The AI operations layer: one brain across your business software

An AI operations layer is a set of AI agents that sits across all the software a business runs (field service, CRM, phones, accounting) and watches for work that falls between them. It doesn't replace any of those tools. It reads what each one reports, notices when something that should have happened didn't, and drafts the next step for a person to approve.

For a home-service company, "something that didn't happen" is concrete. A call came in and nobody booked it. An estimate went out and nobody followed up. A job was marked complete and never invoiced. Each of those lives in a different system, which is why no single tool catches them.

AI operations layer vs AIOps

The terms get mixed up in search results, so it is worth separating them.

AIOps is AI for IT operations: monitoring servers, spotting outages, finding the root cause of a failed deploy. It is a tool for IT teams.

An AI operations layer, as we use the term, is about business operations: leads, jobs, invoices, follow-up. It is for owners, office managers and dispatchers. The closest everyday comparison is a very attentive office manager who can see every screen at once.

Why work falls between systems

Most service companies run five to ten pieces of software. A typical stack: ServiceTitan, Housecall Pro or Jobber for jobs and dispatch. A CRM or GoHighLevel for leads and texting. A phone system. QuickBooks for the books. Google Business Profile for reviews. Ad accounts.

Each tool is good at its own job. None of them is responsible for the handoff to the next one.

Larger companies have the same problem at a bigger scale. Okta's 2025 Businesses at Work report found the average company now uses 101 apps. In a MuleSoft survey of IT leaders at large companies, only 29% of their applications were integrated, and 80% said integration problems were slowing them down. If enterprises with IT departments can't connect their tools, a 30-person plumbing company with one office manager has no chance of doing it by hand.

The phone is where it shows first. Invoca analyzed more than 70 million calls and found that only 52% of callers to home services businesses speak with a person. The rest hit voicemail, hold queues or menus. The call log shows the miss. The CRM never hears about it. Nobody calls back.

Why a chatbot in each tool doesn't fix it

Every software vendor is adding AI. That helps inside each product. A field service AI can summarize a job. A CRM AI can draft a text. But each one only sees its own data.

The missed call is in the phone system. The customer's history is in the FSM. The open estimate is in the CRM. The unpaid invoice is in QuickBooks. A chatbot inside any one of them can't see the gap, because the gap is the space between them.

An operations layer reads from all of them. That is the whole idea.

How an AI operations layer works in practice

Under the hood, it works on events and checks.

1. Listen for events. Most modern business software can send a notification the moment something changes. Housecall Pro offers webhooks for more than 50 events, including jobs created and completed, estimates and invoices, though only on its MAX plan. GoHighLevel sends real-time notifications for 50+ events, including new contacts and inbound messages. Twilio, which runs many business phone systems, sends a status callback when each call ends. ServiceTitan's online booking can send an event when a customer leaves contact details but abandons the booking.

2. Compare across systems. The agent holds simple rules a person wrote down. A call ended with no answer and no job was created for that number within an hour. An estimate was sent five days ago and the CRM shows no reply. A job was completed yesterday and has no invoice.

3. Draft, don't act. When a rule trips, the agent drafts the fix: the callback text, the follow-up email, the invoice reminder for the office. A person reviews and sends. What the agent may do alone is a permission you set, one job at a time. We describe that model in human-in-the-loop AI.

4. Report. Once a day, the owner gets one message: what was caught, what was handled, and what is still open.

Newer standards make the plumbing easier. The Model Context Protocol is an open standard for connecting AI models to business tools and data. It also says the host application must get the user's explicit consent before a tool runs, which fits the draft-then-approve approach.

Examples from a home-service week

This is an illustration of how the layer behaves, not a client story.

• Monday, 8:10 AM. Three calls rang out over the weekend. The layer matches them against the FSM, finds two with no booking, and puts two callback drafts in front of the CSR.
• Tuesday. Eleven large replacement estimates have had no reply for a week. The comfort advisor gets a list with a drafted follow-up for each, written in the company's voice.
• Wednesday. A tech closed out a water heater install, but no invoice was created. The office manager gets a flag before the end of the day.
• Every morning. The owner gets yesterday's booked jobs, revenue and open items in one text, pulled from the FSM, QuickBooks and the ad accounts.

None of those steps is clever. The value is that they happen every day without someone remembering to check.

Risks and how to control them

Connecting AI to every system raises fair questions. Who can see what? What if it sends something wrong?

Three rules we follow:

• Least access. Each agent gets only the records its job needs. The estimate agent doesn't need payroll.
• Approval on anything with money or a promise. Pricing, discounts and upset customers stay with people.
• A written risk process. NIST's AI Risk Management Framework organizes this into four functions: govern, map, measure, manage. A small business doesn't need a thick policy, but it should know who owns each agent and how failures are caught.

Where the layer runs is a choice too. Some companies keep it in the cloud. Others want it on a private server. We cover that trade-off in private AI for service businesses.

How to start

Small firms are behind on AI use. The Census Bureau found AI use didn't change significantly among firms with fewer than 20 employees from late 2025 into 2026, while larger firms kept adding it. The good news is that an operations layer doesn't have to start big.

Start with one handoff that hurts. Usually it's missed calls or quiet estimates. Connect the two systems involved, write the rule, and run it with a person approving every draft for a few weeks. Add the next handoff once the first one is trusted.

That is how Cortex builds it: AI agents with defined jobs, working inside the software you already pay for. See which systems we evaluate on integrations and why we work inside existing software instead of replacing it. The free AI Operations Audit maps where work drops in your stack and which handoff to fix first.