// Resources
Should We Wait for AI to Get Better Before We Invest?
What actually gets cheaper if you wait and what does not, how many businesses your size are using AI today, what starting now and being wrong really costs, the four moves that hold their value whatever you buy later, and the four situations where waiting is the right call.
By Victor Xu, Founder · LinkedIn
Last updated · 2026-10-27
The short version
Wait on the tools and start on everything else. Waiting is a genuinely good bet on anything priced per use or shipped as a feature, because that side of the market gets cheaper and better on a timescale no business can match: Stanford's 2025 AI Index found the inference cost for a system performing at the level of GPT-3.5 dropped more than 280 times between November 2022 and October 2024. What has not got cheaper is knowing which of your processes is worth changing, having data clean enough to use, and having a team that trusts the output. Those take months whatever the technology does, and they are what decides whether the cheaper tool is useful to you the day it arrives. So the answer for most businesses is to spend the next quarter on the parts that do not depreciate, buy as little software as possible while doing it, and make the purchase decisions later, when they are cheaper and clearer. Waiting on the whole question is the one version that leaves you with nothing.
Should we wait for AI to get better before we invest?
Waiting makes sense for the purchase and not for the preparation, which is why this is two decisions rather than one. The software gets better and cheaper while you wait. Your own readiness does not improve by itself at all.
Split the spend into three kinds. There is money you give a vendor for a tool or a per-use service, which is where waiting pays, because that market reprices every few months. There is money you spend working out what should change in the business, which does not get cheaper and is the input to every later decision. And there is money you spend on the plumbing, cleaning up records and connecting systems that should have been connected anyway, which you would want done whether or not any AI followed.
Only the first of those three is a bet on the technology. The other two are bets on your own business, and they hold their value if the tools change, the vendor disappears or the model you planned around is retired.
The practical version of that answer is a quarter of work that costs very little: find out what is already in use, measure one expensive bottleneck properly, and write down what a fix would have to do. Then buy, with the prices and the products of that month rather than this one.
What that gets you is the position where a cheaper tool is useful on day one instead of starting a six month discovery. What an AI roadmap is describes the artifact that comes out of it, and how long AI automation takes answers the separate question of duration once a build starts.
What gets cheaper if we wait, and what does not?
What gets cheaper is compute, capability and packaging. What does not is anything that lives inside your business. The table is the whole argument on this page.
| What you could wait for | Does waiting help? | Why |
|---|---|---|
| Cost per use of a model | Yes, substantially | Stanford's AI Index tracked a fall of more than 280 times at a fixed capability level between November 2022 and October 2024 |
| Capability on a hard task | Yes | Each generation does things the last one could not, at no extra cost to you |
| A feature inside software you own | Yes | Your existing vendors ship AI into their tiers whether you ask or not |
| Knowing which process to change | No | It takes interviews, observation and arithmetic, and nobody has made that faster |
| Data clean enough to use | No | The work is proportional to how long the mess has been accumulating |
| Systems that talk to each other | No | Integration work is the same whether the model is new or old |
| A team that trusts the output | No | Trust is built by use, over months, and cannot be bought later |
| Knowing which vendor to believe | Partly | The field consolidates, though the claims get louder with it |
The pattern is that the fast-moving half is the half you rent, and the slow half is the half you own. A business that waits on both arrives at the cheap tools with none of the conditions that make them work. A business that waits on the first and starts on the second gets the price drop and the readiness.
Stanford's 2025 AI Index states the fall this way: the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024, which it attributes to increasingly capable small models. Read it as a fixed capability getting cheaper rather than every model getting cheaper, because that is the claim: what was expensive and impressive in 2022 is now a rounding error, and the same will be true in two years of what is expensive and impressive today.
How many businesses our size are actually using AI today?
About one in five businesses are using AI today, and the share is moving faster in firms with more than twenty employees. The United States Census Bureau's Business Trends and Outlook Survey, reviewed over data collected from 14 December 2025 to 3 May 2026, found overall use hovering between 17% and 20% of businesses, with 20% to 23% expecting to use AI within six months.
Size is the clearest split in that data. In the same Census survey, firms with at least 250 employees reported 37% use, firms of 100 to 249 reported 32% in the collection period ending 3 May 2026, and firms with four or fewer employees were under 20%. Use rose among firms with at least 20 employees between December 2025 and May 2026, while firms under 20 showed no significant change. One caution on the series: the Census broadened the question in November 2025 from AI used "in producing goods or services" to AI used in "any business function", so figures before and after that change are not directly comparable.
The investment side of the same picture appeared in a Gartner poll of January 2025, reported in its press release of 25 June 2025: of 3,412 webinar attendees polled, 19% said their organisation had made significant investments in agentic AI, 42% conservative investments, 8% none, and 31% were taking a wait-and-see position or were unsure. A webinar audience selects itself for interest in the subject, so treat that as the shape of the market rather than a representative sample of it.
Two readings are available and both are honest. A third of the market waiting means waiting is normal rather than negligent. And four in five businesses not yet using AI means the ones who get this right are not late.
What does it cost to start now and get it wrong?
Starting now and getting it wrong costs a quarter and a small amount of money, provided you scope it so that nothing is irreversible. That is the part most of the advice on this question misses: the risk is a function of how the first project is shaped, not of when you start.
Three rules make a first project cheap to be wrong about. Pick a process with a number attached, so you can tell whether it worked. Keep a person in the loop on anything that reaches a customer or moves money, so a bad output is a correction rather than an incident. And decide in advance what would make you switch it off, which is the decision nobody writes down and everybody needs.
What costs real money is not a failed first project. It is a tool bought for the whole company before anyone has used it, a multi-year contract signed on a pilot, and a build on top of data nobody checked. Those are the expensive versions of wrong, and all three are avoidable at any date.
The barrier owners themselves name is not risk anyway. ServiceTitan's 2026 Residential State of the Trades, a survey of more than a thousand residential specialty contractors reported by ACHR News on 13 April 2026, found 25% currently using AI, with the most common obstacle, named by 53%, being operational complexity or not knowing where to start, and 43% citing lack of trust in output. Both of those get smaller with practice and neither gets smaller with time. Why AI automation projects fail covers the failure patterns in detail.
What should we do first if we are not ready to buy anything?
Do four things first, all of which cost almost nothing and hold their value whatever you buy later. None of them needs a vendor, and all four are inputs to a decision rather than a decision.
Find out what is already running. Check which AI tools your team already uses, what they use them for, and what they paste into them. The five places to look take an afternoon, and the list doubles as a demand signal: people reach for tools where the work is heaviest.
Measure one bottleneck properly. Pick the process that costs the most skilled time and count it: how many times a week, how long each time, who does it, at what loaded hourly cost. A real number changes the conversation from whether to how much.
Write down what a fix would have to do. Not the technology, the outcome. "Every estimate gets a second contact within 48 hours and I can see which ones did" is a specification anyone can quote against, now or in a year.
Fix the data you would need. The records the fix would read, in the state they are in. This is the one that takes longest and the one that is useful regardless, and whether your data is good enough for AI sets out what good enough actually means.
When we built the specimen database for the Universidad Maimónides 3D-scanning lab, that sequence is what the project rested on: a counted bottleneck of about 15 hours a week, roughly $39,000 a year of skilled time at $50 an hour, and a clear statement of what the fix had to do. The case study shows what it produced.
When is waiting actually the right answer?
Waiting is the right answer in four situations, and all four come up in strategy conversations often enough to be worth naming.
The job is on your current vendor's roadmap. If the platform you already pay for is shipping the feature in the next release, buying it twice is waste. Ask them for the roadmap in writing and diarise it.
The process is about to change anyway. A business mid-way through a system migration, an acquisition or a restructure should not automate the process it is about to replace. Wait for the new shape, then look again.
The team is in a crunch. Adoption needs attention from the people doing the work. In a peak season, a filing deadline or a short-staffed quarter, the same project lands badly and gets blamed for the disruption. Six weeks of patience buys a better result.
Nobody owns it. A project with no accountable owner is a project that will stall at the first awkward decision, and the fix is naming the owner rather than starting anyway. Who should own AI in a company without a CTO is the shorter answer to that one.
What none of those is, is a reason to skip the four cheap moves in the previous section. Waiting on a purchase is a decision. Waiting on understanding your own business is a habit, and it is the expensive one.
Frequently asked questions
How much does it cost to start small with AI?
The four preparation moves above cost time rather than money: an afternoon on the tool inventory, a few hours counting a bottleneck, and whatever the data cleanup turns out to need, which is the only unpredictable one. On the build side, what moves the cost is how many systems the work has to touch, how much history has to be cleaned up first, whether a person stays in the loop, and who runs it afterwards. We scope that on a call rather than publishing a number, because the same described problem can be a configuration change or a substantial build depending on what it has to reach.
How will we know when the technology is ready for the thing we want?
Test it on your own work rather than tracking announcements. Take ten real examples of the task, run them through whatever is current, and read the outputs with the person who does the job today. That takes an hour and answers the question for your case, which no benchmark can. Re-run the same ten examples every few months: the cost curve above means the answer changes, and a kept test file is how you notice.
Is it better to wait, run a pilot, or build something properly?
All three are correct at different stages of knowing what you want. Wait when the decision belongs to a vendor roadmap or a process that is about to change. Pilot when you have a candidate process and no evidence, keeping it small enough that being wrong costs a fortnight. Build when the process is understood, the volume justifies it, and somebody has committed to running the result. The common error is a pilot that nobody ever decides about, which quietly becomes a subscription.
Is our industry too early for this?
Industry is rarely the deciding factor, and process shape is. Work that is repetitive, rule-governed, text or document heavy and high volume is ready now in any sector. Work that depends on physical dexterity, real-time judgement about people, or data nobody has recorded is not, in any sector. Our industry boards set out which jobs are proven, which are working and which are early, with the maturity stated per capability rather than per sector.
What if we build on a model that gets retired?
Model retirement is a real operational event and a manageable one, and it is a reason to care about how a system is built rather than a reason to wait. What protects you is that the business logic, the data and the integrations stay yours while the model is a component: a system designed that way takes a swap in days. Ask any supplier what happens when the model they are using is deprecated, and listen for whether the answer is a plan or a shrug. In our case the systems run on our infrastructure and keeping them current as the technology moves is our job rather than yours.
The boards referred to above are manufacturing, home and trade services, legal and accounting, dental and medical and professional firms, and how to get ROI from AI covers making a project pay once you start one.
Spend the quarter on the part that holds its value
We map how your business actually runs, put a number on where time and money leak, and tell you which parts are worth building now and which are worth waiting on. Then our engineering team builds what the roadmap calls for and we run it for you from there. We own the outcome, not the deliverable.