No, and the reason is uncomfortable: the reliability curve is flattening while the capability curve keeps climbing, which means the thing you are waiting for is not coming on the schedule you assumed.
I spent the better part of a year waiting.
Not loudly. I was shipping, teaching, building. But underneath the activity there was a quiet assumption I never examined, which was that the real automation would start once these things got trustworthy. One more model generation. Two at most.
Then I would hand over the work that actually matters.
That assumption was wrong, and finding out why changed how I build.
Here is the short answer before the long one. Capability has climbed hard for eighteen months. Reliability has barely moved. Agents score better on benchmarks and still behave inconsistently across identical runs, and they remain poorly calibrated about their own uncertainty, which is a polite way of saying they are confidently wrong at about the same rate they always were.
So waiting is not a strategy. It is a bet on a curve that is flattening.
What changed for me was realizing that reliability was never the variable I controlled. Task design was. And task design turns out to matter more.
Key Takeaways
- Capability and reliability are two different curves, and only one of them has been climbing over the past eighteen months.
- An AI that is right sixty percent of the time is genuinely useful when the check is cheap and the undo is easy, and nearly worthless when either one is expensive.
- Coding agents work not because models are better at code, but because code ships with tests, diffs, and version history built in.
- You can build the equivalent of tests and undo into any non-technical workflow, and that is the actual unlock.
- Choose your next delegation by how safe it is to check, not by how badly you want it off your plate.
The Year I Spent Waiting
Let me describe the trap honestly, because I was fully inside it.
You try something ambitious. It works four times and then does something strange on the fifth, and the strange one costs you an hour of cleanup. So you file the whole category under not yet. You keep using AI for drafts and summaries, the low-stakes stuff, and you tell yourself you are being prudent.
That is what I did. And I want to be careful here, because prudent was not the wrong instinct. The instinct was right and the conclusion was wrong.
The instinct said do not put something unreliable in a place where a mistake is expensive. Correct. Keep that instinct.
The conclusion said therefore wait for it to become reliable. That is where I went sideways.
I know something about waiting for conditions to improve before you move. I have been in prison. I have been through bankruptcy. I have been through two divorces. Every one of those taught the same lesson from a different direction, which is that the conditions do not improve on your schedule, and the people who do well are the ones who build inside the conditions they actually have.
I just did not connect that to AI for an embarrassingly long time.
[PLACEHOLDER FOR JONATHAN: one specific moment where an AI workflow failed on you and what it cost. Two or three sentences, concrete, with the real number. This draft deliberately does not invent one.]
The thing that finally moved me was not a better model. It was watching how the people getting the most out of this were actually working, and noticing that none of them were waiting for anything.
What The Evidence Says About The Gap
I want to show you the numbers, because this is the part that made it click for me.
Longitudinal work on agent performance over the last eighteen months found something specific. Accuracy on standard benchmarks has risen steadily. Overall reliability has shown only marginal improvement. And as agents get more capable, the reliability gap widens rather than narrows.
Read that last clause again. It gets worse as the models get better.
That is not a contradiction once you understand what reliability means here. Reliability is consistency across runs plus calibration about uncertainty. A model can be substantially more accurate on average and still give you a different answer on Tuesday than it gave on Monday, and still tell you it is sure both times.
Second piece of evidence. Forrester and Anaconda research, replicated independently by a16z and an MIT Sloan CIO panel, found 88 percent of agent pilots never graduate to production. Deloitte's 2026 Tech Trends put it at 89 percent. The blockers leaders named were evaluation gaps at 64 percent, governance friction at 57 percent, and model reliability at 51 percent.
Notice the ranking. Reliability is third.
The two bigger blockers are both about whether you can tell if the thing is working. That is not a model problem. That is yours and mine.
Third. This month Nvidia shipped an Open Agent Safety Platform, built with more than a hundred industry partners, designed to quarantine rogue agents quickly. Separately, a company called Classie launched a platform for real-time tracking and control of enterprise agents.
When the people selling the shovels start selling containment, the industry has quietly agreed on something about failure rates. I do not say that as a criticism. I say it because it is confirmation.
Fourth, and this is the one I keep coming back to. Across the practitioner communities, the recurring finding is that when steps are small, observable, and reversible, an agent running at roughly 61 percent reliability becomes genuinely useful, and most of the real value is landing in assistive paths rather than autonomous ones.
Sixty-one percent. Useful.
That number should have been obvious to me and it was not, because I had been thinking about reliability as a property of the model instead of a property of the system the model sits inside.
Fifth. Goldman Sachs found that 76 percent of small businesses report using AI, but only 14 percent say it is fully embedded in core operations. The gap between those two numbers is almost entirely made of people doing what I was doing, which is using AI for things that do not matter much while waiting for permission to use it for things that do.
What Actually Changed My Mind
Here is the reframe, and it came from watching developers rather than business owners.
Coding agents work. That is not controversial among people who use them. Claude Code and Cursor are doing real work for real teams right now, at reliability levels nobody would call impressive in the abstract.
For a long time I assumed that was because the models are better at code than at business writing or analysis. Plausible. Also wrong.
Code works because of what surrounds it.
Every change is a diff you can read. Every change can be reverted with one command. Tests run automatically and tell you something broke before a human sees it. The entire history is preserved. The unit of work is small by convention.
Small. Observable. Reversible. Not because anyone designed code that way for AI, but because developers built those properties for themselves decades ago and AI inherited them.
That is the whole secret. The environment makes unreliability survivable.
Which means the question is not "when will AI be reliable enough for my work." The question is "what would it take to make my work survive unreliability."
And that question has an answer I can act on this week.
Once I saw it that way, my whole approach inverted. I stopped evaluating AI by asking how often it is right. I started evaluating my own processes by asking three things. How fast can I tell if this output is wrong. How expensive is it to fix. Can I undo it.
Where those three answers are good, I delegate aggressively and the sixty percent model earns its keep. Where they are bad, I do not delegate at all, and I stop feeling guilty about it, because that is not timidity, that is arithmetic.
The best part is that improving those three answers is work I control completely. No model release required.
How To Build This Into Your Own Work
Here is what I would do, in order, starting today.
One: find the steps in your workflow with no undo. Walk through your process and label every step reversible, catchable, or permanent. Be honest about permanent. An email that has been sent is permanent. A post that has been published is permanent. A number that went into an invoice a client already paid is permanent. Most people have fewer permanent steps than they fear and more than they expected.
Two: make the check cheap before you make the task bigger. Before you hand a task to AI, decide how you will know if the output is wrong, and time yourself doing that check once. If verification takes longer than doing the work yourself, you have not found an automation, you have found a trap. Often the fix is just changing the output format so the likely errors are visible instead of buried.
Three: shrink the unit of work. This is the single highest-leverage change and almost nobody makes it. Instead of asking for the finished thing, ask for the piece. One section instead of the whole document. One decision instead of the whole plan. A small wrong step is a correction. A big wrong step is a disaster, and they cost the same to generate.
Four: build your undo. You do not need version control software. You need a copy of the thing before it was changed, somewhere you can find it. A dated folder works. A duplicated document works. The point is that reverting is a thirty-second action instead of a reconstruction project.
Five: make the model tell you when it is unsure. Put it in your standing instructions, in plain words, with the exact phrasing you want it to use. The failure mode that hurts most is not a wrong answer, it is a confident wrong answer. You can partially fix that with instructions, and most people never try.
Six: choose the next task by safety, not by resentment. This one stung when I noticed I was doing the opposite. I kept trying to delegate whatever annoyed me most, which is almost always the highest-stakes judgment work, which is exactly the wrong place to start. Rank your candidate tasks by how cheap the check is and how easy the undo is. Start at the top of that list, not the top of your frustration list.
Seven: measure what you are guessing at. For two weeks, keep a tally of how often the output needed correction and how long the correction took. You will find out you have been either overtrusting or undertrusting, and you will not know which until you count. I was undertrusting by a wide margin, and the only way I found out was writing it down.
Frequently Asked Questions
Is AI reliable enough to use for client-facing work?
For client-facing work with a human review step, yes, and that is how most of it should run anyway. For client-facing work with no review step, no, and the reason is not the model, it is that you have removed the only mechanism that catches errors. Keep the human on anything that leaves your building with your name on it.
How accurate does AI need to be before it is worth using?
It depends entirely on how cheap the check is. At roughly 61 percent reliability, practitioners are getting real value when steps are small, observable, and reversible. The same 61 percent is worthless if verifying the output takes as long as doing the work. Accuracy is not the variable to watch. Cost of checking is.
Why do AI coding tools seem to work better than AI business tools?
Because code arrives with tests, diffs, version history, and small units of work already built in. The environment makes mistakes cheap to find and cheap to undo. The models are not dramatically better at code. The surroundings are better, and you can copy the surroundings for non-technical work.
What should I never hand to an AI agent right now?
Anything permanent that you cannot verify quickly. Sending to a list, publishing, moving money, making commitments to a client, deleting anything. Not because the agent will certainly fail, but because when it does there is no undo, and an action with no undo should have a human on it regardless of what is making the decision.
Will this change when the next generation of models ships?
The capability will change and the reliability probably will not, based on the last eighteen months. Open-weight releases from several labs are expected this month and will likely cut costs more than they cut error rates. Build the checks and the undo now, and every future model release makes your system better instead of making your assumptions wrong.
The Thing I Wish Someone Had Told Me Sooner
I wasted a year being careful about the wrong thing.
Careful is good. I will not apologize for careful. But I was careful about the model when I should have been careful about the process, and those are not the same job.
Here is what I believe now, and I believe it firmly enough to say without hedging. The people who get the most out of AI over the next few years will not be the ones who waited for it to be trustworthy. They will be the ones who built workflows where trust was not required.
That is not a small distinction. It is the entire distinction.
And it is good news, because it puts the work back in your hands. You cannot make the model more consistent. You can absolutely make your process more forgiving. Small steps. Visible output. A cheap check. An easy undo.
Nobody is going to ship that for you. It is yours to build, and it does not take a technical team or a budget. It takes an afternoon and the willingness to stop waiting.
I do not fear AI taking people's work. I never have. What I fear is people standing at the edge of their moment, waiting for conditions they were never going to get, while the conditions they already had were good enough.
That was me for a year. It does not have to be you.
Stop waiting for the tool to deserve your trust. Build the process that does not need it.
P.S. The thing that embarrasses me most is that the answer was sitting in plain sight the whole time. Developers solved this decades ago and called it version control. I was waiting for a breakthrough and the breakthrough was a backup copy and a checklist.
About the author
Jonathan Mast is the founder and CEO of White Beard Strategies LLC. He teaches non-technical business owners how to use AI to amplify the skill and experience they already have, through a Facebook community of roughly 500,000 members, the AI Insiders membership, and live training programs built around the Perfect Prompt Framework. He has been to prison, through bankruptcy, and through building and losing a seven-figure agency, which is the main reason he is suspicious of anyone who tells you to wait for better conditions.
Draft note for review: one placeholder is marked in the opening section for a specific personal failure story with a real number. Nothing was invented to fill it. The Anti-AI Writing Style Guide referenced in the task is not present as a file in this workspace and the linked Google Doc was not retrievable in an automated run, so this draft was written against SELFSCRIBE-3.1.md plus the hard formatting rules in JONATHAN-CONTEXT.md section 5. All statistics are sourced in AI_Influencer_Report_2026-10-02.md.