Something changed in the inquiries OpenLoop gets this year. In 2024 people asked for websites. In 2025 they asked if we "do AI." Now, in 2026, they ask for an AI agent — by name, using that exact phrase, often before they can explain what problem they want it to solve.
I run an AI-first agency, so this should be great news. It's half great news. The other half is a conversation I now have several times a month, and I want to write down what I actually say in it.
Why everyone suddenly wants an agent
The marketing this year is relentless. Every report says 2026 is the year of agentic AI — autonomous digital workers that browse the web, write emails, manage workflows, and make decisions without a human in the loop. The stats being thrown around are wild: small businesses cutting operational costs 30% in a quarter, 80% of enterprise apps embedding agents, the whole category growing at 46% a year.
Even here in Srinagar, business owners are seeing this content. A shopkeeper with a two-person operation watches a reel about an AI agent that "runs your entire customer service" and messages me asking how much one costs. I don't blame him. The pitch is genuinely compelling. If the demos were the product, I'd want one too.
But there's a shift happening underneath the hype, and the more honest reports admit it: 2025 was the year everyone talked about AI, and 2026 is the year clients started asking whether it's actually working. That's the question I care about, because I'm the one who has to make it work.
What people think they're buying
When a client says "AI agent," what they usually picture is an employee that costs nothing. Something that answers every customer, chases every invoice, updates every spreadsheet, and never sleeps. The marketing has trained them to expect autonomy — set it up once, walk away, watch it work.
What's actually reliable in production right now is much narrower. The agents that work are the boring ones: a system that reads incoming WhatsApp orders and puts them into a sheet. A bot that answers the same fifteen questions about pricing and delivery times. A workflow that drafts the follow-up email and waits for a human to hit send.
Notice the pattern. Every one of those has a tight scope, a predictable input, and a human nearby. The moment you ask an agent to handle the open-ended stuff — an angry customer, an unusual order, a judgment call about a refund — you've left the territory where the technology is dependable and entered the territory where it's a demo.
What I actually build when someone asks for an agent
Most of the time, what the client needs is a workflow with one or two AI steps inside it — not an autonomous anything. The skeleton is deterministic: this message comes in, this data gets extracted, this record gets updated, this draft gets written. The AI does the parts that used to need a human reading and typing. Plain code does everything else.
A recent example: a local retailer wanted an "agent to handle customer service." What we shipped was a WhatsApp flow that classifies incoming messages, answers the common questions instantly from a small knowledge base, and routes everything else to the owner's phone with a suggested reply he can edit and send. The AI portion is maybe 20% of the system. The other 80% is the unglamorous plumbing that makes it trustworthy.
Is that an AI agent? By the marketing definition, barely. By the only definition that matters — does the owner save two hours a day and trust the thing — completely. He doesn't care what we call it. He cares that it hasn't embarrassed him in front of a customer.
The constraint advantage, again
Building for small businesses in Kashmir forces a kind of honesty about this. My clients don't have an innovation budget. They can't absorb a system that works 90% of the time, because the 10% failure is a real customer walking away, and they know that customer by name. There's no tolerance here for "it's still learning."
That constraint turns out to be a feature. It forces us to build the narrow, supervised, boring version first — the version that actually survives contact with reality. The clients with money to burn are the ones who end up with an abandoned autonomous agent project and a story about how AI doesn't work. My clients end up with a WhatsApp bot that quietly pays for itself every month.
I wrote earlier this year that AI changed about 60% of how we work and left 40% untouched. The agent wave hasn't changed that ratio much. It's changed the packaging. The same discipline applies: use the model where it's strong, wrap it in deterministic code where it's weak, keep a human where the cost of being wrong is high.
What I tell clients now
The conversation has settled into a script. First: tell me the task, not the technology. "I want an AI agent" tells me nothing. "I spend two hours a day answering the same delivery questions" tells me everything. Second: we start with the smallest version that touches real customers safely, and we expand its autonomy only after it earns trust. Third: if the honest answer is that a Rs 3,000 automation without any AI solves the problem, that's what I'll recommend, because the alternative is a refund conversation in six months.
Some prospects walk away from that pitch. They wanted the demo from the reel, and someone else will happily sell it to them. The ones who stay become the kind of clients who refer three more, because the thing we built still works a year later.
2026 really is the year of asking whether AI is working. My answer, from inside a small agency in a place nobody benchmarks: yes — when you scope it like a tool and not like an employee. The agents are coming, probably. The discipline of starting narrow will still be the right call when they do.
If you run a small business and you're trying to figure out what AI can actually do for you — without the hype tax — get in touch: me@mehranshahmiri.com
Related: What AI Actually Changed (and Didn't) · Why I'm Not Raising Money