Salesforce shipped seven named AI agents on September 11. Casey, Paige, Carter, Hunter, Marshall, Piper and Fin. Each one is scoped to a single job across sales, service, commerce, IT and HR, supply chain, and customer experience. Each runs on the company’s existing Customer 360 data and obeys the permissions already in place. Trade press filed it as an enterprise story and moved on. That reading misses the useful part. The launch is not a product you will ever buy. It is a design pattern, and the pattern costs you nothing to copy.
The pattern is this. One agent. One job. One data source it is allowed to read. One short list of actions it is allowed to take. A name, so everyone in the building knows what it is and what it is not. Compare that to how AI actually shows up in a 12-person local business: one general chat window, open in a browser tab, used by four people who each prompt it differently and none of whom wrote down what it is for.
The Gap Is the Spec, Not the Tool
Adoption stopped being the story a while ago. The 2026 U.S. Chamber of Commerce small business survey puts AI use at 89 percent of small businesses, up from 36 percent in 2023. Generative AI specifically sits at 58 percent, up from 40 percent the year before. Almost everyone is using something.
The results are lumpier. Aufsite Research found that 77 percent of small businesses using AI have no formal system for how they prompt or apply it. Only 23 percent have had any training on the tools they already pay for. Salesforce research puts revenue lift at 91 percent of SMB AI users, and adopters at 2.3 times more likely to report growth than non-adopters. Both things are true at once: the tools work, and most shops are running them without a spec.
A spec is not a strategy document. It is one page, and it answers five questions. What job does this agent do? What does it read? What is it allowed to send or change without a human? What does it hand back to a person, and to whom? What does failure look like, and how would you notice? If you cannot answer all five in ten minutes, the agent is not ready and neither are you.
That page is the entire difference between the 77 percent and the rest. It is also the reason Salesforce bothered to give its agents human names. A named thing with a written job description gets reviewed, corrected and fired. An unnamed browser tab gets used badly forever.
Build the Missed-Lead Agent in Five Days
Pick one workflow. The highest-yield candidate in most local businesses is the lead that came in and never got answered. The after-hours call. The web form at 4:40pm on Friday. The quote you sent 11 days ago that nobody followed up on. That pile is already paid for. Nothing in it requires new demand.
Here is a five-day path a working owner can run without clearing the calendar.
Monday, 30 minutes. Write the spec. Job: follow up on every inbound lead with no reply after 24 hours and every open quote with no reply after 5 days. Reads: your form submissions, your call log, your quote list. Writes: a draft message only. Hands back: the draft, to you, in one place.
Tuesday, 60 minutes. Get the data into one table. A Google Sheet is fine and often better than a CRM migration you will abandon. Columns: name, contact, source, date in, last touch, status, quote amount. Pipe the form and the call log in with Zapier or Make. Zapier’s Professional plan runs about $30 per month at low volume. Make starts near $10.
Wednesday, 90 minutes. Build the draft step. This is one API call or one Zapier AI step per row that qualifies. Feed it the row, your spec, and three follow-up messages you actually sent and were proud of. Output goes to a draft folder or a second sheet tab. Nothing sends. Model cost at this volume is real but small, and the numbers below show why.
Thursday, 60 minutes. Shadow run. Point it at last month’s leads, the ones whose outcome you already know. Read all 40 drafts. Mark the ones you would have sent as written. If that number is under 70 percent, your spec is too vague or your three examples were too few. Fix the spec, not the model.
Friday, 20 minutes. Turn it on with a human gate. Drafts land in one place every morning at 7. You or your front-desk person approves and sends. Keep the gate for at least three weeks. Auto-send is a decision you earn with data, not a setting you flip on day one.
What the Token Price Collapse Actually Buys
The pricing floor moved hard this year, and it moved in your favor. Flagship models from the major vendors converged around $3 to $5 per million input tokens and $12 to $30 per million output. The cheap tier went further. Google’s nano model prices at $0.075 per million input. Gemini 3.8 Flash landed September 2. Meta shipped Muse Spark 1.3 the same day near a blended $0.10 per million.
Run the math on the missed-lead agent. A lead row plus spec plus examples is maybe 2,000 input tokens. A drafted follow-up is maybe 300 output tokens. At 200 leads a month on a mid-tier model, you are looking at single-digit dollars. Your real monthly cost is the automation plumbing, so $20 to $60 all in, plus the 4 hours you spent building it.
Now look at the other end of the market. OpenAI has floated agent products at $2,000 per month for knowledge-worker roles and $10,000 for a software developer agent. Anthropic prices its agent runtime at $0.08 per session hour. The spread between those two worlds is not capability. It is who wrote the spec. Enterprise buyers pay five figures to have someone else define the job, own the integration and carry the risk. You can define the job yourself in half an hour because you already know it better than any vendor would.
That is the whole arbitrage available to a local operator right now, and it has a shelf life.
The Signal to Watch Before Mid-October
Watch what your existing software vendors do with agent pricing in the next 30 days. The field-service and SMB CRM platforms are all building agent layers, and there are only two ways to charge for one: fold it into the seat price, or bill per action taken. Seat-price bundling means the capability becomes table stakes and your edge evaporates by spring. Per-action billing means the vendors expect volume, and the operators who already know which workflow to point it at will pay far less per outcome than the ones who turn it on everywhere.
Either way, the window where a written one-page spec and $40 a month beats a $2,000 agent subscription is open now and will not stay open. Salesforce named seven agents because naming forces clarity about the job. Name one. Give it one job. Read its first 40 drafts yourself. That is the entire assignment this week, and it fits between the jobs you already have on the board.
Atlas Unchained writes one of these every weekday, one pillar at a time, for operators who would rather ship than read. Subscribe and get the next one in your inbox before the shop opens.
