In May 2026, reports said Microsoft had cancelled most of its internal Claude Code licenses. The rollout was only six months old, and cost was part of the reason. Separate reports said Uber had used its full 2026 budget for the same kind of tool in four months. Both stories spread fast, and CTOs everywhere asked the same question. Could this happen at your company?

Neither company lacks technical skill. Their budgets were set in fall 2025, before agentic tools pushed usage far above plan. A sound artificial intelligence strategy expects that kind of surprise. It sets limits, owners, and review dates before the invoices arrive. That’s the work a CTO can start this quarter.

This guide shows CTOs how to rank AI bets, fund them in stages, and govern them without slowing delivery. It draws on 2025 and 2026 cases from Microsoft, Uber, and Deloitte. It also uses recent survey data from BCG and CIO.com.

Each section ends with a check a team can run in a week. An AI investment strategy that survives a budget review needs that kind of evidence.

Where AI Money Goes Wrong

Most failed projects don’t start with bad models. They start with a plan that missed one of two things, which are cost and quality.

• Budgets Written Before Usage Arrived

Uber’s COO told Business Insider that AI token spending was getting harder to justify. More usage wasn’t producing proportionally more useful features. That’s a pointed comment from a company that had encouraged the usage.

Agentic tools explain part of the problem. They run many steps and repeat queries, so one task can burn far more tokens than a chat. A budget built on chatbot habits breaks quickly. Who at your company watches usage by team today?

• Deloitte’s Report and the Missing Review Step

In October 2025, Deloitte Australia agreed to repay part of its fee. The report covered the government’s welfare compliance system. The contract was worth about AU$440,000. A researcher found fabricated citations and other errors linked to a generative model. Deloitte revised the report and repaid close to AU $98,000.

The firm employs experienced reviewers. The lapse points to a process with no clear check for machine-written references. One simple rule, that a person opens every citation, would test for that.

• What the Two Cases Share

Neither case turns on model capability. In both, they approved the use without a limit or a review step. That’s a planning failure, and it costs the same whether the tool is large or small.

The CIO.com State of the CIO survey for 2026 points the same way. Only 19% of IT leaders said their AI initiatives had met or exceeded business goals. And 32% blamed ill-defined ROI metrics for slowing scale. Plans without numbers can’t earn trust.

• Gartner’s forecast adds a warning. In June 2025, it said over 40% of agentic AI projects would be cancelled by 2027. It named rising costs, unclear value, and weak risk controls as the causes. Those map to three checks in this guide. They are a cost meter, a value number, and a risk tier.

Choosing Where to Invest First

Good AI prioritization begins with fewer ideas. Teams that fund everything end up funding nothing well.

Fewer Bets, Chosen Harder

BCG’s September 2026 spending pulse surveyed 423 senior IT buyers. It found that 66% planned to raise AI and machine learning spending. The same report said CIOs and CTOs were moving from broad experimentation to fewer, higher-impact use cases. The average number of use cases per company fell at every maturity level.

That pattern fits what the failures show. Money is rising, but patience for loose projects is falling. Boards now ask why each use case exists, and a ranked list answers that question. A CTO who can explain why project four ranks below project two earns trust. A longer list earns less. Each rank needs a reason in plain language. Finance teams read those reasons first.

A Four-Question Scoring Table

A scoring sheet with eight criteria usually gets gamed. Four questions are harder to fake.

Disagreements matter more than the totals. Say a CTO rates data readiness at five and the data lead says two. That conversation is the real project.

Fund the Next Stage, Not the Whole Project

The strongest enterprise AI strategy releases money in steps. Each step has a question to answer and a number to hit. If the number is missed, the money stops.

• Stage one tests whether the data can support the task.
• Stage two tests the output on real cases with real users.
• Stage three tests running cost at expected volume.
• Stage four approves the rollout.

Uber’s budget shows why stage three matters. A pilot with ten users says little about a thousand. So the third stage should price usage before launch and again at one month. The second price check matters most, since real users behave differently from test users. A few heavy users may account for most of the spend. A cap, or a cheaper model for simple tasks, can bring the bill down.

Building AI Governance That Doesn’t Slow Delivery

Strong AI governance isn’t a committee that says no. It’s a short set of rules that lets low-risk work move fast.

Match the Review to the Risk

CIO.com found that 83% of surveyed organizations had, or planned within a year, a cross-functional steering group for AI. Only 53% had an official approval process for AI projects. A group without a process becomes a talking shop.

A tiered path fixes that. Low-risk uses get a light check. High-risk uses get a full review. The tiers also tell teams what to expect. A product team knows a meeting-notes tool needs only an owner and data rules. It can ship in days, while a claims tool takes weeks. Clear tiers remove the waiting and the guessing. They also stop reviewers from spending equal time on unequal risks.

A report drafted with AI help fits the medium tier. That tier asks for a quality sample, and a sample of citations tests exactly what failed at Deloitte.

Keep an Eye on Regulation Dates

The EU’s Digital Omnibus package got final Council approval on June 29, 2026. It moves the deadline for stand-alone high-risk systems from August 2026 to December 2027. Embedded high-risk systems move to August 2028.

Transparency duties under Article 50 did not move. They applied from August 2, 2026. One exception exists. Providers of generative AI systems already on the market before that date have until December 2, 2026, to add machine-readable marking.

Companies that sell in Europe still need to know which of their tools carry which duty. A delay isn’t an excuse to wait. Regulators ask for testing records and human oversight rules. Those same documents help a CTO defend a budget.

Put a Cost Meter on Every Use Case

BCG reported that high-maturity firms see AI and token costs down to the team or user level. They use that view to allow broader use. Visibility builds confidence. A cap set per team stops one enthusiastic group from spending a quarter’s budget in a month. Three controls cover most of the need.

• A monthly usage cap for each team
• An alert when a team reaches 80% of its cap
• A finance review of cost per outcome every month

Running Delivery After Approval

Approval is the start of the work. Owners, old systems, and honest measures decide what happens next.

• Legacy Systems Cancel More Projects Than Models Do

A 2026 study by GFT Technologies and Wakefield Research reached 945 CIOs and CTOs. They worked at large firms in 19 countries. It found that 84% said legacy system limits had led to a cancelled AI pilot. So the readiness check belongs before approval. A short list of questions covers most of it.

• Can the tool read the data it needs in a usable format?
• Can it write results back to the system of record?
• Who maintains the connection when a source system changes?
• What happens to the workflow if the tool goes offline?

Some CTOs bring in AI strategy services at this point, mostly to map integrations before a vendor is picked. Others handle it in-house with a two-week architecture review.

• One Owner and One Number per Project

Every approved project needs a business owner who sees the bill. It also needs one number that proves value, such as cost per resolved case. Staff time saved sounds good but is hard to bank. CIO.com found that operational efficiency led the metrics list at 40%, ahead of revenue growth at 27%.

• Measure Honestly, and Report Misses

A monthly report should show misses beside wins. If a tool saves time in one team and adds review work in another, the report says so. Finance leaders trust a CTO who shows the bad month. They discount one who shows only the best.

A useful report fits on one page. It lists cost per outcome, quality sample results, and the status of each stop line. If one item is missing, the missing item is the finding.

• Treat Big Programs as a Series of Small Changes

Large AI transformation programs fail when they try to change everything at once. Teams lose track of what worked, and sponsors lose patience. Small changes in one workflow teach a team more. A claims team adopts one tool, with new roles and new checks. That team then becomes a working example for the next one.

The first 90 days can look like this.

• Month one covers the idea list, scoring, and owner names.
• Month two covers data tests and the first stage gate.
• Month three covers the cost meter and the governance tiers.

Teams that want an outside view often use AI business consulting for the first scoring round. An outsider has no stake in any pet project.

What to Do Before the Next Budget Cycle

Microsoft and Uber showed that cost surprises reach even careful companies. Deloitte showed that quality lapses do too. A plan with stages, tiers, and a cost meter catches both early. Together they form a practical AI investment strategy that a board can read.

Open your current AI portfolio and ask three things, using one page and a pen. Which projects have a named owner? Which have a number that proves value? Which have a stop line written down? Any project that fails all three is a candidate for pause.

An outside review can help here. Good AI strategy consulting reviews the portfolio like a skeptical finance lead, with the same blunt questions. It ends with a ranked list and a stage plan. A CTO who walks into the budget meeting with both will face easier questions. Which of your current projects would you fund again today? Start the review this month, while the next budget is still a draft. Waiting for the invoices to arrive costs far more than a week of questions.

Book a working session to review your AI portfolio before the next budget meeting.