Company AI spending is outrunning budgets because it shifted from per-seat licenses to per-token usage, so a small business should budget for usage rather than seats. The November 6 episode of The AI Daily Brief walks through a new Menlo survey that measures this gap in detail, and this post translates those findings for owners of firms with 5 to 50 people in Fort Wayne, Auburn, and the rest of Northeast Indiana, including Allen County.
Key Takeaways
- Menlo estimates enterprise AI spending near $177 billion in 2026, up from about $33 billion in 2025, and reports that growth accelerated.
- Average spending per enterprise rose about 4.7 times, roughly 2.5 times the growth companies had budgeted for.
- Most traceable AI money came from net new budgets rather than from software that companies already pay for.
- Buyers moved toward frontier models even though small-model prices fell by more than 80 percent.
- A usage bill follows the work a teammate performs, so an owner needs a ceiling, a named reviewer, and monthly visibility.
- Ask every vendor which costs are fixed, which scale with usage, and how a model change would affect both.
What did the latest enterprise AI survey find about spending?
The latest survey estimates that enterprise AI spending reached roughly $177 billion in 2026, after climbing from about $33 billion in 2025. Menlo, which produced the estimate, polled roughly 400 US enterprise decision-makers in mid-September, and the host of the show treats the figures as directionally useful rather than precise.
The most important feature of the data is not the total itself but the rate of change. Most maturing markets decelerate as they expand, and this one conspicuously did not. Growth ran about 5 times from 2023 to 2024 and about 3.9 times from 2024 to 2025, then accelerated in 2026 past even the early rate, despite beginning from a much larger base. For a business owner, that trajectory matters more than any single annual figure, because it signals that the cost curve has not yet flattened.
Breadth of adoption barely moved, while depth of spending increased substantially across the organizations that already used the technology. The share of enterprises using AI rose only from 55 percent to 65 percent, which leaves roughly a third of enterprises reporting no use at all. Among the companies that do use AI, average spending per enterprise rose about 4.7 times over 2025. A small-business owner should read that pattern carefully: the competitive question is less whether rivals have adopted AI and more how much each adopter now consumes.
The episode also places the survey against a volatile market backdrop. Shares of major technology companies fell sharply in the same week, and the host attributed much of that decline to market structure, including leveraged funds unwinding positions, rather than to a sudden drop in demand for AI. Owners should keep the two stories separate, because a budget ought to respond to the spending trend and not to the stock market's mood.
Why did spending grow faster than companies had budgeted?
Spending outpaced plans because companies built their budgets around seats while their bills began following tokens. The episode reports that the average company planned to grow its AI spending about 1.9 times this year. A firm that spent $10 million on AI in 2025 therefore expected to spend about $19 million in 2026. Actual spending rose about 4.7 times per enterprise, which works out to roughly 2.5 times the growth the average firm had budgeted.
The host cautions against interpreting this as uncontrolled spending by default, because the underlying cause is largely experimentation. Many organizations were still discovering what the newer tools could accomplish, and that discovery naturally produced heavier use. The back-and-forth between teams pushing usage upward and finance departments pulling it back down reflects a mismatch between a per-seat forecast and a per-token reality. Forecasts built on headcount and license counts could not anticipate spending that scales with the volume of work each person delegates.
Small firms encounter the same mismatch in a proportionally smaller form. A bookkeeping office, a clinic front desk, or a roofing company that adds an AI teammate can watch its monthly bill rise and fall with the number of quotes, messages, and follow-ups it processes. Without a plan for that variable, the first warning often arrives on an invoice rather than in a forecast.
The host also draws a comparison that owners will find familiar. Budgeting for AI looks somewhat like budgeting for cloud computing, yet even that analogy breaks down under scrutiny. AI spending represents a new category with its own cost drivers, which means it deserves planning rules of its own rather than a place squeezed into an existing budget line.
Why does AI spending behave like usage instead of seats?
AI spending behaves like usage because the most capable tools now perform long, multistep work, and the price of that work depends on how much text the model reads and writes along the way. A seat license charges the same fee whether someone opens the software twice or two hundred times. A usage bill rises with every task the software completes, and the most useful tools complete far more tasks than a casual one.
A token is a small chunk of text, and vendors bill for the number of those chunks a job consumes, both going in and coming out. The episode notes that the frontier models enabling the newest work also carry the highest per-token prices. Claude Opus 5.5 is listed at $4 per million input tokens and $20 per million output tokens, while GPT-6 Astro is listed at $10 and $60. Those figures come from the episode, and a careful buyer should verify current vendor pricing before building a budget around them.
The host's broader argument is that spending now tracks what the technology accomplishes rather than how many people hold a login. That renders a monthly estimate derived from staff counts unreliable, since one employee who hands a large batch of follow-ups to a teammate can generate more usage than an entire department did a year earlier.
The practical response is to measure the work itself, rather than the number of seats, before any budget receives approval. Count the tasks the AI handles each week, such as quote requests answered or invoices reconciled, and compare that count against the bill. Once a firm knows its cost per completed task, it can decide which tasks justify automation and which should remain manual.
Are buyers paying more for better models, and does it matter to you?
Buyers are paying more for better models, and that choice shapes the cost of every task a business automates. The episode reports that frontier models' share of enterprise spending rose from 35 percent to 63 percent. Mid-tier, small, and older models all lost ground, even though labs cut small-model prices by more than 80 percent during the same period.
The host reads this as evidence that the market remains in an early phase. Businesses are racing to adopt capabilities that arrived with the newest frontier releases, so they accept premium prices to get the work done. The host expects cheaper and more efficient models to assume a larger role as adoption matures and the surrounding software stack grows more complex, which suggests today's premium will not remain a permanent requirement.
For a small firm, the useful question is not which model wins the headlines. The useful question is whether a particular task needs the most capable model at all. Summarizing a quote request, sending an appointment reminder, or sorting an inbox may not require the top tier, whereas a complex estimate or a delicate customer complaint might. A setup that can move work between tiers as conditions change protects the budget far better than a single fixed choice.
Model changes also carry risk for the instructions themselves. If a tool's behavior depends entirely on one vendor's model, a price increase or an outage can force an expensive rebuild. A sturdier design keeps business facts, approval rules, and access boundaries separate from the model underneath, so that a switch changes the engine without rewriting the work.
The episode also highlights growing concentration among a small number of model suppliers. Anthropic's share of enterprise API spending rose from 12 percent in 2023 to 55 percent in the current survey, while OpenAI fell before recovering to 31 percent. A small business does not need to predict the eventual market winner. It does need to know which models its tools use, where the work is processed, and what changes if a supplier revises its terms.
Is most AI money coming out of existing software budgets?
Most AI money appears to come from new budgets rather than from software companies already pay for. According to the survey, 55 percent of traceable AI spending originated in net new budgets, while only 21 percent came from existing software budgets. Software growth slowed only modestly, from 13 percent in 2025 to 11 percent in 2026, which suggests that AI has so far added to existing software spending instead of displacing it.
That finding should shape how an owner justifies a new tool. When AI spending is mostly new money, the owner needs a concrete goal for it, such as cutting response time on estimate requests or reducing the hours spent chasing unpaid invoices. Without a measurable goal, a new line item can appear to be waste, even when the underlying work is genuinely useful.
The host adds a warning that deserves attention. The next challenge to traditional software may come from AI tools that operate existing applications directly on a user's behalf, and the host believes that shift has barely begun. Owners who pay for several software subscriptions should carefully watch whether a teammate can use those tools directly, because that development could eventually change which subscriptions remain worth keeping.
Finally, the host reminds listeners that the surveyed companies are large organizations with dedicated budgets, sprawling teams, and extensive vendor relationships. A dozen-person firm cannot replicate those figures, but it can adopt the underlying framework without modification. Separating new spending from replacement spending and keeping each category on its own line gives a small business the same clarity that large enterprises are now trying to build.
What else was in the November 6 episode?
The episode opened with a headlines segment that covered several AI stories beyond the enterprise survey. The host reported that the Magnificent Seven technology companies lost about $1 trillion in combined market value in a single session, with Nvidia down 9.3 percent. Prediction markets also began pricing a delay to an Anthropic public listing, with odds of a listing before Thanksgiving falling from about 85 percent to 60 percent.
Several other items concerned safety evaluation and product design decisions. METR, an independent evaluator, published a report on Anthropic concluding that the pace of model improvement had roughly doubled over the past year, and it attributed much of that acceleration to AI assistance in research. Critics noted that the report left sandboxing and other safety measures unexamined. Anthropic also relaunched Claude for personal finance with read-only bank access through Plaid, and Meta added voice and video conversations to its Muse assistant.
Two additional items bear directly on the small-business customers that owners serve every day. OpenAI's newest image model can now generate short animated GIFs and accept rough sketches as reference material, which lowers the cost of simple marketing graphics. Meta's update illustrates how quickly consumer AI features are becoming everyday conversation, and that expectation will soon reach the customers who call, text, and email your business.
The sponsor messages and promotional segments in the episode have no bearing on the budget question, and this post relies on none of them. Readers who want the full context can start with the show notes and transcript on the original episode page.
What should a small business budget for AI in practice?
A small business should budget for AI as a monthly usage ceiling tied to one specific workflow, rather than as an open-ended subscription. Begin by choosing one task that consumes significant staff time each week, then estimate how often that task occurs in a typical month. Multiply that frequency by a realistic cost per task, add a margin for growth, and record the product as the ceiling.
Next, define in writing what the tool is allowed to do, because the permissions determine both the bill and the level of supervision required. The three common levels differ in ways that matter to a budget:
- Read-only: the tool sends nothing and costs the least to oversee.
- Draft-and-approve: the tool prepares work that a person reviews before anything goes out.
- Full access: the tool acts on its own, which carries the greatest risk, so it should wait until the owner has watched the tool handle real examples without incident.
Designate one accountable owner for the invoice. That person should review usage every month, compare it with the ceiling, and decide whether the scope of the work needs to change. A bill nobody reviews quickly becomes a surprise, while a bill with an accountable reviewer becomes a dependable input to the next planning cycle.
Finally, plan deliberately for the trade-off between cost and output quality. The episode's central lesson is that cheaper models are not automatically the right choice, and premium models are not automatically necessary. Test the same task on two model tiers, compare the output quality, and keep the less expensive tier for work where its results remain correct. The objective is a predictable cost for work done well, not the lowest possible price per token.
For more context, read how owners weigh whether AI tools will get more expensive and how to work out cost per task in a small business budget. Both posts apply the same per-task logic described here.
What does an AI teammate do with this, and what stays with you?
An AI teammate handles the drafting and sorting work while you keep control of approvals, which keeps the cost of each task visible and the risk contained. Hey Button sets up an AI teammate around your real workflows, and the default arrangement is draft-and-approve, which keeps a person in the loop. The teammate prepares the work, and nothing reaches a customer, a vendor, or a payment system until you approve it.
Hey Button offers three distinct access levels, namely read-only, draft-and-approve, or full access. Customer messages, spending, and sensitive changes stay behind the approval rules you choose. Because AI can make mistakes, review is a built-in part of setup, and you watch the teammate work on real examples before it touches a customer.
Changing the underlying model does not reset the arrangement you established. If the first-choice model is slow or unavailable, the system can fall back to another capable model. A model switch does not change the teammate's instructions, the business facts it was configured with, or the access boundaries you set, which is exactly the separation a budget needs while the underlying model market keeps shifting.
Here is an illustrative example, not a result. A customer asks, "Can you give me a quote for a spring cleanup?" The teammate organizes the request, identifies the missing details such as the address and photos, and drafts a reply requesting them. You approve that reply before anything is sent, and the usage for the task appears in the monthly record you already review.
Setup and ongoing support are quoted individually, around your specific workflows and integrations. Any software subscriptions or usage charges are identified before you commit, so the budget you build from this post contains no surprise line items. If you run a Northeast Indiana business and want to start with a single workflow, start the Hey Button questionnaire and tell us which task costs your team the most time each week.
Sources
- The AI Daily Brief, "Companies Are Spending 5x More on AI Now," November 6, 2026 (episode, show notes, and transcript): https://aidailybrief.ai/e/2026-11-06
- Menlo, 2026 state of enterprise AI survey, as summarized in the episode (about 400 US enterprise decision-makers, fielded mid-September 2026).
Frequently Asked Questions
- What should I do in the first 30 days before committing to an AI teammate?
- Start by writing down the one workflow you want help with and how often it occurs each month. Then request a written list of fixed costs and usage charges. Watch the teammate handle real examples before approving access beyond drafts, and compare the first month of usage with your ceiling.
- How do I compare two AI vendors that price differently?
- Convert both quotes into the same unit, such as expected cost per completed task at your real monthly volume. Include setup, subscriptions, and usage in that calculation. A vendor with a lower per-token rate can still cost more once the full volume of your work is counted.
- Is most AI spending new money or money moved from old software?
- Mostly new money, according to Menlo's tracing. About 55 percent of traceable AI spend came from net new budgets, while 21 percent came from existing software budgets. Justify AI as its own line with a clear goal, not as a replacement for a subscription you already pay.
- Does the big-company spending trend apply to a 10-person business?
- The direction applies, but the scale does not. Menlo surveyed about 400 US enterprise decision-makers, many running large teams and dozens of vendors. A 10-person shop should begin with one workflow, a monthly spending ceiling, and one named owner who reviews the invoice.
- What should I ask an AI vendor before I commit to usage charges?
- Ask which costs are fixed and which rise with usage, and what a typical month looks like for a business your size. Ask what happens during a usage spike and whether you can set a cap. Ask which models do the work and whether your setup survives a model change.
Hey Button
Want a teammate that handles this for you?
Answer a few questions about your business and Hey Button will show you what an AI teammate would take off your plate, with you approving anything that goes out.




