AI Agents
ChatGPT waits for you to type. Cortex agents work inside the systems you already run. They watch for problems, handle follow-up and build your reports, and a person approves anything that needs judgment.
A business AI agent is software with one defined job, access to the specific systems that job needs, and rules for what it may do on its own and what it must hand to a person. In a service company, the jobs are plain: log every missed call and draft the callback, watch estimates that went quiet and draft the follow-up, flag finished jobs with no invoice, and send the owner one report each morning.
The agent works inside the software you already pay for (your field service software, CRM, phone system and QuickBooks). It doesn't wait for someone to type a prompt. It watches for the event, does the routine part, and puts the decision in front of the right person.
A software worker with a defined responsibility, access to specific systems, and rules for what it can do automatically versus what requires approval. Not a chatbot. Not a pile of gimmicks.
CRM, shop or field software, ads, inbox, calendar. Only the records you approve. Named tools are examples we evaluate, not a certified list we pretend already works.
Draft a message. Log a note. Assemble a report. Alert a person. What it can send without you is a permission you set, not a surprise on Friday.
Money, promises, and messy judgment stay with people. If you would not hand it to a new hire on day one, the agent drafts and a person decides.
The job is written before a model is chosen. Permissions are a design choice. The off switch ships with the path.
One workflow. One definition of done. One person who owns what goes wrong. “Watch aging estimates and draft the next follow-up” is a job. “Do AI” is not. If we cannot write it in a sentence a new hire would understand, we do not build it.
Limited logins. Only the access it needs. The estimate agent reads your field software and CRM. It does not get the books, the ads account, or a master login. We confirm what we can connect in the audit. Named tools are examples, not a promise.
Draft a message. Log a note. Assemble a report. Alert Slack. Sending, promising, discounting, and anything that spends money stay behind approval until you trust it. Permissions are set up front, not added later.
We test it on real stalled repair orders, quiet collision files, and old quotes, not a vendor demo account. Our checks confirm the draft sounds like your company, does not invent a promise, and hands off when the file is messy.
Logs, a way to shut it off, and a person who can take the work back. If a vendor ships a breaking API change, managed operations keeps the agent running. If you change your membership plan, we update the agent. We do not leave it running on last year’s rules.
Most of what gets sold as AI to service businesses is a chat window or a simple automation. The real follow-up work doesn't live there.
The work that slips happens when nobody is looking. An aging HVAC quote, a skipped stop on a pest route, a quiet collision file. Those do not type themselves into a chat window. An agent watches the workflow it was built for. A chatbot answers when someone remembers to ask.
If-this-then-that works well when the trigger is clean and the action is the same every time. Open estimates need the last note, the job type, whether the customer already said no, and a call on whether to reach out. That takes context and a permission, not a zap.
One agent that watches unsold installs beats a wall of green lights. We start with the work that costs you most when it slips (usually follow-up or the morning report) and add the next job once you trust the first. Your team still keeps pricing calls, angry customers, and anything you would not hand a new hire on day one.
Most "AI agents for small business" lists mix three different things. Here is how we tell them apart.
| Chatbot | Automation (Zapier-style) | Business AI agent | |
|---|---|---|---|
| Starts when | Someone types a prompt | A fixed trigger fires | An event or a condition across systems |
| Sees | What you paste in | The fields in one step | The records its job needs, in several systems |
| Handles messy input | Yes, but only in the chat | No, it breaks on surprises | Yes, and it escalates when unsure |
| Acts | It doesn't | Runs the same step every time | Drafts or acts within permissions you set |
| Owner of the outcome | Whoever was chatting | Nobody, until it breaks | A named person per agent |
Automations are still the right tool for fixed steps. We wrote more about the difference in AI agents vs automations.
Mostly writing. In the Federal Reserve's 2025 Small Business Credit Survey, firms using AI most often used it for writing or marketing (83%). Only 29% used it for customer service and 21% for process automation. And just 7% of AI users said they had fully integrated it into the business. Smaller firms are further behind: the Census Bureau found AI use didn't change significantly among firms with fewer than 20 employees from December 2025 to May 2026.
That gap is where agents fit. A ChatGPT tab helps one person write faster. It doesn't notice that three calls rang out on Saturday or that a big replacement quote has had no reply for a week. That work lives in the systems, and it usually lands on office staff. The Bureau of Labor Statistics puts the average dispatcher at HVAC, plumbing and electrical contractors at $50,090 a year. Their time is better spent on the calls that need a person than on checking lists.
Be skeptical of the label. Gartner estimates that only about 130 of the thousands of agentic AI vendors are real, and calls the rest "agent washing": chatbots, assistants and older automation sold under a new name. The same Gartner release predicts over 40% of agentic AI projects will be canceled by the end of 2027 because of cost, unclear value or weak risk controls (SD Times covered the release in detail).
We explain the approval model in human-in-the-loop AI, and why we build inside existing software instead of replacing it. See how an implementation project works.
Agent runtime
Watch one worker read the systems you already run, do the approved next step, and stop when a person is required.
$ █
without cortex · someone still has to remember to look
with cortex · the layer watches, drafts, and waits for you
Detail Driven is the named deployment. The other three are common jobs we map in audits, not invented case studies or made-up revenue numbers.
Detail Driven · repair
Example · HVAC
Example · pest control
Example · plumbing
Agents can handle repetitive work across sales, operations, customer experience, marketing, and management, while your team keeps the decisions that require judgment. Yours will differ. That is the point.
Lead qualification. Estimate follow-up. Pipeline monitoring. Customer reactivation.
Job monitoring. Schedule exceptions. Stalled-job alerts. Handoffs between systems.
Routine updates. Review monitoring. Issue detection. Escalations.
Campaign monitoring. Lead attribution. Performance alerts. Review management.
Daily operating summaries. KPI monitoring. Unusual activity alerts. Cross-system reporting.
Software with one defined job, access to the specific systems that job needs, and rules for what it can do alone and what it hands to a person. It acts on events in your systems instead of waiting for a prompt.
A chatbot answers when someone types to it and sees only what you paste in. An agent watches your systems for a condition, such as an estimate with no reply, does the routine part and routes the decision to the right person.
They are when the job is narrow and the leak is real: missed calls, quiet estimates, unbilled work, daily reporting. Small companies feel those leaks most because the same few people answer phones, dispatch and bill.
Ask what its one job is, which systems it reads and writes, what it hands to a person, and who watches it after launch. If the answers are vague, it is likely a chatbot with a new label.
Not until you say so. By default, agents draft and alert you. Sending, promising, and anything that spends money stay behind a person until you trust it.
Built-in AI is made for that vendor’s screen. The work that slips lives between systems: field software, CRM, ads, inbox, books. Cortex agents are built for that work. They do not replace the software you already paid for.
Approvals, logs, and a person in the loop on anything that matters. We treat errors as a design problem: tighten the workflow, add a check, or take that action back to a human.
The goal isn't more agents. It's fewer manual bottlenecks.