most leaders expecting bigger budgets, with a warning that more AI money without stronger data foundations, governance and operating models produces duplication, fragmentation and technical debt rather than results. Gartner’s counsel is to redesign the work before adding people, and to train for business acumen, critical thinking and judgment. FutureIOTGartner
The mid-market version is simpler than the enterprise version. A CIO-centric model, where the CIO owns AI and its governance, holds up when the CIO reports to the CEO and controls the budget; a mid-sized manufacturer that already has a CIO does not need to add a chief AI officer on top. If you do not have that seat, the accountable-leader row can be fractional; it cannot be empty. Alongside it: a training line sized to the whole company rather than to IT, and a pipeline from citizen AI to production AI with a promotion rule attached. What your employees built in a browser this year either gets an owner, a control, and a budget, or it gets retired. For an illustrative $40 million distributor with a dozen employee-built workflows touching customer data, the honest exercise promotes three, retires six, and sandboxes three. The promotion budget is the real AI budget. The license count was never it. Digital Chiefs
Security followed the agents
The security row moved for a reason that has nothing to do with vendors. Agents hold credentials. They act inside your systems of record, they follow instructions, and instructions can be forged. Non-human identities are multiplying faster than the human ones, and the fraud that used to arrive as a misspelled email now arrives as a voice on a phone or a face on a video call asking your controller to approve a wire. None of that needs a statistic. It needs three additions to a row you already have: identity and access management that treats an agent as a principal with least privilege and an expiry; monitoring of what agents actually do, not only what they were told; and an incident runbook that covers AI-specific failures, the model that changed under you, the agent that acted on a spoofed instruction, the workflow that quietly drifted. A fourth addition costs nothing but a rule: no wire, no vendor bank change, and no payroll change on a single channel, however convincing the voice.
The method changed, not only the rows
The largest change is not a row. It is that roll-forward budgeting assumes a stable portfolio, and the portfolio is no longer stable. Under AI, things get promoted, retired and repriced inside a single year. Gartner’s 2027 advice to CIOs runs to containment: curate what people can buy, kill the AI projects nobody uses, and demand proof inside ninety days. Our version is a gate. Production AI gets funded through an economic decision point, in tranches, against a benefit claim that a named executive signs, and the next tranche depends on the last one showing up in the ledger. Gartner
That gate sorts the whole sheet into three kinds of rows. Stabilize rows, meaning security, hardware and base platforms, get repriced this year and defended. Optimize rows, meaning the variable software line, vendor discipline and consumption governance, get owners. Monetize rows, meaning production AI with revenue attribution and data treated as an asset, get gates. Where the dollars concentrate depends on the shape of your industry. In a Rising Bar sector the recovered dividend gets spent beyond table stakes, on the wedge that wins share. In a Slow Melt sector the money goes to efficiency and vendor discipline, because margin is the prize. In a Rollup sector it goes to data unification. In a Falling Floor sector it goes to autonomy, because the moat is the point. The 2026 AI Dividend Map places your industry; the budget should agree with it.
What a turned budget looks like
Take the illustrative $60 million services company again, with a $2.4 million technology budget spread across six familiar rows: infrastructure, applications, security, staff, projects, and a small line labeled AI pilots. The roll-forward at six percent produces $2.54 million and buys last year’s company at this year’s prices.
The turned version has nine rows and totals somewhere between $2.7 million and $2.85 million. Infrastructure is up for one year on the hardware price and flat after. Applications are split into $600,000 committed and a $150,000 to $260,000 variable band with a cap and an owner. Security carries about $60,000 more for agent identity and monitoring. Staff is flat in dollars and different in roles, with a $40,000 training line beside it. Governance is a row of its own. Data readiness is a row of its own, and the first draw on the production AI fund, which is $250,000 released in three tranches through the gate. The last row is the one most budgets have never had: a monetize line, holding the data-rights review across every renewal and at least one AI initiative measured by revenue attribution rather than cost avoidance. The increase over the roll-forward is roughly $160,000 to $310,000. What it buys is the difference between a company that adopted AI and a company that owns what AI produced.
The call
Our dated call, September 2026: by the fall 2027 planning cycle, most mid-market technology budgets will carry governance and data as standing rows and software as a committed-plus-variable pair. The budgets that do not will belong to the companies whose AI spending rose fastest and whose EBITDA did not, which is exactly the outcome the pricing-unit switch predicts. We will score that call in next year’s Innovation Vista Report.
AI cannot lead itself, and a budget cannot turn itself. Someone has to decide which rows go variable, which experiments get promoted, who owns the inventory, and what the company’s data is worth to someone else. If that person is not on your payroll, that is a row too, and it is a smaller one than the mistake it prevents.
