Move Now or Wait for AI to Mature? The Timing Question SMB Leaders Keep Getting Wrong

By Victor Hoang · Updated Oct 10, 2026

Every owner and executive of a 50 to 500 person company eventually asks the same thing: should we start now, or wait until AI settles down? The tools change every quarter, the vendors contradict each other, and nobody wants to be the company that bought the wrong thing early.

Waiting looks like the careful choice. Our view is that it's the riskier one, but only for part of what you might do. The tools will keep improving. What won't improve on its own is your organization's ability to use them: which tasks fit, where output can't be trusted, how a workflow should change, who on your team has the instinct for it. That's learning, and you can't buy it later at any price. So move early on low-risk workflows, and wait only on heavy, hard-to-reverse platform bets.

Key Facts

• In a study of 5,179 customer support agents, a generative AI assistant raised issues resolved per hour by about 14% on average, with a 34% improvement for novice and low-skilled workers and minimal impact for experienced, highly skilled ones.
• The same paper reports suggestive evidence that the tool spread the practices of more able workers and helped newer workers move down the experience curve faster.
• Economists describe a "productivity J-curve": general purpose technologies require large complementary investments in processes, business models and worker skills, so measured gains lag in the early years and arrive later.
• In an experiment with 758 BCG consultants, AI users did better on tasks inside AI's capability frontier (12.2% more tasks completed, 25.1% faster) but were 19 percentage points less likely to get a correct answer on a task outside it.
• Waiting doesn't remove the risk of using AI badly. It only means employees learn on their own, without guardrails (see the governance gap).

The wrong question

"Has AI matured yet?" has no stable answer. There's always a next model, a cheaper tier, a new category of agent. If maturity is your trigger, you'll be waiting in 2028 too, and you'll be waiting with a team that has spent two more years not learning.

The better question is narrower: for this specific workflow, what's the cost of being wrong, and what do we learn by trying? Some workflows are cheap to try and cheap to undo. Others are expensive to commit to and painful to unwind. Treating "AI" as one decision lumps those together and produces either reckless spending or total paralysis.

Three things hide inside "AI maturity":

1. Model quality. Improves continuously and is available to everyone at the same time. Waiting buys you nothing here that you wouldn't get by starting now, because a workflow you've already built can switch to a better model. We cover how to design for that in an AI strategy that survives model churn.
2. Vendor and platform stability. Genuinely uncertain. Products get acquired, repriced, or shut down. This is where patience pays.
3. Your organization's readiness. Not improved by waiting at all. It only improves through use.

Waiting helps with the second, does nothing for the first, and actively hurts the third.

Why learning is the asset you can't buy later

When a technology changes how work gets done, the tool is the cheap part. The expensive part is everything around it. The productivity J-curve research by Brynjolfsson, Rock and Syverson makes this point at the level of the whole economy: general purpose technologies require significant complementary investments, much of it intangible (new processes, business models, worker skills), and that's why measured productivity gains tend to arrive well after the technology does. The authors also note that the effects of AI-related intangible capital on measured productivity are small but growing.

For a smaller company, translate that into plain terms. Your team has to learn what to ask, what to check, and what to leave alone. Your managers have to learn to redesign a handoff instead of just bolting a tool onto the old one. Your documentation has to get good enough that a model can use it. None of that arrives in a vendor's box, and none of it transfers from a competitor who figured it out first.

The customer support study is a useful picture of how this learning shows up. In that study of 5,179 agents, productivity rose about 14% on average, but the gains were uneven: 34% for novices and low-skilled workers, with minimal impact for experienced ones. The authors suggest the assistant disseminates the practices of more able workers and may help newer workers move down the experience curve faster. The point for a decision maker isn't the headline number, which came from one company's support operation and won't carry over to yours. It's that the benefit depended on how the tool was folded into a real workflow with real people, and that kind of fit is something you learn by doing.

There's a second reason learning compounds: it's hard to see the edges of what AI can do from the outside. The BCG experiment found that consultants using AI performed better on tasks inside the frontier and worse on one task placed outside it, by 19 percentage points on correctness. The authors describe the frontier as jagged: tasks that look similar in difficulty can fall on opposite sides of it. People learn where the edge is by working near it, and by catching the tool being confidently wrong a few times at low stakes. A team that has never had that experience will meet its first confident error on something that matters.

What waiting actually costs

Waiting isn't free, and the cost doesn't show up on a budget line, which is why it gets ignored.

Shadow adoption fills the vacuum. Your people are not waiting. In most companies of this size, some of them are probably already using consumer tools on their own, possibly with customer data, under no rules. A decision to wait is, in practice, a decision to let that happen without guidance. If you're worried about risk, this is the largest source of it. Our piece on the governance gap walks through how that plays out, our piece on employees building their own apps with AI argues for a light registry over a ban, and the library guide on AI security and compliance covers what to put in place first.

You lose the baseline. If you start measuring a few workflows now, in a year you'll know what AI changed. If you wait, you'll be comparing against a memory. That matters because measuring AI ROI beyond "time saved" is hard enough with good before-and-after data.

You postpone the boring work. Cleaning up process documentation, defining who approves what, agreeing on what "good output" looks like: this is the work that makes any later AI investment pay off, and it takes calendar time. Starting it late delays everything downstream.

Hiring and retention drift. A company that has said "not yet" for two years is making a statement about its operating pace to candidates and current staff, whether it means to or not. We look at the related hiring version of this reflex in wait-and-see and the hiring freeze.

When waiting is right

None of this is an argument for buying everything now. There are real reasons to hold back, and they cluster around one trait: the move is expensive to reverse.

Wait, or at least go slowly, when:

• The commitment is long. Multi-year contracts, annual prepayment, or deep integration into your core systems of record.
• Switching means rebuilding. If your data, prompts, or automations only live inside one vendor's proprietary format, you're taking on lock-in.
• The category is still being defined. If vendors in a space can't agree on what the product is, the winners probably haven't emerged.
• The task is high stakes and hard to check. Legal commitments, financial reporting, hiring decisions, and anything customer-facing where an unreviewed error is costly. These need a human in the loop and a track record before you extend trust, which is another argument for building experience on safer tasks first. The difference between assistance and autonomy is covered in AI copilots vs. agents.

Notice what's not on the list: "the models might get better." A running workflow benefits when they do.

A decision table: move now or wait, by workflow type

Use this to sort the specific things on your list, not AI as a whole.

A rule of thumb that covers most of the table: start where a mistake costs you an afternoon, and wait where a mistake costs you a quarter.

Worked example (illustrative)

Take a hypothetical 120-person services company. The COO is choosing between two moves.

Option A: give the account management and delivery teams approved AI tools for meeting summaries, status-report drafts, and internal research, with a one-page rule on what data can go in. Cost is small, the work is reviewed by a person every time, and it can be switched off in a day.

Option B: sign a three-year contract for an AI-native platform that would replace the company's project management and billing tools.

Following the table, A starts this month. B waits. Six months in, the company has a team that knows where the tools are reliable, a short list of workflows where the time savings are measurable, documentation that's cleaner than before, and a clearer picture of what it would actually want from a platform. When the platform decision comes, they're a far better buyer than they would have been if they'd spent those six months waiting. The money saved by delaying B was never the point. The time spent learning on A was.

What to do next

You don't need a program. You need a short list and a calendar.

1. List ten candidate workflows. Ask team leads for tasks that are repetitive, text-heavy, and reviewed by a person already. Don't start with the most exciting idea. Start with the most reversible.
2. Score each one with the table above. Two questions: how hard is it to undo, and what does an error cost? Anything cheap on both is a yes.
3. Pick two or three and name an owner for each. Someone accountable for how it's used, what's measured, and when it's reviewed.
4. Write the one-page rule. What data can and can't go into AI tools, and who to ask. Keep it short enough to be read.
5. Record a baseline before you start. Time per task, rework rate, a quality check by a manager. Even rough numbers beat none. The ROI measurement piece covers what to track beyond hours saved.
6. Set a 90-day review and a standing "stop" rule. If a workflow isn't earning its place, drop it. A cheap experiment you can end is the whole advantage of starting small.
7. Hold off on the heavy bets and write down your conditions. For each, note what would make you say yes: price stability, a clean exit path, a track record. Revisit when the conditions are met, not on a date.

If your bottleneck is coordination rather than tools, coordination cost in operations is a useful companion read.

Learn More

• Build an AI Strategy That Survives Model Churn
• Measuring AI ROI Beyond 'Time Saved'
• The Governance Gap: What Leaders Get Wrong About AI at Work
• AI Copilots vs. AI Agents