This week the markets started asking the question most small business owners asked a year ago: does AI actually pay? China's homegrown AI and chip index has lost more than 40% since June. Alphabet just had its worst trading day on record. Apollo's chief economist says the "core economic flaw" in the AI trade is that compute costs more to produce than it returns to the people using it. He is right about the companies he is describing. He is describing the wrong way to buy AI.
What the selloff is actually saying
Start with China, because it is the clearest example. The country's AI exports are booming. Chips, servers and robots helped produce a record trade surplus and a third quarter that beat expectations. The stock market fell anyway. The CSI 300 is in a bear market, Hong Kong's Hang Seng is down 12% on the year, and Unitree, the humanoid robot maker that rose 600% on its August debut, has given half of that back. Beijing's $19 billion of ETF buying bought about a week of calm.
The analysts CNN quoted all land in the same place. Morningstar's Lorraine Tan said sentiment measures are short-lived "if there are no policies to address underlying key issues." Natixis called it "a repricing of crowded AI trades." In plain terms: a lot of money chased AI names because they were AI names, and nobody could point to the earnings.
The same thing is playing out in the United States at a slower speed. The Nasdaq has slid for two weeks. Alphabet dropped because investors saw search advertising growth slowing at the same time its AI spending kept climbing. The common thread is not that AI does not work. It is that enormous sums were spent before anyone agreed on what the spending was supposed to replace.
Why most AI spending loses money
Torsten Slok's argument at Apollo is worth reading closely because it applies at every scale, from a hyperscaler down to a four-person law office.
The core economic flaw in the AI bull case is that the marginal cost of producing compute exceeds the marginal benefit it delivers to users. How much compute and how many data centers are needed depends on the returns firms are able to generate with that compute. So far, the AI payoff for firms has been slow to materialize.
His data point is token spending. Growth in the amount businesses pay for AI usage is slowing, and when prices move even slightly, buyers throttle usage or switch to cheaper open-source models. Demand looks unlimited only while compute is close to free. Companies will keep paying a high price only if it shows up as more revenue, higher profit or real savings, and at scale there is little sign of any of the three.
That matches what we hear from owners every week, and it comes down to three mistakes.
1. AI got added on top of everything else. The business kept its CRM, its e-signature tool, its note-taker, its answering service and its bookkeeper, then bought an AI subscription too. Nothing was cancelled. Costs went up by exactly the price of the subscription.
2. The tool had no job. "Try ChatGPT" is not a plan. A chatbot with no assigned work, no access to the business's systems and no report at the end of the month cannot produce a return, because nobody defined what a return would look like.
3. The bill grows with use. Per-seat and per-token pricing means the more the tool helps, the more it costs. The business that finally gets value out of AI is the one that gets punished for it on the next invoice.
What AI that pays back looks like
The fix is to flip the order. Do not start with the AI. Start with what the business already spends, and only deploy an agent where it cancels a bill, covers work you would otherwise hire for, or brings in revenue you can count. Six places that money comes from, roughly in the order most owners find it:
Cancel the software stack. Most small businesses pay a dozen separate subscriptions: CRM, e-signature, meeting notes, proposals, outreach tools, bookkeeping software. One platform of agents that does those jobs, with usage included, replaces most of that list.
Cover work without hiring. Answering every call, taking notes, chasing follow-ups, prepping the books, screening applicants. These are the hires most owners put off, and the overtime and temp costs they pay instead. Agents take the routine part, and staff do the work that needs a person.
Cut outside professional fees. Fractional CFO hours, compliance consultants, routine legal review, the annual penetration test. Agents do the preparation, and the outside professional only reviews and signs off. You pay for judgment, not for the hours of assembly before it.
Lower IT and security cost. Break-fix bills, a pile of separate security tools and a scramble at cyber insurance renewal get replaced with one flat per-user price and agents that watch the systems around the clock.
Get leads cheaper. Search, local listings and AI-answer visibility run by agents, so customers find you on Google and in chatbot answers without an agency retainer and a separate web developer.
Bring in new revenue. Outbound email, LinkedIn and AI voice calls working the prospect list every day, with proposals ready when someone says yes. That is an SDR hire, a list vendor, a dialer and a sequencing tool, and the owner's evenings.
Notice what every item has in common. Each one is measured against a cost the business already carries. If the agent does not let you cancel or avoid more than it costs, it does not belong on the list. That is the whole discipline, and it is what the companies in this week's headlines skipped.
A ten-minute worksheet
You do not need a consultant to do this. Write down what you pay today for each of the following, per month:
1. CRM, e-signature, note-taker, proposal and outreach software.
2. Bookkeeping software plus bookkeeper hours.
3. Receptionist, answering service, or the value of the calls you miss.
4. IT support, security tools and backups.
5. Compliance consultant, pen test, fractional CFO.
6. Marketing agency, SEO and website hosting.
7. Prospecting: SDR, list data, dialer, sequences.
Add it up. That number is the only honest benchmark for any AI proposal you receive. If the proposal cannot show you which of those lines it removes, and by how much, it is the "added on top" mistake with a new logo. Only count costs you can actually cancel or avoid. If the math does not work for your business, the right answer is to not buy.
Keeping agents on a short leash
The other half of the market's worry is control. The biggest banks are now planning for AI agents that act without being asked, after customers started handing login credentials to agents that in some cases took actions on their own. Jamie Dimon put it bluntly in Chicago this month.
You're assuming they're going to act properly on your behalf and they may not. You can't stop it. You just have to, like, manage the hell out of it.
He is right, and the answer is not to avoid agents. It is to run them the way you would run a new employee with access to the bank account: one job each, a fixed list of what they are allowed to do, a person who approves anything involving money, contracts or an upset customer, an off switch you control, daily volume and spending caps, and a record of everything they did. An agent built that way from the start is cheaper to run and easier to trust than one you try to fence in later.
How Acclivity Labs helps
Acclivity Labs builds and runs AI agents for small and mid-sized businesses, and every engagement starts with the worksheet above. On the first call we list what you pay today, mark what our agents replace, and the difference is your savings. If it does not save you money, we say so.
One platform, usage included. Our CRM, eSign, Scribe, Proposal Generator, CoPilot for LinkedIn, Acclivity Books and the rest run on one private platform, on hardware we own and operate in the United States. There are no per-token or per-seat bills that grow when the agents are useful. Done-for-you packages start at $1,947 per month for Front Desk (an AI receptionist plus meeting notes) and $2,547 for Back Office (bookkeeping, accounting and CFO reporting), and a business that would rather not rent can buy its agents outright.
Every agent has a job and a report. You get a monthly outcome report for each agent: calls answered, notes filed, follow-ups sent, invoices prepared. Returns are measured against the bill they replaced, not against a feeling.
Short leash by design. Agents prepare and a person approves. Payments, contracts, refunds and anything legal go to a human first. Calls are recorded and transcribed, changes to the books carry a history of who and what, and every agent has an off switch and a daily cap. Each client runs in its own private tenant, never pooled with others.
Marketing and IT from affiliated partners. Search, local listings and AI-answer visibility come from Acclivity Digital Marketing, our affiliated agency. Help desk, patching, monitoring and 24/7 threat response come from our affiliated managed IT partner at a flat per-user price.
We run our own company on it. Our CRM, books, e-signature, meeting notes, LinkedIn outreach and AI calling all run on our own stack, and we retired the paid tools they replaced. We will show you the same dashboards and outcome reports we use ourselves.
The AI trade on Wall Street may or may not recover. For a small business that is beside the point. AI that cancels a bill, covers a hire or books a meeting pays for itself whether the Nasdaq is up or down, and AI that does none of those things was never an investment to begin with.