Yes, a small business can use AI price competition to get better terms, but only if it can actually switch providers. Competition is pushing prices down, and the businesses that benefit most are the ones holding a credible option to walk away from any single vendor.
The November 17, 2026 episode of The AI Daily Brief assembled three separate studies of how organizations purchase artificial intelligence. Examined individually, each study appears to report a narrow finding, but examined together, they describe a marketplace in which the buyer holds considerably more negotiating leverage than most owners recognize. The host concluded that enterprises probably have far more leverage than they are exercising, provided they have built the operational ability to switch.
This article explains what that leverage looks like in the data, what it means for a company with five to fifty employees, and what an owner in Northeast Indiana can do about it this month. It also identifies where the evidence is weak, because each study carries real methodological limits.
Key Takeaways
- Business AI usage expanded far faster than spending in one recent month, which points to falling prices per unit of work.
- Most companies using a major AI gateway now rely on two or more model providers, so switching has become a routine decision.
- Price per token is a weak guide to cost. Cost per finished job describes what AI actually contributes to your business.
- The largest spenders in the data are probably outliers, yet they hint at where the typical buyer may eventually move.
- Your leverage is real only when you can transfer your work to another provider without losing the instructions, records, or results you have built.
Why does the AI market now favor the buyer?
The market favors buyers because the price of AI is declining while competition among providers is intensifying. Across OpenAI and Anthropic, the average price per million tokens fell 21% in October. Over the same period, the Ramp AI Index recorded total token consumption across its businesses rising 36% in a single month, while spending rose only 2% during that interval.
The host attributed part of that price decline to the mere presence of inexpensive models from Chinese laboratories. A company does not need to adopt those models in order to benefit, because their availability pressures the large providers to lower prices, and those providers also pursue efficiency for their own competitive reasons. The decline therefore reflects several forces operating at once, rather than a single cause.
Growth in spending has also slowed considerably, as Ramp's token spend grew 1.05% in August, 0.22% in September, and 1.65% in October. If the October rate persisted, it would compound to roughly 22% annually, which represents a substantial deceleration from earlier growth, yet it still falls within the 20% to 40% increase that surveyed Global 2000 leaders anticipate for 2026.
For a small business, the practical reading is straightforward: a given dollar now purchases more finished work than it did a year earlier, and vendors have strong incentives to keep their prices competitive. Ramp, however, draws from a narrow and technology-forward customer base, so an Indiana contractor should not expect its own invoices to decline at the same rate. The three studies point in the same direction, and that agreement is what makes the trend worth acting upon.
How common is switching between AI providers?
Switching is considerably more common than most owners assume. In July, 64% of companies using Databricks' Unity Gateway relied on a single model provider, and by early November, 72% used two or more. That transition occurred within roughly four months, which is remarkably rapid for a purchasing habit that once seemed sticky.
The Arcana Research survey of Global 2000 leaders confirms the same direction: a third of OpenAI customers said they would consider moving, and many of them named Anthropic as the likely destination. Among GitHub Copilot customers, 43% expressed openness to switching, and many named Claude Code as the alternative they would consider. Preference is shifting between providers, not merely within a single vendor's product line.
Vendor share reveals the same pattern from another angle: just under 45% of US businesses on Ramp pay Anthropic, compared with 40.6% for OpenAI. On Unity Gateway, Anthropic received 68.8% of proprietary model tokens, and when Arcana asked respondents to name an exciting AI vendor, 81% named Anthropic. These figures describe a market in which buyers compare alternatives carefully before committing their budgets.
Switching is useful only when the work itself can move, and a company that stores its instructions inside one vendor's settings, and keeps its customer records there as well, possesses little practical leverage. A company that maintains its prompts, checklists, and records in files it controls can evaluate a second provider within an afternoon. That difference, rather than the published price list, usually determines who receives the better arrangement.
Should you judge AI by price per token or cost per finished job?
Judge AI by cost per finished job whenever you possibly can; Databricks advances a version of this argument, holding that the raw price of a token is the wrong figure to monitor, and that the more meaningful measure is the expense of completing a task successfully. A low-cost model that requires several retries can ultimately cost more than a higher-priced model that delivers acceptable work on the first attempt.
Consider a routine quoting task in which an inexpensive model drafts a quote that a staff member must rewrite twice before it can be sent. The token charge appears modest, yet the staff time, the delay, and the prospective customer who waited all represent genuine costs. A model that costs more per call but produces a usable quote once may be the economical choice for the business as a whole.
You can measure this without a dedicated data team by selecting three recurring tasks, such as quote requests, follow-up messages, or scheduling replies. Each month, total the AI spending attributable to each task type, then divide that total by the number of tasks completed correctly, and record how many of them required rework. After three months, the resulting pattern usually reveals which tool deserves the budget. An earlier Hey Button article on budgeting AI by cost per task describes a simple version of that tracking sheet in greater detail.
This approach also strengthens your position during negotiation: when a vendor announces a lower price, you can ask what happened to the price per finished job. If the vendor cannot answer that question, the reduced unit price may not change your economics in any meaningful way.
Why do the biggest AI spenders look so different from the median?
The gap between the largest spenders and the typical company is enormous, and in the Ramp data, the median firm spends $16.32 per employee per month on AI, while the top 1% of firms spend more than $10,000 per employee per month, roughly 625 times the median. Databricks observed a similar pattern among its own users, where the median user spends $2.49, while its 99th percentile user spends nearly $700 per week, a difference of about 281 times.
These figures require careful interpretation, because Ramp's customers are generally technology-based and technology-forward organizations, so its highest spenders represent a highly specialized group. Databricks observes only customers who route their traffic through its gateway, and Arcana's Global 2000 sample is comparatively small. Each dataset carries caveats, and none of them describes a typical Indiana trades company with a dozen employees.
Nevertheless, the question the host raised deserves serious consideration. Are the top 1% outliers, or do they preview where ordinary companies are heading? The most plausible answer lies somewhere in between. For a small business, the useful lesson is not to imitate the heaviest spenders. It is to recognize that spending concentrates in a few teams and tasks, and to decide deliberately which of those tasks merit additional investment.
Where does AI money go, and why can't most leaders see it?
Most AI spending flows toward a relatively small set of uses, and Ramp found that 72% of its AI spend originated in engineering teams. Customer support teams increased their share from 0.8% of spend in June to 13.3% in November, which makes customer service the fastest-growing business application in the data. Arcana reported that 38% of its Global 2000 respondents already deploy AI for customer service.
Visibility represents the weakest point in the picture. Arcana found that two-thirds of Global 2000 leaders cannot see how much they spend on model and API usage. That is primarily a large-enterprise problem, yet it reappears in small companies in a less obvious form. A ten-person firm with three AI tools billed to three separate personal cards may have no accurate picture of its own costs either.
The episode also observed that autonomous tools generate considerable excitement but comparatively little deployment. Arcana ranked autonomous tools fourth on its heat index of what companies are enthusiastic about, yet last on its adoption index of what companies have actually put into production. Put plainly, many businesses want the capability and have not yet established the routines that make it dependable.
A straightforward remedy suits most small firms: maintain one shared inventory of every AI tool you pay for. Record what each tool performs, who authorizes its access, what it costs each month, and when its contract renews. Review that inventory before every renewal, because that list forms the foundation for each negotiation that follows.
What else was in the November 17 episode?
The episode opened with headlines that surround the business story without belonging to it directly. Reporting indicated that Nvidia's board was concentrating on an independent investigation into Jensen Huang's conduct and was not discussing his return. Sam Altman argued that the AI industry would remain sound without Huang, citing strong revenue and demand. Separately, the Commerce Department issued a rule requiring data center operators in Singapore, Malaysia, Thailand, Indonesia, and Brazil to obtain export licenses for advanced AI chips, with existing facilities receiving ninety days to comply.
Other headlines concerned safety oversight and infrastructure resilience. OpenAI agreed to permit outside evaluators, including Apollo Research, METR, Transluce, and Goodfire, to work with its internal models. Apollo stated that the access and disclosure rights in its pilot did not yet satisfy its standard for full embedded evaluations, while the host contended that labs should not select their own auditors without commonly agreed standards.
Meanwhile, a major banking outage halted logins, ATM withdrawals, and transfers at several large institutions, and JPMorgan Chase attributed the disruption to a software defect in on-premises hardware rather than to a cyberattack.
These developments matter to a small business chiefly as background context. Chip export rules and the fortunes of major vendors can alter what AI costs and which tools remain available over time. A company that depends on a single provider should monitor news about that provider as carefully as it would monitor the financial health of any critical supplier.
What can a small business do with this leverage?
Begin with the inventory described above, and add a column indicating whether you could transfer that work to a second provider within one week. Mark any task that would require a month or longer to relocate. Those rows reveal where your leverage is weakest, and they deserve attention before anything else.
Next, run one real task through a second model. Choose work you already perform every week, such as drafting a follow-up message or summarizing a service visit. Supply both models with identical inputs, then compare each output against the standard you would accept from a team member. A small test grounded in your own work tells you far more than a published benchmark. Keep the test compact enough that you can repeat it every quarter without disrupting operations.
Then ask each vendor three questions in writing. How is usage billed, and what constitutes a billable unit? What happens to my data and my instructions if I leave? Can I export my records in a format I can actually use elsewhere? Vendors who answer plainly are easier to trust. Vendors who hesitate have revealed something valuable about the terms you would be accepting.
Finally, establish a review rhythm by comparing spending against finished work each quarter, and reexamining the market at every renewal date. Prices are moving quickly, so an arrangement that looked fair last spring may no longer match what competitors offer today. Expect the strongest terms to go to businesses that compare options, measure outcomes, and remain prepared to move.
What does an AI teammate do with this?
An AI teammate handles the everyday work on which a switching review or cost analysis depends. It can read incoming requests, organize the relevant details, track which tasks are complete, and draft replies for your approval. The default arrangement is draft-and-approve, so nothing reaches a customer until you authorize it. That structure keeps the business in control of its data and its customer relationships while routine work proceeds more quickly.
Access levels should correspond to the trust you have established. Hey Button offers read-only, draft-and-approve, and full access, and you select the level appropriate to each category of work. Customer messages, spending, and sensitive changes remain behind the approval rules you define. Those rules form part of the initial setup, and they simplify any later change of tools, because the instructions and approval rules belong to your business rather than to any single model.
The teammate is configured around the way your business actually operates, and model routing selects the least expensive model that still performs the job correctly, and it can fall back to another capable model when the preferred choice is slow or unavailable. A model switch leaves the teammate's instructions, the business facts it was configured with, and the access boundaries you set entirely unchanged. For a practical discussion of matching the model to the job, see this Hey Button article on when a cheaper model is good enough.
Setup and ongoing support are quoted according to your specific workflows and integrations. Any software subscriptions or usage charges are identified before you commit, so no surprise line items appear later. The outcome is a teammate you can keep, adjust, or replace on your own timetable.
How can you start without betting on one provider?
The starting point is the one the episode implies: buyers gain power when they can actually move their work. Hey Button builds that position from the first day, because your instructions, your records, and your approval rules remain with your business while an AI teammate drafts the routine work for your review. Our short questionnaire asks how work arrives today and which tools you already use, so the setup can begin with your real workflows rather than with a generic template.
Sources
- The AI Daily Brief, "How Business AI Use is Changing," November 17, 2026: https://aidailybrief.ai/e/2026-11-17
- Ramp AI Index, November and October editions, as cited in the episode: token usage, spend growth, price per million tokens, and spend concentration figures.
- Databricks Unity Gateway and its new AI index, as cited in the episode: provider mix, token share, and user spend figures.
- Arcana Research, first annual study on the state of enterprise AI, as cited in the episode: Global 2000 spending expectations, spend visibility, switching intent, and vendor preference figures.
Frequently Asked Questions
- Does a drop in AI token prices mean my bill will fall?
- Not automatically. A lower price per token can still produce a higher monthly bill when usage grows faster than prices fall, which is what one recent Ramp month showed. Ask your vendor how usage is billed, then compare total monthly spend against the work you actually finished.
- How do I know whether my business can switch AI providers?
- Check three things. Your prompts, instructions, and business records should live in files you own, you should be able to run the same job on a second model, and your contract should allow exit with a clear data return. If any answer is no, your leverage is smaller than the market suggests.
- What is cost per successful task, and why does it matter?
- It is total AI spend for a job divided by the number of jobs finished correctly. A cheap model that needs three retries can cost more per finished quote than a pricier model that gets it right once. Track it on a simple monthly sheet with one row per task type.
- Should a small business expect its AI spending to keep rising?
- Probably, but at a slower pace than the fastest years. One recent Ramp month's growth rate, if sustained, compounds to about 22% a year, which sits inside the 20 to 40% increases surveyed leaders expect. Budget a range, review it each quarter, and avoid assuming next year repeats last year.
- Who should own the AI vendor decision in a small company?
- The owner should, with input from whoever uses the tool every day. Keep one written list of every AI tool you pay for, what it does, who approves its access, and when its contract renews. Review that list before each renewal rather than after the bill arrives.
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.




