
How our AI agents work is one shared loop. An email, call, form, bank feed or finished job starts a run. Each agent reads only the systems its job needs, and a person approves outward steps your rules leave open.
We build six of these loops, one for each pile of repeat work: the Sales, Customer, Admin, Ops, Marketing and Finance Agents. Each one has its own page under AI Agents.
New to the idea? Start with what an AI agent is. The software that builds and tests each agent has its own write-up in how our AI engine works.
How do our AI agents work under the hood?
Every agent runs the same loop: a trigger starts it, it reads the systems it is allowed to see, it drafts or acts, and a rule decides whether a person must approve. Each step lands in a log. What differs from one agent to the next is the trigger, the systems and the approval point.
What starts a run
A trigger is the event that starts an agent, and it is always a record your business already makes.
- A message. A new email, text, web chat or Facebook message.
- A call. A customer phones in, with transcription where enabled.
- A lead. A website, Google, Facebook or form inquiry from a possible buyer.
- A record change. A deal marked won, a booking confirmed or a job marked complete.
- A data feed. Bank and processor transactions as they arrive.
- A schedule. A morning brief on the jobs board or a weekly marketing brief.
What each agent may read
Each agent runs with least-privilege access, which NIST defines as giving each entity only the minimum resources and authorizations its function needs (opens in a new tab).
Passwords, API keys and other secrets are never sent to the model or written to logs. Agents are isolated, so one customer's details never reach another.
We also sell Hardened Agent Templates, a separate security-first product line with these limits built in, as Least Privilege AI Agents: Security Built Into the Setup shows.
Where the models run
A model is the AI engine that reads the text and writes the draft. Every agent runs on private or on-premise models, rather than open public endpoints.
Each part of a job goes to the engine that handles it best. Nothing you give us is used to train AI models.
A newer model reaches an agent only after it passes an evaluation set built from your own past cases. Our guide on whether AI is safe for your business data covers the data side.
Where a person signs off
Prices, promises, payments, filings and anything legal or financial wait for a named person to approve them. You write down, up front, what each agent may do on its own.
A few outward steps run alone inside those rules: first replies to new leads, bookings into open slots, the email classes you enable for auto-send and invoices your rules allow. Everything else waits as a draft.
Those limits matter because agents can be tricked. In NIST's technical blog on agent hijacking (opens in a new tab), the strongest new attack reached an 81% success rate, against 11% for the strongest baseline one.
What the log records
Agents write each action to an append-only audit log, so any later edit is visible. That is what NIST calls an audit trail (opens in a new tab), a chronological record that lets someone reconstruct the events behind an action.
We also make a separate product, the Agent Audit & Risk Dashboard, that turns an agent's log into a risk list and a weekly summary, as Audit Trail for AI Agents: See Every Action and Risk explains.
How a person takes over
A customer can ask for someone on your team at any time, and the agent hands the conversation over with the full thread.
The Customer Agent sends angry, injury or legal messages straight to a named person with the whole thread, then stops replying. In an emergency it pages your on-call person.
If anyone may be hurt or in danger, the only instruction it gives is to call 911. On the jobs side, the Ops Agent waits for your approval, with a pause switch always in reach.
Sales: from first reply to won deal
The Sales Agent answers each new lead from your website, Google, Facebook or forms inside your target response window, after hours included. It qualifies the lead against your published rules, books meetings on the live calendar and keeps every CRM stage current. A person keeps the close, any discount and every change to scope or price.
A single lead moves through five steps, and each one is written to the CRM, the customer record your team keeps, as it happens:
- Read the lead. It classifies the source and the intent.
- Reply. It writes the first message in your brand voice.
- Qualify. Your rules decide fit: service area, job type, budget band, timeline and exclusions. An out-of-area lead is turned down politely.
- Book or nurture. A good fit is booked on the live calendar, inside its rules and buffers. A lead that is not ready starts a follow-up sequence.
- Brief the closer. A hot or exception lead gets a one-screen brief: the source, the ask, the promises made and a recommended next step.
Follow-up on open quotes runs on the schedule you set. It stops on an unsubscribe, a not-interested reply or a hostile one, and the log shows it.
Proposal drafts come only from your approved templates, price list and the facts pack for each job type. Any discount or custom term goes to a person, as does any date promise or off-menu sale.
Our guide to the AI sales agent covers the follow-up side in more depth.
Where does a customer inquiry go?
Calls, texts, web chat, Facebook messages and email all land with the Customer Agent. It answers from your own policy book rather than the open internet, checks live availability and fills open slots inside your rules, including after hours. Anything it cannot place inside those rules waits on a morning exception queue for a named person.
A booking is checked against your service area, the job type and your deposit rules before it lands. It always writes to the shared calendar, never a side calendar.
Before it answers a status question, it checks that the person asking is on the account, then replies from the job record in approved wording. If the record does not say, it escalates instead of guessing.
It explains a bill from the invoice your office sent and routes price questions to the Sales Agent. When the books are involved, it opens a case for the Finance Agent.
It never promises a refund, a price change or a legal outcome, and it never invents a price or an arrival window.
The Admin Agent's inbox and calendar
Every new email gets one of three outcomes from the Admin Agent: filed, drafted for a person or flagged as urgent. It keeps one shared calendar free of clashes and builds documents from your templates plus the job record. A missing field blocks the send, and nothing that commits the business goes out without a person.
Auto-send is the one narrow exception. You pick the message classes, such as confirmations, got-it replies and scheduling, and anything missing from that allowlist stays a draft.
On the calendar, it places holds, moves meetings, protects focus time and keeps travel buffers, so the closer and the crew are never double-booked when your rules forbid it.
It files each finished document, from a work order to an insurance certificate, under a consistent name in the shared folders. It never copies one customer's details into another customer's thread.
It reminds the owner before insurance, license, lease and annual report dates come due. It never files a legal or tax form, changes pay or sends as the owner without a person's approval.
Ops keeps the job record complete
Once work is sold, the Ops Agent runs the path from booked to scheduled, done and invoiced. It checks each booked job for scope, access, photos and timing, then flags missing details before the crew sees the job as ready. Completion on the job board is the only automatic trigger for an invoice draft.

When something changes in the CRM, it takes the next step in the schedule, purchasing, the job folder or team chat. While work is underway, it watches for blockers: missing info, parts, weather, no access and jobs stuck in progress.
It books and moves jobs and sends the right crew by skill and location. When a tech calls in sick, it shows who is open so the crew lead can decide.
Crew judgment, quality and safety stay with your people. Delay notices go out in your approved wording only.
Change orders, purchase orders and vendor bids stay drafts until a person approves the prices. It never moves money and never promises materials missing from your source system.
Marketing drafts, a person publishes
The Marketing Agent keeps an editorial calendar tied to what you sell this month and to the open capacity that Ops reports. It builds posts, short video scripts, ad drafts and landing-page drafts from your brand kit, for channels you already pay for. Nothing net-new goes live without a recorded human approval.
It checks every asset against your logo, colors, fonts and voice examples, and holds back any public claim it cannot back up. Each campaign plan names an audience, an offer, a goal and a stop point.
Spend stays inside your weekly limit, a campaign pauses the moment you say, and only a person can raise a cap. A job type or zip flagged as full for the week drops out of the plan.
It drafts replies to new reviews in your voice. You approve every public reply, and a bad review goes to you.
The weekly brief says what worked, what to drop and which angle Sales should use. It counts campaigns by the leads they hand over rather than by clicks alone.
How far does Finance go on its own?
The Finance Agent categorizes each bank and processor transaction against your chart of accounts, drafts invoices from finished jobs and builds the Monday numbers. It prepares packets for your accountant, sales tax and payroll, but never the filing itself. Refunds, large payments and bank-detail changes need sign-off from the owner and a second person.
It reconciles your bank to your books every day and lists anything that does not match for a person. It flags duplicate vendors, fake invoices, odd refunds and completed jobs that were never billed.
Each categorization and each send gets an audit row: who, what source, when and what changed. An invoice draft starts only when a job is marked complete, and it goes out only when your fitted rules allow.
Past-due reminders stop the moment a reply turns into a dispute, a hostile message or legal language, and a person takes over. Bills are matched to orders and receipts, and a person approves every payment.
Above your two-person limit, money moves only with dual control: a named person plus a confirmation. No code path submits a tax, state or lender filing without a recorded human sign-off.
The Monday pack is built from the live books and job data. It shows cash in, cash out, unpaid invoices by age, completed jobs next to billed jobs and unapplied deposits.
Six agents side by side
Each row below pairs one agent with the event that starts it and the point where it stops for a person. Read the last column as the part of each job that always stays with your team. The systems each agent reads are listed in its own section above.
| Agent | What starts it | Where a person signs off |
|---|---|---|
| Sales Agent | A new lead or a stale deal | The close, discounts, dates and off-menu work |
| Customer Agent | A call, text, chat or email | Refunds, price changes, upset or legal threads |
| Admin Agent | A new email or document request | Anything outside your auto-send allowlist |
| Ops Agent | A confirmed booking or schedule change | Prices on orders, crew judgment and safety |
| Marketing Agent | The month's offer and open capacity | Net-new campaigns, public posts and new spend |
| Finance Agent | Bank transactions or a finished job | Every filing and every payment |
How does work pass between agents?
Most handoffs start on a record change, such as a deal marked won or a job marked complete, and carry a set list of facts for the next step. The agents also share one calendar and one place for filed documents. The chain below follows one sale from campaign to invoice.
- Marketing to Sales. A qualified click or campaign inbound arrives with the offer, the angle, the source and the hook promised.
- Sales to Admin. A booked meeting or a requested proposal sends over the calendar hold and the document pack.
- Sales to Customer. An accepted job arrives with the booked work, the promises made and the qualification notes.
- Sales to Finance. A won deal arrives with its terms and the deposit expectation.
- Customer to Ops. A confirmed booking arrives with the scope, access, photos and timing.
- Ops to Finance. A finished job arrives, with any billable extras flagged for a person to review.
Two side channels run alongside the chain. Ops shares capacity by job type, zip and week with Sales and Marketing, so nothing is sold or advertised that the crew cannot run.
The Customer Agent also passes invoice questions and deposits taken to Finance. The finance side sits at the end of the chain and hands nothing on.
We turn the agents on in three waves. Sales, Customer and Admin come first, because the rest need a calendar and a file. Ops follows, then Marketing and Finance once capacity is published and delivery is trustworthy.
Our look at how an agent team runs daily operations shows the same chain from the owner's side.
Frequently asked questions
Can one agent see another agent's systems?
Agents see each other's systems only where the design shares one on purpose. Access is scoped to each agent's job, and work moves between agents as a set list of facts.
Those facts include scope, access, photos and timing for a booked job. The shared places are the one calendar and the filed documents that the jobs and finance sides both read.
Do the agents send anything without approval?
Automatic sends happen only inside rules you write down first. They cover first replies to new leads, bookings into open slots, the email classes you enable and invoices your rules allow.
Anything net-new in marketing waits for a recorded human approval, and every price, promise or payment waits for a named person.
What happens when an agent is not sure?
An unsure agent asks a person instead of guessing. On a job, it asks when it cannot pick the right playbook.
On a status question, it escalates when the job record cannot answer. After hours, it queues the inquiry for the morning with full context.
Who decides when an agent gets a newer model?
Our team decides, after testing the newer model on your own past cases in shadow beside the current one. The rollout is staged, and the old version stays pinned for rollback.
We report the switch in your scheduled review. Your business rules, such as prices and refunds, change only with your approval.
Next step
Book a free consult meeting and bring the one job your team repeats most. We will sketch its trigger, the systems it needs and where your people approve, then show which agent fits.
Our standards
- The examples in our guides are illustrations, not customer stories.
- Your AI runs on private and on-premise models, not public AI tools. Nothing you give us is used to train AI models.
- Agents are built so each action is logged in a record that shows any later edit.
- Texts go only to people who gave you permission, and every message says how to stop.
- If anyone may be hurt or in danger, the only instruction an agent gives is to call 911.
- Depending on the investment package you choose, you get either the right to use the software we build for you or full ownership of it.
- We work from the Kansas City metro and meet owners face to face or by video call.
- Same team. Same jobs. More capacity.
About the ScaleAlphaSi team
ScaleAlphaSi is a Kansas City AI advisory firm specializing in Advanced Marketing and AI Engineering. Our team drives demand through SEO, AEO, GEO and paid ads management. Once that demand arrives, we build specialized AI agents to scale operations for profitability. Each guide draws on engagements with owners and operations teams, covering where to start, where people stay in charge and how to scale one workflow at a time.
Meet the Team











