Because cost was never your constraint. Judgment was. Here is the exact question this article answers: if AI is now cheap enough to run on almost anything, how do you decide what to actually point it at?
The Week Intelligence Got Cheap
On July 31, 2026, OpenAI cut the price of its Luna model by 80 percent. Twenty cents per million tokens in. A dollar twenty out.
Read that again, because the number is doing something people are not noticing. A million tokens is roughly 750,000 words. For twenty cents, you can now hand a capable model the equivalent of ten novels and ask it to think about them.
Here is the part that should stop you cold. OpenAI published research showing that the efficiency gains came from one of its own models, Sol, rewriting the GPU code that serves the family. Fifteen percent more efficient. Twenty percent cheaper to run. The model made itself cheaper, and the savings went to customers.
So here is the direct answer to the question in the headline. Your business is still stuck because the price of intelligence was never what was holding you back. What is holding you back is that nobody in your company has decided which decisions are actually worth improving. Cheap intelligence pointed at the wrong problem is just a cheaper way to be busy.
I have watched this happen in dozens of businesses. The owner finally gets access to good tools, feels the relief of the price drop, and then spends six months generating more of what they were already producing too much of. Nothing changes because nothing was decided.
Key Takeaways
- OpenAI cut Luna pricing by 80 percent on July 31, 2026, and the efficiency gains came from one of its own models rewriting its serving code.
- MIT’s NANDA initiative found that 95 percent of enterprise generative AI pilots produced no measurable profit and loss impact, and the failure was in approach rather than technology.
- Roughly half of corporate generative AI budgets went to sales and marketing, while the clearest returns appeared in unglamorous back-office automation.
- The right first move is a decision audit, not a tool purchase, because AI improves decisions and most businesses have never listed theirs.
- Pick three recurring decisions with real financial weight, document how you make them, and only then automate.
The Problem Nobody Says Out Loud
Most business owners cannot tell you what decisions their company makes.
They can tell you what their company does. They can list tasks, roles, deliverables, and deadlines. Ask them to name the recurring decisions that determine whether the month is good or bad, and the room goes quiet.
I have been on the wrong side of this. In the early days of White Beard Strategies I could describe our workflow in exhausting detail and could not have told you the five judgment calls that actually moved revenue. I optimized process. I never touched judgment. And process optimization on top of unexamined judgment is how you get very efficient at doing the wrong thing.
This is not a small failure. It is the failure. MIT’s NANDA initiative studied this directly, drawing on 52 executive interviews, surveys of 153 leaders, and analysis of 300 public AI deployments. The finding was blunt: roughly 95 percent of enterprise generative AI pilots delivered no measurable impact on profit and loss. Lead author Aditya Challapally and his team were clear that the problem was not model capability. It was integration and misaligned priorities.
That study is worth sitting with, because the natural reaction is to assume those companies had bad tools or bad luck. They did not. They had the same models you have. What they lacked was a decision worth improving.
I know how uncomfortable that is to hear if you have already spent money and time on this. I have spent both. But there is something freeing in it too. If the constraint were technology, you would have to wait. If the constraint is clarity, you can fix it this week.
What the Research Actually Shows
Four findings, taken together, tell a coherent story.
The money goes where the excitement is, not where the return is. The MIT work found that about half of generative AI budgets went into sales and marketing, while some of the clearest measurable returns showed up in dull back-office automation. Nobody posts a case study about their invoice reconciliation. That is exactly why it is underserved.
Feeling faster and being faster are different things. METR, a nonprofit research organization, ran a randomized controlled trial with 16 experienced open-source developers across 246 real coding tasks. Developers using AI tools completed work 19 percent slower. Afterward, those same developers estimated that AI had made them roughly 20 percent faster. The sample was small and METR now labels the finding historical rather than current, so do not treat it as a permanent law. Treat it as a warning about self-report. If you are measuring your AI results by how the work felt, you are not measuring.
Human judgment is where the value concentrates. Anthropic analyzed roughly 400,000 Claude Code sessions between October 2025 and April 2026. The pattern that emerged is that people make most of the planning decisions while the model makes most of the execution decisions, and domain expertise amplifies how much useful work the model produces per instruction. The person who knows the business well gets dramatically more out of the same model.
The people building this are asking for a brake. On July 29, 2026, more than a thousand employees across OpenAI, Anthropic, Google, Meta, and Thinking Machines signed a public statement called Pacing the Frontier, asking governments to build the ability to deliberately slow AI progress if it becomes necessary. Both OpenAI and Anthropic publicly endorsed it. This is not a doom signal. It is a signal that the pace is real, which means the window where clarity beats capital is open right now and will not stay open.
Put those four together and the conclusion is not subtle. The models are ready. The pricing is ready. The gap is entirely on our side of the table.
The Decision Audit
Here is what changed things for us, and for the operators I work with.
Stop asking what AI can do. Start asking what decisions your business makes.
A decision is not a task. Sending an invoice is a task. Deciding which client to send a proposal to first is a decision. Writing a follow-up email is a task. Deciding whether this lead is worth a follow-up at all is a decision. Tasks are where the hours go. Decisions are where the money goes.
When you audit decisions rather than tasks, three things happen almost immediately.
First, the list is shorter than you expect. Most small businesses run on somewhere between eight and fifteen recurring decisions. That is a list you can actually work through.
Second, the financial weight is wildly uneven. Two or three decisions on that list account for most of the variance in your results. The rest are noise you have been treating as strategy.
Third, you discover that you already have the inputs. This is the part that surprises people. The information required to make your highest-value decision well is usually already sitting in your CRM, your email, your invoices, or your call notes. It is not missing. It is scattered, which is a completely different and much cheaper problem.
That third point is where the price drop finally matters. When intelligence costs almost nothing, pulling scattered context together and reasoning across it stops being an expensive engineering project and starts being a Tuesday afternoon. The economics of context assembly changed. Most people are still spending the savings on content generation.
I am not going to tell you this makes the work easy. Documenting how you decide something is genuinely hard, because the parts you do best are the parts you cannot immediately explain. But it is the work that makes everything downstream function.
Practical Steps
1. List every recurring decision in your business. Not tasks. Decisions. Go week by week and write down every point where somebody chooses between options. Include the ones you make so automatically you forget they are decisions.
2. Attach a number to each one. Estimate what changes financially when that decision is made well versus poorly. You do not need precision. You need relative weight. Ranking by feel is fine on the first pass.
3. Cut the list to three. This is the step people skip and it is the whole game. You cannot improve fifteen decisions this quarter. You can improve three. Choose the three with the highest financial weight, not the three that annoy you most.
4. For each of the three, write down the inputs a perfect version would need. What would you have to know to make this decision correctly every time? Be exhaustive. Then mark which of those inputs already exist somewhere in your business and which genuinely do not.
5. Start with the decision where you already have the most inputs. Not the most valuable one. The most ready one. You need a working example before you take on the hard one, and a fast result buys you the patience for the slow result.
6. Build the smallest possible workflow and keep a human in the review seat. Pull the inputs, produce a recommendation, and have a person approve it before anything happens. The Anthropic session data is clear that this split, human planning and review with model execution, is where the results actually live.
7. Measure with a number you set in advance. Decide before you start what result would count as success, and write it down. Then check it honestly. The METR finding exists because smart people trusted how the work felt instead of what it produced.
Frequently Asked Questions
Does the price drop mean I should switch to the cheapest model for everything?
No. It means you should route deliberately. Use cheap fast models for high-volume routine work and reserve expensive models for tasks requiring real judgment or nuance. Run a blind comparison on your three most common tasks first. In many cases the difference will not be visible to your customer, and that is where the savings are real.
How do I tell a decision apart from a task?
A task has one correct execution. A decision has multiple defensible options and requires choosing. If two competent people could reasonably do it differently and get different business results, it is a decision. If they would do it identically, it is a task.
What if my highest-value decision depends on information I do not have?
Then your first project is collecting that information, not automating the decision. This is a good outcome, not a setback. Most businesses discover one or two inputs they should have been tracking for years, and starting today compounds.
Should I hire someone to do this or do it myself?
Do the decision audit yourself. Nobody else can name the judgment calls that make your business work. You can delegate the building. You cannot delegate the deciding, because that knowledge only exists in your head right now.
How long before I see a result worth talking about?
If you pick a narrow decision where the inputs already exist, thirty days is realistic for a working version and ninety days is realistic for trustworthy results. If you are still nothing after ninety days, the scope was too wide. Cut it in half and run again.
The Close
The model rewrote its own code to make itself cheaper. A thousand of the people who build these systems signed their names to a request for a brake pedal. Both of those things happened in the same week.
That should tell you something about the moment we are in. The tools are no longer the bottleneck and they are not going to slow down enough for you to catch up by waiting.
So stop waiting for the technology to be ready. It is ready. Stop waiting for the price to come down. It came down 80 percent while you were reading your email.
Sit down this week with a blank page and write the header “decisions we make.” Then fill it in. It will take you two hours and it will be the least glamorous two hours of your quarter. It will also be the only two hours that change what the rest of the year looks like.
Cheap intelligence is not an advantage. Everyone has it now. Knowing exactly where to point it is the whole advantage, and that has always been a human job.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs turn AI from a novelty into an operating system for their business. He is the creator of the Perfect Prompt Framework, a speaker, and a father of eleven, which is where most of his opinions about systems and delegation actually came from.