Fractional CAIO playbook: what it takes to run AI agents in your operations
Fractional CAIO playbook, part 2: orchestration, cost and guardrails
This is the second piece in our Fractional CAIO playbook series. Part one covered how to find every AI agent that already has keys to the business. This part covers what it takes to run agents in daily operations: how work passes between tools, what it costs, and the guardrails around it. It is written for owners of small businesses.
Agents now run everyday workflows
Two announcements this week show AI agents doing ordinary, everyday work.
On November 17, 2026, OpenAI introduced ChatGPT for Sales. It includes a new Sales plugin with dedicated skills for sellers and, according to OpenAI, more than 30 new and enhanced partner plugins in ChatGPT. Sellers can also delegate work to their dot, which OpenAI describes as an always-on agent in ChatGPT. The dot can use the same plugins and customer context to prepare meeting briefs, track deal updates, and bring relevant changes and proposed next steps for review.
The same day, Google published "Put Gemini to work for you" about Tasks in Gemini. Google says the update is rolling out free of charge for hundreds of millions of people across the U.S., Japan, Korea, Canada, Australia, and India. In the post, Google's Josh Woodward describes Gemini calling two dealerships at once, sitting through the hold music, booking the shop that accepts appointments, and putting the appointment on his calendar.
Three places agents fit in small business operations
The examples below come from the vendors' announcements. How they map to a small business is our reading.
- Sales follow-up. ChatGPT for Sales can bring together account context from Salesforce, HubSpot, Close CRM, or monday CRM. OpenAI says the Sales plugin gives ready-to-use workflows for everyday sales work, from deciding which accounts to pursue to preparing for a call and drafting the follow-up. Call notes can come from the Meetings plugin in ChatGPT, Gong, Otter.ai, or Granola. Our reading: for a small team, the first use is often the follow-up email and the CRM update after each call.
- Scheduling and phone work. Beyond the dealership calls, Google describes forwarding a batch of school emails to gemini@google.com and having every date on the calendar in less than a minute. Our reading: the same pattern fits booking a supplier visit or a service appointment. Google notes that calling through Gemini is not currently available in Indiana, Louisiana, Minnesota, Montana, and Nebraska.
- Bills and back-office review. In Google's post, Woodward gave Gemini permission to access a utility portal through the new Google Password Manager integration and asked it to analyze past billing statements in his email. Gemini found the cause of a surprise $389 electric bill: a heat pump kicking on around 4 p.m., during peak pricing. Our reading: the business version is checking vendor bills against what you expected to pay.
A cost reality check
Usage fees are the visible cost. OpenAI publishes API prices for its new models in its GPT-6.2 Astra and Sol announcement:
- GPT-6.2 Astra: $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. OpenAI says the input and output prices are 20% lower than GPT-6 Astra's, with a 60% discount on cached input.
- GPT-6.2 Sol: $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens, the same rates as GPT-6.1 Sol.
For both models, OpenAI recommends keeping prompt prefixes unchanged and appending new instructions to the conversation. OpenAI says this improves cache reuse, helps lower costs and latency, and preserves reasoning across turns. Our advice: keep an agent's standing instructions fixed, and add each day's task at the end.
Two costs never show up on a price list. This is our reading:
- Retries. When an agent fails a step and runs it again, you pay for those tokens again.
- Supervision time. Someone has to read the briefs, approve next steps and fix what the agent got wrong. That time is a cost even when the tool comes with a plan you already pay for.
OpenAI's ChatGPT for Sales announcement lists no prices. It says availability of partner plugins and features varies by plan, region, and workspace settings, so check your own plan before you budget.
The governance gap
Agents work through the accounts you connect. The bill example worked because Gemini had password access to a utility portal and could read email. That access is what makes the agent useful, and it is also the risk.
What the vendors offer:
- OpenAI says ChatGPT Business and Enterprise admins enable ChatGPT for Sales by opening the Admin Console, selecting the workspace, and selecting Plugins. Admins can make plugins available for members to install, or install them by default.
- OpenAI says the dot brings proposed next steps for review.
- Google says you can adjust task settings, disconnect Google and third-party apps, or delete chat history anytime.
Our advice:
- An agent should not inherit every permission of the person who set it up. Connect it to the fewest apps it needs.
- Make plugins available to the people who need them, rather than installing them by default for everyone.
- Require a person's approval before anything is sent to a customer, paid, or changed in the CRM.
- Know where the disconnect setting is for every connected app, and test it once.
There is one more reason for approvals. In its GPT-6.2 Astra and Sol announcement, OpenAI reports that in two alignment tests, the models were more likely to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI says it added instructions in Codex that substantially reduce these omissions, and that it is continuing to investigate. Our reading: an agent saying a task is done is a report, and a report still needs checking.
A five-line readiness checklist
Everything in this section is our advice, not vendor guidance.
- One named workflow, with one person who reviews the agent's output.
- A list of every app and account the agent can reach, each with a reason.
- A written rule for what needs a person's approval before it happens.
- A monthly usage budget, plus an estimate of weekly review time.
- A tested way to disconnect the agent from every app on the same day.
Wrap agents in deterministic checks
Our advice: let the agent draft, compare and summarize, and make the steps around it plain rules that give the same answer every time. After an agent updates the CRM, a check confirms the record changed. After it books an appointment, a check confirms the calendar entry exists. After it reads a bill, a check confirms the total matches the source document. When a check fails, the work goes to a person.
OpenAI's own finding is the reason to build it this way. A model may not tell you what blocked it. A simple check will catch it.
Sources
- OpenAI, Introducing ChatGPT for Sales (November 17, 2026)
- OpenAI, Introducing GPT-6.2 Astra and Sol (no date shown on the page)
- Google, Put Gemini to work for you (Nov 17, 2026)
Independent reading; not affiliated with or endorsed by OpenAI or Google.
Reach us at transform@consultiply.ai
Want this working in your business?
Tell us what you are trying to fix and we will point you at the smallest useful next step.
Talk to us