Strategy

A Nickel a Task: The Overnight Shift a 15-Person Company Can Afford Now

B

Benjamin Hopwood

Operations Scaling | Agentic AI Orchestration

September 10, 2026|5 min read
A Nickel a Task: The Overnight Shift a 15-Person Company Can Afford Now

Six Days as a Nobody

Most service businesses lose revenue the same quiet way: quotes go out, nobody follows up, and the customer hires whoever called back. The follow-up shift you cannot afford to staff is the work AI just got cheap enough to do.

Here is what changed, in one story. For six days in late August, the most-used AI model on OpenRouter, the platform where developers pick AI models the way travelers pick flights, was anonymous. No brand, no launch post. Developers routed real work through it and kept coming back on results and price alone. On August 26 the Chinese lab Z.ai revealed it as GLM-5.3-Flash, released the same day with open weights under an MIT license, and matching Claude Opus 4.8, one of the strongest Western models of the past year, on Artificial Analysis's Intelligence Index.

About the nickel: Artificial Analysis prices that index's benchmark task at roughly $0.045 at launch rates. A "task" there is one benchmark problem, so read it as a unit price, not your bill. The honest version for a business: an overnight run of a few hundred real tasks, drafted replies, triaged messages, follow-ups, a morning briefing, costs a few dollars in tokens. A year ago, capability at this level cost roughly ten times more.

What the Trades Are Already Doing

This is not theoretical, and you do not have to be a tech company. Bill Joplin's Air Conditioning and Heating, a Texas HVAC contractor founded in 1978, turned on the AI voice agent inside ServiceTitan, software it already ran. Per ServiceTitan's release: about 20 minutes to set up, and over roughly three months it handled 1,300-plus calls, booked 90 percent of them, completed 72 percent with no human touching the call, and generated over $74,000 in incremental booked revenue.

A vendor case study of a 22-person Tampa HVAC contractor tells the custom-build version: $52,000 in year-one build costs for an after-hours voice agent, live in five weeks. After-hours answer rate went from 62 to 98 percent, booking conversion from about 6 to 31 percent, roughly $340,000 a year in recovered revenue. And Wilson Companies, a large electrical, HVAC, and plumbing operator, reported two weeks of onboarding, three to four hours of its call-center manager's time, and about $30,000 a month in savings as abandoned calls fell from 18 percent to 7.

These are vendor-published numbers, so treat them as a best case. The pattern across them still holds: weeks not months to live, and payback measured against missed calls and unchased quotes. ServiceTitan's 2025 industry survey found 46 percent of trades contractors already using or trying AI, and the barriers it found were organizational, not technical.

The Honest Cost Line

There are two ways to buy this, and they cost very different amounts.

The first is switching on the AI features inside software you already pay for. Jobber sells an AI receptionist add-on at about $99 a month; ServiceTitan prices its by quote. Setup is minutes to hours. The capability is whatever the vendor built, and for phones and follow-up it is now genuinely good.

The second is a custom build. Real money: the Tampa contractor's $52,000, or, at bigger scale, a disclosed cost breakdown for a 200-employee manufacturer's contract-analysis agent came to roughly $120,000 to $160,000 in year one, $22,000 to $34,000 a year after, and 14 to 18 weeks to production.

Then there is the line almost nobody prices: review. Every deployment that works has a named person who checks the output, twenty to thirty minutes a morning is a realistic ask, and that person has to exist before you build. If your office manager reviews AI drafts, she is not scheduling crews for that half hour, so the loop has to hand her back more time than it takes. Notably, almost no vendor case study publishes its review-hours number. Ask for it. A vendor who cannot tell you how much human checking their system needs is telling you something.

Is My Data Going to a Foreign Lab?

Fair question, since the cheapest capable models right now are Chinese releases. Open weights means the model is a file you can run anywhere: through an American hosting provider, or on hardware you own, where nothing you type leaves your building. Self-hosting is real but not free; a machine that serves a model of this class well is a five-figure box plus someone to run it, so most small companies should start with a reputable domestic host and price self-hosting later. If you are in a regulated field, health, legal, finance, ask your compliance advisor before customer data touches any model, foreign or domestic.

What to Do, With Dates Attached

  1. This week: count last month's quotes that never got a follow-up, multiply by your average job value. That number is your case for doing anything at all.
  2. This quarter: switch on one AI feature inside software you already pay for, and compare booked jobs sixty days before and after. Cheapest possible experiment, and you may find you already own the tool you were about to buy.
  3. Before signing anything custom: demand the all-in number, build cost, monthly run cost, and review minutes per day, plus the week-two kill criteria. For calibration, our own first build starts at $6,500 all-in, with a $1,500 fully refundable deposit and nothing else invoiced until you accept the work. Ask any vendor for their equivalent sentence; the ones who have delivered can say it quickly.
  4. Name the reviewer before you start. No named owner with twenty spare minutes a morning means you are not ready yet, and fixing that costs nothing.

One more thing, honestly: your competitors can buy the same shelf. When factories first got electricity, the ones that swapped the steam engine for one big motor gained almost nothing; the gains went to the ones that redesigned work around the new power, and the laggards took decades to copy them. The model is now a commodity. The redesign of your work around it is where the next year's advantage sits.


Agentic Solutions builds first agent workflows for small and mid-size companies: the process, the guardrails, and the review habit that make cheap intelligence into finished work. First builds from $6,500 all-in, deposit fully refundable. Start the conversation.