Go Back

How Many AI-Agents Work For You?

by Alim Uderbekov

Originally published on Substack

There’s a workflow I trace with every retail program leader I meet. We start with one store opening, just one, and we count.

• Roughly 100,000 coordination messages.
• 67+ signed documents.
• 30+ stakeholders across as many companies.
• One million mouse clicks across a dozen systems.
• Zero agents.

That’s the count for one store. Multiply it by your annual program, 50, 100, 200 stores, and you have the operating reality of store development in 2026.

The title of this piece is a literal counting exercise. Walk one store opening end to end and count the agents doing autonomous work alongside your team. The number is almost certainly zero. That isn’t an indictment. It’s the starting line.

One workflow, today

Before going any further, let’s look at one workflow every retail real estate team runs: facility management.

• A store manager notices water dripping through the ceiling near the back of the sales floor.
• They open ServiceChannel (or any other soft) and type a ticket.
• The facility manager reads it and starts the triage.
• Who’s responsible? Most leases put structural and roof on the landlord. HVAC routine is on the tenant. Warranty might cover it if the unit is recent.
• They pull the lease to check. They check the warranty list.
• They draft an email to the right party. They send it.
• They track the response.
• They follow up when nobody replies.
• They close the loop.

Nine Steps! For one ticket. Now multiply by 80 stores.

If this sounds familiar, it’s because this is how most retailers run it today. The systems are good. ServiceChannel, Fexa, Corrigo they all do exactly what they were designed to do. They route tickets to people. The bottleneck isn’t the software. It’s that every ticket still requires a person to read it, decide, draft, send, and chase.

A century-old operating system

Now zoom out from the ticket. Look at a full store program, the kind with milestones, dependencies, multiple trades, a permitting timeline, a soft-opening date that can’t slip. What you see is a Gantt chart. Tasks plotted across a horizontal timeline, dependencies drawn between them, a critical path running through the middle.

The Gantt chart was invented around 1910 by Henry Gantt. The Critical Path Method came in 1957. PERT in 1958. Every project management concept your team uses today was invented before the transistor. Procore, Lucernex, Tango, they all digitized these methods. None of them replaced them. The work being managed hasn’t changed in a century.

Except for one thing – AI-agents are arriving.

Why software is the right teacher

That early-twentieth-century moment Gantt, Taylor, the rise of scientific management is when project management became a discipline. From there, it developed in every industry that produced projects. Construction. Manufacturing. Aerospace. Defense.

But one industry has produced more projects than all the others combined. By orders of magnitude. Software.

Project management is a skill that compounds with reps. The more projects an industry runs, the sharper its instincts become about how work flows, where it stalls, and what unblocks it. Software has run more projects in the last 30 years than the rest of the world combined ran in the previous 100. That makes it the natural place to look when something new shows up in how projects get done.

Even SpaceX, the company most associated with disrupting a physical industry, runs program management closer to a software shop than a traditional aerospace prime. Iterative build-test-learn loops [source]. A flat hierarchy. A proprietary internal ERP they wrote themselves. Boeing runs waterfall. SpaceX ships rockets like software.

What just happened in software

In October 2025, the CEO of Linear, a tool used by most modern software teams to manage their projects, published a short post titled “Issue tracking is dead.”

The evidence behind the statement: coding agents are now installed in 75% of Linear’s enterprise workspaces. 25% of new issues are being authored by agents, not humans. Agent-completed work grew 5x in three months.

This isn’t a forecast. It’s the operating reality of software engineering in 2026.

The role shift behind the numbers matters more than the numbers. The PM’s job in software has moved from tracking status to orchestrating AI-agents and quality of inputs from humans. So the work has moved up the value chain from coordinating execution to engineering coordination.

A mobile app and a retail store

A mobile app has a frontend, a backend, infrastructure, a release process, and a maintenance cycle.

A store has a storefront, MEP, fixtures, a turnover process, and a facilities cycle.

Both are built by people in different companies, working from different systems, on different schedules, who have to agree on what’s true at the end of every week.

The shape of the work is the same. The tools we use to manage it shouldn’t be a century apart.

Now look at that same ticket — with agents

Let’s walk back through the facility ticket. Same store. Same leak. Same systems underneath. But this time, watch what happens at each stage.

Stage 1. The intake agent

The store manager doesn’t open ServiceChannel and type a structured ticket. They speak, or type a single line: “Water coming through the ceiling near the back stockroom.”

An agent, sitting on top of ServiceChannel, not replacing it, asks the right follow-ups. → How long has it been happening? → Is it active right now? → Is the floor wet?

It classifies the issue. It attaches the photo the store manager just took. It formats the ticket for the facility team.

This is already happening. ServiceChannel shipped intake AI in late 2025. Fexa has a version. Most companies that have deployed AI in facility management have stopped here and called it done.

It’s a real improvement. But it’s one agent. Single shot.

Stage 2. The routing agent

Now a second agent sits with the facility manager. It reads the incoming ticket, suggests who’s likely responsible, and drafts the email. The facility manager reviews, edits if needed, and sends.

Two agents now. The store manager’s intake is faster. The facility manager’s triage is faster. Fewer keystrokes. Fewer errors.

The ROI on these two alone isn’t enormous. But the asset you’ve built is. Every interaction those agents have with your team is data. They’re learning how your facility manager actually thinks. Which vendors get certain trades. Which landlords need a specific tone. The agents become more useful with every ticket, not because the model got smarter, but because the context did.

Stage 3. The lease check agent

Here’s where it gets interesting.

The routing agent doesn’t actually know who’s responsible. It guesses based on patterns.

But responsibility in retail facilities isn’t a pattern – it’s a clause.

Most leases put structural and roof on the landlord. But if the leak is from a tenant-installed HVAC unit that punctured the roof, suddenly it’s the tenant’s problem. And every lease is different. Every amendment is different. Every store is different.

This is where context becomes the unlock.

A third agent reads the actual lease for that specific store. It finds the relevant clause. It tells the routing agent:

“Roof penetrations are landlord-responsibility. The HVAC unit causing this leak is tenant-installed under Amendment 3. Repair cost is split, landlord covers structural patch, tenant covers HVAC.”

Three agents. And now the facility manager isn’t relying on memory or a half-updated lease abstract. They’re working from the actual lease, parsed in seconds.

The same data your facility manager already has access to is what makes this agent possible. You don’t need new data. You need an agent that can read the data you already have.

Stage 4. The dispatch agent

Context isn’t enough. The dispatch agent needs skill, knowing what to do with the context.

It pulls the right contact for that landlord at that property. It drafts an email in the tone that landlord responds to (some want detailed photos and lease references; some want a one-liner). It attaches the right documents. It cites the right clause. It sets a follow-up.

If the landlord doesn’t respond in 48 hours, it sends the chase email. If they push back on responsibility, it surfaces the disagreement to the facility manager with the relevant lease language already pulled.

Four agents.

Stage 5. The orchestrator

You now have four agents. They don’t talk to each other.

The intake agent doesn’t know the lease check agent exists. The routing agent doesn’t know the dispatch agent’s contact list. Each one is good at its job, but the workflow still depends on a human stitching them together.

The orchestrator is the fifth agent, the one the store manager actually talks to.

The store manager drops a single line: “Roof leak. Back stockroom.” The orchestrator takes it from there. It pings the intake agent for follow-up questions. It hands the ticket to the routing agent. It calls the lease check agent for the responsibility answer. It feeds all of that to the dispatch agent for the email. It tracks the response and reports back.

Agent talks to agent. Agent works with agent.

The facility manager sees the result, not the process.

The reframe

Your facility manager’s job has changed.

They’re no longer processing tickets. They’re designing how the agents handle them. They’re an agent behavioral engineer or agent maestro, defining the logic, the escalation rules, the tone, the exceptions. They navigate the agents through the logic, not through the actions.

You don’t need an army of facility managers anymore. You need one who understands how the agent stack works and can shape it.

One workflow. Five agents. Working 24/7. And that’s just facility tickets.

The bigger thing nobody is talking about yet

Here’s what most teams miss when they think about agents.

The agents don’t just do the work. They remember it.

Five facility tickets become a hundred. A hundred become a thousand. And every one of them is now structured data: which HVAC unit, in which store, in which climate zone, under which landlord, with which response time, at which cost. The facility agent isn’t just closing tickets. It’s building the most accurate operating dataset your portfolio has ever had.

Now imagine what your real estate team can do with that.

Your deal-maker is negotiating a new lease. The site has a 20-year-old rooftop HVAC unit. Deal-maker’s AI agent, the one helping them evaluate the deal, has access to the same data the facility agent has been collecting for years.

It surfaces a quiet warning:

“This HVAC model fails on average within 18 months. There’s no certified service provider within 200 miles of this market. Estimated cost-of-ownership over the lease term is $340K higher than your portfolio average.”

That’s not a facility insight anymore. That’s a deal insight. It changes how you negotiate the TI allowance. It changes which obligations you push to the landlord. It might even change whether you sign the deal.

The pattern repeats across every workflow in store development. The bid leveling agent learns which GCs over-promise on schedule. The permit agent learns which jurisdictions move slowly in Q4. The construction agent learns which stores opened under which conditions hit their first-year revenue targets.

Agents don’t just coordinate. They turn coordination into intelligence. The facility manager becomes a data source for the dealmaker. The dealmaker becomes a data source for the construction PM. Every workflow feeds the next. That is the part that changes the industry, not the speed, the intelligence.

Why now

You’ve probably tried this before. A pilot in 2023 or 2024 that underwhelmed. A demo that looked great and a deployment that didn’t work. The pilot was right to fail. The infrastructure wasn’t ready. Models hallucinated on long documents. Tool integrations were brittle. There was no orchestration layer.

Then between mid-2024 and late 2025, model performance on real-world software engineering tasks went from 1.96% to over 80% on the SWE-bench benchmark. Reasoning models replaced next-token prediction. Standard protocols replaced custom integrations. Agents can now run for hours on a single task, recover from their own mistakes, and call other agents to fill in what they don’t know.

If you tried this 18 months ago and it didn’t work, you owe yourself a second look. Not because the vendors are better. Because the ground moved underneath you.

What isn’t changing

Steel-toes. Concrete cures on its own clock. Permits wait their turn. Inspectors sign on their schedule. The store opens because someone in the field did the work.

This article isn’t about the work that opens the store. It’s about the layer above it, the layer that decides whether the field gets to build the right store, on time, with the right information.

What isn’t changing

Steel-toes. Concrete cures on its own clock. Permits wait their turn. Inspectors sign on their schedule. The store opens because someone in the field did the work.

This article isn’t about the work that opens the store. It’s about the layer above it, the layer that decides whether the field gets to build the right store, on time, with the right information.

So how many agents?

Right now, the honest answer is almost zero.

In two years, that changes. Not just in retail real estate. In every industry where work is mostly the coordination of people, documents, and decisions across companies that don’t share systems. The teams winning that shift won’t have bigger departments. They’ll have smaller, sharper ones, designing how a workforce of agents handles the work they used to do by hand.

The good news: it’s day one. You haven’t missed it. You’re early.

If this resonated

We work with retail real estate, store development and facility teams who are exactly in this position, running complex programs on spreadsheets and disconnected systems, and asking how to start hiring their first agents.

If you want to see what this looks like in practice, on workflows that look like yours, we’d love to show you.

Reach out at m@surfaice.pro or connect with me directly on LinkedIn.

Alim Uderbekov is the CEO and co-founder of Surfaice, an AI platform for retail store lifecycle.