Start with AI now or wait 6 months? The right question is a different one

Your committee has had this on the agenda for six months. Every time it comes up, someone says "let's wait for the technology to settle" and everyone nods. And it comes up again the next month. It's the most deferred decision in leadership committees so far in 2026, and deferring it isn't free.

What follows is an honest conversation with no propaganda: four situations where waiting is the right decision, four where it isn't, and the single question that closes the loop.

Why adopting AI is the most deferred decision in leadership committees in 2026

The conversation repeats on an almost identical script in dozens of leadership committees. The CFO asks how much it costs. The COO asks what problem it solves. The CIO says more clarity is still needed. Someone mentions the AI Act. Another person recalls a competitor's case. And the conclusion, month after month, is that it's best to keep observing.

The fatigue is real and the arguments for waiting have merit. The technology is changing fast. European regulation applies in full on August 2, 2026 with consequences still without precedent. Vendors are restructuring — several announced headcount adjustments in April simultaneously with their AI growth metrics. Whoever decides to wait isn't being lazy. They're reacting to a noisy market with a reasonable response.

The problem isn't waiting. It's waiting without having defined what would have to change to stop waiting. That's the question most committees aren't asking themselves, and it's the only one that closes the loop.

The three legitimate reasons to wait

It's worth starting by acknowledging that waiting can be the right answer. There are three weighty reasons we see repeated that, in their strong form, justify deferring.

1. Regulatory reason: high risk with no closed legal analysis

If your company's activity falls within the high-risk catalog of the European Artificial Intelligence Regulation — consumer credit banking, hiring with automated screening, medical devices, critical infrastructure, migration management, justice, education with automated assessment — and your legal team still hasn't finished the analysis of the impact that has on your processes, waiting three to four months to know what applies to you is prudent, not procrastination.

2. Operational reason: data not ready for production

If your data isn't in a state to be used — six versions of the same table across different systems, data governance with no clear owner, integrations done by hand every month — the first artificial intelligence project on dirty data is a predictable failure that's expensive to reverse. Starting before having a minimum of order is building on sand.

3. Organizational reason: nobody with the judgment to lead

If nobody in your organization has enough judgment to lead a serious conversation with an AI provider — you don't need a technical degree, you need the judgment to tell a promise from a delivery — starting means delegating the direction to the provider. The provider ends up deciding what to build, how to measure it and when to declare it a success. It's the worst possible configuration.

These three reasons, in their strong form, are legitimate. What isn't legitimate is invoking weak versions — "we don't have perfect data", "we want to wait for the definitive version", "we'd rather someone did it first" — as cover for a decision that in reality isn't being made.

The three legitimate reasons to start now

On the other side there are also three weighty reasons, none of them an urgency invented by a vendor.

1. Organizational learning curve

A company that has completed a serious artificial intelligence project — even a small one, even one with stumbles — has something the one that has only read about the topic doesn't: a trained committee, a common vocabulary, calibrated evaluation criteria, a baseline of how much it costs and how much it returns. That learning isn't bought; it accumulates. Every month you wait, the organization of the competitor who did decide is learning something you'll have to learn later, probably in a hurry.

2. Competitive window

If your market already has a visible competitor with a customer-facing artificial intelligence case — chatbot, recommender, automated quoting, conversational service — the window to enter in parallel is counted in quarters, not years. The difference between arriving third and arriving seventh is structural; the difference between arriving second and arriving third is usually anecdotal. If the window is open, waiting closes it.

3. Talent

Professionals with real artificial intelligence experience — not certifications, measurable operational experience — choose active companies. A company that defers the AI decision for a year has an implicit message for any candidate with judgment: we don't do it here. The fight for talent isn't won with a salary package; it's won with a recognizable project. Waiting shrinks the pool of who could lead the first serious project.

The empirical record: what the companies that started on time report

There's a relevant empirical asymmetry in this conversation, and it's worth facing head-on. The evidence published so far in 2026 materially favors those who decided to start on time and well — especially when the first project was applied to internal corporate processes, where vendor noise is lower and returns are measured with the company's own accounting.

The concrete cases published this quarter point in the same direction and are starting to have enough volume to not be anecdotal. A large European integrator reported in its first-quarter results that 11% of its total signed bookings already comes from agent-based projects — the first time a hard financial metric for agentic work is reported to an investor. An American integrator communicated the same week that the cost of its internal artificial intelligence transformation program — applied first to its 350,000 employees before selling it externally — returns between 200 and 300 million dollars a year in measured operational efficiency.

A global boutique published a legacy banking code modernization case in which a process that historically took three months now closes in four hours, and a pilot in a different industry where automating 1% of the warranty claims flow returns approximately one million dollars a year. Another firm reported administrative process automation projects where closing time drops from months to weeks and the percentage of tasks the system resolves without human intervention reaches 70%.

These cases aren't marketing material. They're figures published in investor results reports, official announcements and specialized press coverage — verifiable if your team wants to read them.

The conclusion a prudent committee should draw isn't "everyone is winning with artificial intelligence": the same Bain report notes that most are still in pilot with no clear return, and the consultancy Gartner anticipates that more than 40% of agent-based projects will be cancelled before the end of 2027 due to escalating costs or diffuse business value. The correct conclusion is more nuanced and more useful: organizations that start well and on time capture structural operational advantages; those that start badly or late pay the cost without the return. The four conditions we detail below are precisely the ones that separate starting well from starting badly.

The four contexts where waiting is the right decision

Combining the three reasons to wait with each company's reality, there are four contexts where the answer is objectively "not yet".

The four contexts where starting now is the right decision

By symmetry, there are four contexts where the answer is "now".

The two traps that invalidate the decision

Trap 1: waiting with no review date or concrete trigger

The most frequent trap when you decide to wait is not setting a review date or a signal that activates the decision. The conversation is left as "we'll pick it up when the time comes" and renews itself every month. After half a year the situation is identical. Waiting is legitimate only when it comes with a date — three months, six months — and with a concrete trigger: "if a competitor appears with a visible case, if an important customer asks, if legal closes the AI Act analysis, we start before the date".

Trap 2: starting with no pilot exit metric

The symmetric trap when you decide to start is not setting an exit metric. The pilot is approved with enthusiasm, executed with reasonable difficulties, and keeps getting extended. After a year there's a perpetual pilot nobody wants to kill and that produces no clear result. Starting is legitimate only when it comes with a small pilot, a concrete metric — reduction in time to production, improvement in response time, ratio of cost to verified savings — and an evaluation date.

The question that closes the loop

The honest conversation your committee needs to have isn't "wait or start?". It's:

If the answer is "nothing concrete, we'll look at it next quarter", the decision isn't being made. What's being taken is the default option, which is to wait another month. And next month it'll come up on the agenda again. And it'll be deferred again.

If the answer is "we wait until June because legal closes the AI Act analysis then, and we start earlier only if our biggest competitor launches something public first", the decision is made. Future uncertainty is bounded. The committee is free to take the item off the agenda until one of the two things happens.

That's the question most committees are skipping so far this year. It isn't a question of artificial intelligence. It's a question of how your committee decides.