Conversations about AI in business almost always start the same way. Chatbot. Product descriptions. Support tickets. All useful. All pretty small compared to what's actually possible.

The more interesting question is not what AI can help with. It's which parts of your business AI can actually execute.

There's a meaningful difference between AI as a tool you reach for and AI as a layer that runs part of your operation. Most businesses are still in the first camp. According to McKinsey, 72% of organizations have adopted AI in at least one business function — up from 55% in 2023. But adopting AI in one function and redesigning your operation around what AI can own are two very different things. The businesses pulling ahead are figuring out the second.

This post is in two parts. First, what AI can execute in ecommerce specifically — the workflows already being automated in stores today. Second, how the same logic applies to any business, with real examples from finance, SaaS, media, and our own operation.

Part 1: AI in ecommerce

In ecommerce, the workflows being handed to AI agents right now aren't experimental. They're operational — running in live stores, owning real parts of the day-to-day.

Inventory and reorder logic

AI monitors stock levels, predicts demand based on seasonality and historical data, and triggers purchase orders — without someone manually checking a spreadsheet every Monday. The agent doesn't just surface the information. It acts on it. Salesforce data shows automated order processing reduces order-to-fulfilment cycle times by 40–60%; the same principle applies to replenishment logic built on the same data.

The value here isn't just speed. It's that the decision happens continuously, based on live data, rather than weekly, based on whenever someone remembered to look. Stockouts and overstock both come from the lag between what the data shows and when a human acts on it. An agent removes the lag.

Post-purchase flows

Not just the confirmation email — the full sequence. Tracking updates, review requests timed to actual delivery, upsell logic based on what was bought, win-back triggered by behaviour rather than a calendar. Each step responds to what the customer actually did, not just what day it is.

Salesforce describes AI taking on the routine work of tracking, collecting customer feedback, sending shipping and tracking links, and managing returns — the entire post-purchase experience, running on triggers rather than schedules. Done well, the customer feels attended to at every step, and no one on your team touched any of it.

Analytics and reporting

Most teams have analytics data spread across four or five tools. Few look at it consistently — not because they don't want to, but because pulling numbers from each source takes time and rarely ends with a clear list of what to do next.

This is one we built ourselves. Our analytics agent pulls data from five sources, analyses it, and delivers a structured report to a Slack channel on a fixed schedule. No one has to run it. It just runs. For Chekly, a SaaS company building tooling that lets Shopify stores offer local payment and delivery methods, it replaced over 8 hours of monthly manual reporting at under €0.05 per run in API cost. The full technical build is documented in that post.

Customer support routing

Not replacing support — routing it. AI handles repetitive tier-one questions, flags the ones that need a human, and ensures nothing falls through. The point isn't full automation. It's putting the right query in front of the right responder, automatically.

Pandora, using AI to automate routine inquiries like order status and product questions, saw a 10% lift in net promoter score. The routine questions get answered instantly; the complex ones reach a human faster because the queue isn't clogged with password resets.

Merchandising and pricing

Which products go in which collection, in what order, based on conversion data. Most stores still make these decisions manually, on a cadence that has nothing to do with what the data is showing. An agent with access to your analytics can make these decisions continuously.

The same applies to pricing. Salesforce describes AI-driven dynamic pricing — adjusting prices in real time based on competitor data, inventory levels, demand, and behaviour — as producing smarter pricing decisions, better conversion, and higher margins, without the manual research cycle that used to sit behind every price change.

The bigger shift in ecommerce

Beyond individual workflows, the customer journey itself is changing. Salesforce reports AI assistant-driven traffic to retail sites grew 119% year-over-year in the first half of 2025. Customers increasingly begin product discovery inside AI assistants — ChatGPT, Gemini, Copilot — before they ever reach a storefront.

During the 2025 holiday season, AI influenced 20% of global online sales, worth $262 billion, and retailers running their own shopper agents grew sales 59% faster than those on the sidelines. This is why product data quality and store infrastructure matter more than they used to — a point we cover in depth in our post on ecommerce content strategy. If an AI can't read your catalogue clearly, it can't recommend your products.

Part 2: AI in the rest of the business

Everything above is ecommerce-specific. But the underlying logic applies to any operation where work is repetitive, data-driven, and structured enough to hand off. The industry doesn't matter — finance, SaaS, healthcare, media, professional services. The question is always the same: which decisions are structured enough to delegate, and which need to stay with a human?

Here's what that looks like in practice, across sectors and inside our own operation.

Finance: document processing at a scale humans can't match

The clearest example of AI owning a real workflow comes from finance. JPMorgan Chase built an AI system called COIN (Contract Intelligence) that interprets commercial-loan agreements — work that previously consumed around 360,000 hours of lawyer and loan-officer time every year. The system reviews the documents in seconds, and it's less error-prone than the manual process it replaced.

What makes this a good example isn't the scale of the hours saved. It's the type of work: high-volume, highly structured, rules-based document interpretation, where the same task repeats thousands of times with small variations. That's the profile of work AI owns well. The lawyers didn't disappear — their time moved from mechanical review to the judgment-heavy work that actually needs a lawyer.

Customer support: the Klarna lesson, in full

Klarna's AI customer service assistant, launched with OpenAI in February 2024, handled 2.3 million conversations in its first month — two-thirds of the company's chats, equivalent to the work of 700 full-time agents. By its Q3 2025 earnings call, Klarna reported the assistant doing the work of 853 agents and saving around $60 million a year, with resolution times cut by 82%.

But the Klarna story has a second chapter that's more instructive than the first. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg the cost-cutting automation push had "gone too far," and Klarna began reintroducing human agents so customers could always reach a person. The AI still handles the bulk of routine volume. What changed is that humans came back for the complex, high-stakes, emotionally difficult cases — disputes, hardship situations, anything where a confident-but-wrong answer causes real damage.

That's not a story about AI failing. It's the clearest public map anyone has of where the boundary sits: AI owns the high-volume, low-ambiguity tier; humans own the value tier; and the line between them is a design decision, not a fixed rule. Every business automating support is now, in effect, working out its own version of that boundary.

Media: recommendation as an operational engine

In media and streaming, AI-driven recommendation isn't a feature bolted onto the product — it is a large part of the product. Platforms like Spotify use AI to curate playlists, surface new releases, and personalise discovery for hundreds of millions of users continuously. No human team could manually merchandise content at that scale or that frequency. The recommendation engine is an operational layer that runs the core experience, adjusting to each user's behaviour in real time.

The lesson generalises beyond streaming: when personalisation at scale is central to your product, it's a workflow AI is uniquely suited to own — provided the underlying data is clean and the feedback loop is well-designed.

Inside Base X Tech: how we actually use AI

We build AI systems for clients, so it would be strange if we didn't run our own operation on them. A few examples of where AI owns or accelerates real work internally:

Client communication health. We use an internal agent to keep an eye on the state of ongoing client conversations — flagging when a thread has gone quiet on our side and needs a response, so nothing slips through during a busy period. It's a safeguard that runs quietly in the background, making sure our communication stays as responsive as we want it to be.

Pre-project store analysis. Before we take on a store, we use AI to do a first-pass competitor analysis and technical pre-audit — surfacing what's underperforming so our team's deeper review starts from evidence rather than a blank page. It replaces part of the manual review labour and gets us to the substantive findings faster.

Analytics reporting. The Chekly agent described in Part 1 started as an internal tool for our own website before we extended it for a client. It still runs for us — pulling our own analytics into a scheduled report so the team sees what moved without anyone running the numbers.

Code review. We use AI to handle part of the code review load — catching the routine issues quickly so our developers' attention goes to the decisions that actually require human judgment. It speeds the process without removing the human checkpoint.

Pitch decks and custom offers. Producing client-facing proposals and presentations is faster with AI handling first drafts and structure, leaving the team to refine the substance rather than build every document from scratch.

Some of these are polished; some are still evolving; others aren't ready to talk about yet. That's the honest picture of AI adoption in a working business — a mix of production tools and works in progress, not a finished suite.

Where most people underestimate the work

Here's what the conversation usually misses, in ecommerce and everywhere else: when AI writes a product description and gets it wrong, you edit it. When AI updates an order status, issues a refund, or changes a record in your CRM — a mistake is an operations problem, not a copy problem.

The governance design is the real work. Permissions. Human approval checkpoints. Rollback logic. Clear boundaries around what the agent does autonomously and what requires a human. Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027 — due to escalating costs, unclear business value, or inadequate risk controls. Not because the technology doesn't work. Because the implementation wasn't designed to handle failure.

AI doesn't fix broken processes. It amplifies them. If your data is fragmented, your systems don't talk to each other, and your workflows are unclear — an AI agent will just make the mess move faster. The Klarna correction is the same lesson from a different angle: the technology worked, but the boundary between what to automate and what to keep human had to be designed deliberately, not assumed.

Three questions worth answering before building any agent, in any industry:

1. Is the workflow well-defined? Can you write down every decision point, input, and expected output — including the exceptions?
2. Is the data clean and accessible? The agent is only as good as what it can read.
3. What happens when it goes wrong? Every consequential action needs a defined failure mode — how errors are caught, who's notified, what gets rolled back.
   Get those three right before writing any code. The agent itself is the easy part.

The question worth sitting with

The competitive gap isn't opening between businesses that have heard of AI and businesses that haven't. It's opening between those who've identified which parts of their operation are ready to hand off — and built the infrastructure to do it safely — and those still treating AI as a tool they pick up when they remember to.

The question isn't "how do we add AI to our business." It's: which decisions in our operation are repetitive enough, structured enough, and safe enough to hand off — and what needs to stay with a human?

That's where the actual work begins. At Base X Tech we build custom integrations and AI agents for ecommerce and beyond — from the analytics agent we shipped for Chekly to the internal tools that run our own operation. If you're working out which parts of your business are ready to automate, get in touch.

FAQ

What is AI integration in business?
AI integration means connecting AI systems — agents, models, or automation layers — to your operational workflows so they can execute tasks, not just assist with them. In ecommerce this looks like automated inventory logic, post-purchase flows, and merchandising decisions. In other sectors it's document processing, support routing, or personalisation at scale. The common thread is that the system acts on data and triggers outcomes without a human initiating every step.

What's the difference between an AI tool and an AI agent?
An AI tool responds when you prompt it — you initiate, it responds, you decide what to do with the output. An AI agent monitors conditions, makes decisions based on defined parameters, and takes actions — including triggering other systems — without a human starting the process each time. Tools augment individual tasks; agents own workflow segments.

What can AI actually execute in an ecommerce store?
AI agents can monitor inventory and trigger purchase orders, run the full post-purchase communication sequence based on customer behaviour, route support tickets between automated handling and human escalation, generate scheduled analytics reports across multiple data sources, update merchandising based on live conversion data, and adjust pricing dynamically. Each is a workflow the agent owns end-to-end.

How are businesses outside ecommerce using AI?
Across sectors, the pattern is consistent — AI owns high-volume, structured, repetitive work. JPMorgan's COIN system reviews commercial-loan contracts that once took 360,000 lawyer-hours a year, in seconds. Klarna's AI handles two-thirds of customer service chats. Media platforms use AI for recommendation at a scale no human team could match. In each case, the routine gets automated and human judgment gets elevated to the work that needs it.

Is it safe to let AI take actions in my business systems?
It's safe when the governance layer is designed alongside the agent — defining what it can do autonomously, what needs human approval, how errors are caught, and what gets rolled back. Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027, largely due to weak risk controls and unclear value — not because the technology fails. Klarna's 2025 rebalancing toward human agents for complex cases is the clearest real-world illustration: automate the volume tier, keep humans on the value tier.

What do I need in place before automating a workflow with AI?
Three things: a well-defined workflow (every decision point, input, and output documented, including exceptions), clean and accessible data (the agent's output is only as good as what it can read), and a defined failure mode for every consequential action. These prerequisites are harder to establish than the agent itself — and skipping them is the most common reason projects fail in production. AI amplifies whatever process it's dropped into, so the process has to be sound first.

How does Base X Tech use AI internally?
We run several internal AI workflows: an agent that monitors the health of ongoing client communication, an AI-assisted pre-project analysis that flags what's underperforming on a store before we start, a scheduled analytics reporting agent, AI-assisted code review, and AI support for producing proposals and presentations. Some are mature, some are still evolving. We build AI for clients, so we run our own operation on it too.

How does Shopify support AI integration?
Shopify's API layer provides access to order data, inventory, customer records, and product catalogues — the foundation agentic workflows depend on. Custom Shopify integrations connect a store to analytics platforms, CRM systems, inventory tools, and AI agents. Shopify's Spring 2026 Edition also introduced the Universal Commerce Protocol, designed to make stores machine-readable for AI agents operating across discovery surfaces.

Base X Tech builds custom Shopify integrations and AI agents for ecommerce operations and beyond. If you're working out which parts of your business are ready to automate, get in touch.