Last Year: Invest in AI. This Year: Cut Costs.

Last year: invest in AI. This year: cut costs.

Gartner says tech budget growth is at a decade low. Down 6% from last year. If you run engineering at any size, you felt this already. The spreadsheet conversations started earlier this year. The tone changed.

Twelve months ago, every board deck I saw had the same slide. Some version of "we must invest in AI or get left behind." Money moved fast. New teams got hired. GPU clusters got bought. Proof of concepts got funded with almost no scrutiny.

Now the same boards want to know where the return is. They want cost reduction. They want both, actually. Keep the AI momentum but also cut 15% from the tech budget. These two things do not fit together, and pretending they do will get you into trouble.

I have been on both sides of this conversation. I ran a $5M consulting P&L and managed $50M in annual tech investment at a global logistics company that operated in 130 countries. I have sat through budget cycles where the strategy reversed in a single quarter.

The answer is not "do both"

When your CFO says "invest in AI and cut costs," the temptation is to spread the remaining budget thin. Fund eight AI projects at 60% each. Keep everything alive. Show progress on all fronts.

This is the worst thing you can do.

At the logistics company, we went through exactly this. The board wanted AI-driven route optimisation and predictive maintenance and customer service automation and demand forecasting. All at once. All with a shrinking budget.

I pushed back. We picked three bets. Three projects that had a clear line to either revenue or cost savings within 12 months. We funded those properly. Everything else went on a "not now" list. Not a "never" list. A "not now" list.

The three we picked: predictive container maintenance (because we were losing $8M a year in unplanned downtime), automated customs documentation (because it was pure labour cost we could measure), and demand forecasting for capacity planning (because it directly affected pricing). Each one had a number attached to it before we started.

We cut six other AI projects that were interesting but had no clear payback timeline. The data lake modernisation got paused. The internal chatbot got paused. The computer vision project for warehouse inspection got paused.

Two of our three bets paid off within nine months. The third took 14 months but eventually delivered. If we had spread that money across nine projects, I think none of them would have delivered anything usable.

FinOps is not optional any more

Most engineering teams I talk to still do not know what their AI workloads actually cost. They know the monthly cloud bill. They do not know the cost per inference call. They do not know what a single model retraining run costs. They cannot tell you the cost per decision their ML pipeline makes.

This matters because you cannot defend investment you cannot measure.

Three practices that worked for us.

Per-decision cost tracking

Every AI model we deployed had a cost tag. Not just compute cost. The full cost: data prep, training, inference, monitoring, and the engineering time to keep it running. When someone asked "is the demand forecasting model worth it," I could say "it costs us $14,000 a month to run and saves us $180,000 a month in better capacity allocation." That conversation is very different from "we think it is probably saving us money."

GPU spend guardrails

We set hard limits on GPU spend per team per month. Not soft limits. Hard limits. If your training run was going to exceed the budget, you had to make a case before starting it, not after. This sounds harsh. It saved us from two runaway experiments that would have cost six figures each.

Pre-deployment cost gates

Before any model went to production, it went through a cost review. What does it cost to run at current scale? What does it cost at 10x scale? Is there a cheaper architecture that gets 90% of the accuracy? We killed two models at this stage that worked well but were too expensive to run. One of them we redesigned with a simpler approach that cost a fifth as much.

Your "not doing" list matters more than your roadmap

Something most CTOs get wrong with boards. They present what they are building. They never present what they chose not to build.

Boards are not stupid. They read the same articles you do. They know about agents and copilots and all the rest. If you do not tell them what you decided against, they will wonder why you are not doing those things. They will assume you have not thought about it.

I started bringing a "not doing" list to every board meeting. One page. Three columns: what we are not investing in, why, and what would have to change for us to reconsider.

For example: "We are not building an internal AI coding assistant. Reason: off the shelf tools are good enough and improving fast. The build cost is $400K and the advantage over commercial tools is marginal. We will reconsider if commercial tool pricing exceeds $X per seat or if we need custom model behaviour for our proprietary codebase."

This did two things. It showed the board I had considered these options seriously. And it stopped the "why aren't we doing X" questions that eat up half of every board meeting.

The argument that saved our AI platform budget

In the second round of cuts, the CFO wanted to reduce the AI platform team. Not the projects. The platform underneath them. The shared infrastructure, the ML ops tooling, the data pipelines that fed every model.

This is the one thing you cannot cut. Cut projects, fine. Cut the platform and you lose the ability to run any of them.

The argument I made was simple. I showed the cost of rebuilding platform capability for each new project versus the cost of maintaining it centrally. Without the platform, every project team would build their own data pipelines, their own model serving infrastructure, their own monitoring. The duplication cost was roughly 3x the platform team cost.

I also showed what happened at two other companies in our industry that cut their platform teams during the 2020 downturn. Both spent 18 months rebuilding what they had dismantled, at higher cost, with worse results.

The board kept the platform team. They cut it by two headcount instead of eight. I could live with that.

So what do you do on Monday morning

Get your actual AI costs. Not the cloud bill. The real, full cost per workload. If you do not have this, nothing else matters because you are making decisions blind.

Then rank every AI initiative by payback period. Not by how interesting it is. Not by how much the team loves working on it. By when it will return more money than it costs. Be honest about the ones where you cannot estimate payback. That gap tells you something.

Pick your top three. Fund them fully. Pause the rest. Tell the board what you paused and why. Build that "not doing" list and bring it to the next meeting.

The boards have not changed their minds about AI. They still believe it matters. They just want to see that you can be trusted with the money. Show them the numbers. Show them the discipline. And show them you know when to say no.