How Automation Gets Work Done (and where AI fits in)

November 14, 2026 06:25 PM

There's a lot of talk about AI agents doing work that used to take a person. They can browse websites, fill out forms, retrieve information, and finish multi-step tasks.


What gets lost in the excitement is that software has been doing many of these things for decades. What's changing is how computers figure out what needs to happen, how they interact with other systems, and how much judgment we can hand them.


If you've used Zapier to connect two apps, you've already used automation. If you've watched an AI assistant navigate a website and complete a task, you've seen a different kind. Neither makes the other obsolete. To see where AI fits, it helps to know how automation actually works.

Three questions that describe any automation

    1. What starts the work? An event, a schedule, a person, or another automation?
    2. How does it decide what to do? Predefined rules, AI interpretation, or both?
    3. How does it interact with software? An API, a browser, or a tool made available to an AI assistant?

The answers are independent of one another. An AI agent might use an API. A traditional automation might operate a browser. A workflow might use AI to read a customer email, then follow strict rules to decide what happens next.

    the automation landscape

    APIs: software talking directly to software

    An API (application programming interface) gives software a defined way to interact with another system. All an API is, really, is a funny-looking email that one app sends to another. Or think of a service counter. Instead of walking into the warehouse and searching the shelves yourself, you submit a specific request and get a structured response.


    An automation might ask an order management system, "What's the shipping status of order 12345?" and get back the carrier, tracking number, and ship date. Nobody opens the app, navigates to the order, or copies anything off a screen. In promo, commonsku works this way. It has an API, so an automation can look up or update orders without anyone logging in. On the supplier side, PromoStandards web services play a similar role for things like order status and shipment notifications. Because the interaction follows a defined structure, APIs are generally among the most reliable ways to automate a process.

    Where Zapier fits

    Zapier (and Make, n8n, and others) is an automation platform that connects apps and coordinates workflows. When a Zap creates a contact in Zoho CRM, it usually talks to Zoho's API behind the scenes. Zapier can also use webhooks, code, and other methods, so it's an orchestration platform rather than a single interaction technology. Many promo tools plug in directly. Dynamic Mockups, for example, has a Zapier integration, so a Zap can request a product mockup as one step in a larger workflow.

    What about webhooks?

    An API call usually asks a system to do something or return information. A webhook lets one system tell another when something happens. When a customer places an order, the ecommerce platform can send the order data straight to an automation.

    It's the difference between calling a supplier to ask whether your order has shipped and getting an automatic shipping notification. Webhooks let you start workflows without repeatedly checking for updates.

    RPA: software operating the interface

    What if an app has no API that supports what you need? One option is robotic process automation, or RPA. There's usually no physical robot involved. RPA software works with applications much as a person would: it clicks buttons, types into fields, navigates menus, and copies data between systems.


    Say an accounting app requires someone to key in invoice details. An RPA bot could open the app, go to the invoice screen, fill in the fields, and save the record.


    Traditional RPA follows predefined instructions and doesn't necessarily understand the business purpose of what it's doing. If a button moves or the app changes, the automation may break. Automating a person's actions isn't the same as understanding their decisions.


    Newer RPA tools are adding AI features, such as reading invoices and pulling data from scanned forms. The core idea is still the same, though: repeat steps that someone defined ahead of time.


    RPA also doesn't have to be expensive. Zoho RPA is a lower-cost option with a free plan and a free trial, and it works with other Zoho apps. If you already live in the Zoho ecosystem, it's worth a look for jobs that have no API.

    Browser automation: related, but different

    Browser automation is similar to RPA but limited to websites and web apps. Tools like Playwright and Puppeteer control a browser programmatically. They can navigate pages, click elements, fill in forms, download files, and extract information.


    A promotional products distributor, for example, might need data from a supplier portal that has no suitable API or PromoStandards service. A browser automation could log in, find an order, read its status, and pass it to another system.

    Where cloud automation platforms fit

    You don't have to run browser automation on your own computer. Some platforms host automated programs for you in the cloud, so they run on a schedule or on demand without a machine of your own to keep running. Apify is one example that we use. Its programs are called Actors.

    These programs aren't all browser automation. Some drive a browser to work with websites. Others use APIs, HTTP requests, or other techniques to collect and process information. Browser automation and web scraping are common uses, but the category is broader than either.

    They're also a handy way to fill gaps. If an app's Zapier integration lacks a trigger or action you need, a hosted program can automate the app directly and hand the result back to your workflow.

    Is browser automation the same as scraping?

    Not quite. Browser automation describes how software interacts with a website. Scraping describes pulling information out of one. A browser automation might submit an order without extracting anything. A scraper might fetch product details without ever opening a full browser. They overlap often, but the terms aren't interchangeable.

    AI computer use: AI operating the interface

    Traditional browser automation follows instructions written in advance. AI computer-use systems can interpret a task, look at an interface, and decide what to do. Instead of scripting every click, you might say, "Find the three orders that haven't shipped and prepare a summary for the customer." An AI system with computer-use capabilities could navigate the order management interface, find the relevant information, and work through the task. Claude and other AI systems can do this, depending on the tools and environment available to them.


    The difference isn't just that AI can click buttons. It's that AI can decide which buttons to click based on what it sees and what it's trying to do. That flexibility helps when tasks aren't fully predictable.


    It also adds uncertainty. AI can misread an interface, pick the wrong record, or take an action nobody intended. Interfaces change, and computer-use systems can be slower and less predictable than structured integrations. For a repeatable process with a reliable API, having AI navigate the interface may add complexity instead of removing it.

    four ways to interact with software

    MCP: giving AI access to tools

    You may have started hearing about MCP, the Model Context Protocol. It's an open standard that lets AI applications connect to external tools and information sources through a common interface.


    Say you ask an AI assistant to pull customer information from your CRM. Without a suitable integration, it may have no way to get it. With an MCP server exposing the right tools, the assistant can discover what's available and request that a tool be used. An MCP server might offer tools to search for a customer, retrieve open sales orders, look up shipment information, or create a follow-up task. The assistant picks the one that fits the request. Servers can also expose data and reusable prompts, not only tools.

    Does MCP replace APIs?

    No. An MCP server often calls existing APIs to do the actual work. The API connects to the underlying system, and MCP gives AI applications a standard way to discover and use that capability. Think of an API as a service an app offers, and MCP as a common way for AI assistants to find and request those services.


    MCP doesn't make an automation intelligent. The model supplies the reasoning, and the MCP tools supply the capabilities.

    WebMCP: websites exposing tools to AI agents

    WebMCP takes a related idea into the browser. Instead of making an AI agent work out a site by inspecting and clicking its interface, a website can expose structured tools that describe what actions are available.

    Imagine a supplier site offering a tool called checkOrderStatus. An agent could discover it, pass in an order number, and get the result without clicking through the site's order lookup screens. The website still controls what it exposes. That could make interactions with websites more structured and predictable for agents. Unlike a traditional API integration, these tools are registered by the page itself and called by an agent working in that browser.

    WebMCP is early. It's a draft proposal from a W3C community group, not a finished standard, and so far Chrome is the only browser testing it, through an origin trial. Don't assume any given website supports it, and expect the details to keep changing.

    The distinction that matters

    An agent could get an order status several ways:

      • Call an API directly.
      • Use an MCP tool that calls the API.
      • Use a WebMCP tool the website exposes.
      • Follow a scripted browser automation.
      • Navigate the website visually with AI computer use.

    The result might be identical. The mechanism isn't, and the mechanism affects reliability, security, maintenance, and cost.

    Deterministic automation vs. AI interpretation

    So far we've covered how automation interacts with systems. The other question is how it decides what to do.

    Deterministic automation

    A deterministic workflow follows explicitly defined rules, such as "If the order total exceeds $500, send it for approval." The condition and the action are clear, and given the same inputs and rules, the workflow is designed to produce the same result. Much of traditional business automation works this way. It's predictable, testable, and usually easy to audit.

    AI interpretation

    Instead of relying only on predefined rules, an AI model interprets information and produces a response, classification, recommendation, or decision. For example: "Read this customer's email and determine whether they're asking about shipping, pricing, or a product problem."


    Customers don't all use the same words. One writes, "Where's my stuff?" Another asks, "Do you have an ETA on those polos?" A model can recognize that both are about order status, which a simple keyword rule wouldn't catch. That makes AI useful when information is ambiguous, unstructured, or hard to capture in fixed conditions. But its outputs can be wrong or inconsistent.

    Hybrid automation: often the best of both

    Consider this workflow:

      1. A customer emails asking about an order.
      2. AI identifies the intent and pulls out an order number.
      3. An API retrieves the order.
      4. Deterministic rules confirm the order belongs to the customer.
      5. AI drafts a reply from the verified shipment information.
      6. The reply is sent automatically or routed to a person for approval, depending on the workflow's rules.

    AI handles the language and interpretation. Traditional automation handles the structured operations and the safeguards. This is one of the most useful ways to bring AI into an existing business process.
    Rules, Reasoning, and Hybrid Automation

    What actually makes something an AI agent?

    The term "AI agent" gets attached to almost anything involving a language model. But calling a model once isn't the same as giving it agency.

    Suppose an automation sends an email to a model with the instruction "Classify this message as a quote request, order inquiry, or complaint." The model returns a classification and a predefined workflow handles the result. That's AI-powered automation, but the model isn't deciding how to complete a larger task.

    Now suppose an assistant is told, "Investigate why this customer's order is late and prepare an update." It might look up the order, check the shipment history, identify the supplier, retrieve the supplier's latest status, decide whether it needs more information, and write a response from what it found.

    An agent can choose actions, use tools, look at the results, and decide what to do next, within the capabilities and instructions it's been given. It doesn't necessarily need a browser. It might use APIs, MCP tools, computer use, or a mix. And it doesn't necessarily work without oversight, since an agent can be required to ask for human approval before certain actions. What defines it is that it decides which steps a goal requires, instead of running a fixed sequence.

    One workflow, several technologies

    Here's a familiar promotional products scenario. A customer emails a distributor: "Hey, any idea when those shirts are going to arrive? We need them for Friday."

      1. An event starts the workflow. The incoming email triggers the automation through an email integration or webhook.
      2. AI interprets the message. The model recognizes a delivery status question and picks out any useful order details.
      3. The automation retrieves the order. An API, such as commonsku's, searches the distributor's order management system.
      4. Additional information is collected. If the supplier's shipment status isn't available through an API or a PromoStandards service, the workflow might use browser automation to get it from a supplier portal. Or an AI agent might use an available MCP tool or computer use.
      5. Rules decide what happens next. If the order has shipped and tracking confirms delivery before Friday, the workflow proceeds. If the date is uncertain or the order looks delayed, it escalates.
      6. AI drafts the response. Using the verified information, it writes a clear customer update.
      7. A person approves when needed. Routine inquiries can be answered automatically. Exceptions go to a coordinator for review.

    No single technology does all of this. The value comes from combining them.

    One Order Inquiry, Multiple Automation Technologies

    A quick word on security

    The more freedom you give AI, the more it matters what it can touch. One risk is worth knowing by name: prompt injection.

    AI works by reading text, and it can't always tell your instructions apart from instructions hidden in the content it's reading. A web page, an email, or a document can contain hidden text telling the AI to do something else, like send data somewhere or click the wrong button. With a chatbot, that might mean a bad answer. With an agent connected to real tools, it can mean a bad action.

    This applies to AI computer use, MCP, and WebMCP alike, because all three put an AI model in contact with content and actions you don't fully control. Fixed rules don't have this weakness. They don't take instructions from the content they process, which is one more reason to keep them in the loop.

    A few habits reduce the risk:

      • Give the AI the least access it needs. Start with read-only.
      • Require a person to approve anything that sends, deletes, pays, or changes a record.
      • Use fixed rules for checks that must never fail, like confirming an order belongs to the customer.
      • Connect only tools and servers you trust.
      • Keep a log of what the AI did.

    Which method should you use?

    There's no universal winner. This table is a rule of thumb, not a guarantee.

    Which method fits? A rule of thumb, not a guarantee.
    MethodWhat it doesReliabilitySetup effortBest for
    API or webhookSoftware talks directly to softwareHighLow to mediumRepeatable tasks when an API exists
    RPAA bot repeats the clicks and typing a person would doMedium (breaks when screens change)MediumApps with no API
    Scripted browser automationA script drives a browser step by stepMedium (breaks when the site changes)Medium to highWebsites and supplier portals with no API
    AI computer useAI looks at the screen and decides what to clickLower and less predictableLow to start, but needs close oversightMessy, variable tasks nobody has scripted
    MCPGives AI assistants a standard way to use your toolsDepends on the API underneath and the AI's choicesMediumLetting an AI assistant look things up or take actions in business systems
    WebMCPWebsites offer tools directly to AI agents in the browserPromising, but earlyDepends on the site supporting itA future option, not widely available yet

    A simple way to choose

    which method to choose

    Here's the same thing in words. First, decide how the work will reach the system. Take the first answer that fits.

      1. Does the system have an API or webhook that does what you need? Use it. It's the most reliable and the easiest to maintain. commonsku, or a supplier that offers PromoStandards services, are examples.
      2. No API, but the task is the same every time? Use browser automation for a website, or RPA for other apps. Zoho RPA is worth checking if budget matters.
      3. No API, and the task varies or needs judgment about what's on screen? Try AI computer use, with a person checking the results.

    Then add whatever the process needs on top:
      1. Do you want an AI assistant to use your business systems when asked? Add MCP, usually on top of the API.
      2. Does a decision have to follow strict business rules? Keep that step as fixed rules, whichever method moves the data.
      3. Is the input messy human language, like customer emails? Use AI to interpret it, then hand off to rules and a person where it matters.

    The goal isn't to use as much AI as possible. It's to use the right kind of automation for each part of the process.

    The bigger picture

    More business software can now be operated by AI, not just integrated through traditional automation platforms. That's a real shift, but it doesn't erase the value of APIs, webhooks, RPA, or the deterministic workflows businesses already depend on. It adds options.


    The most capable systems won't be the ones that use AI for everything. They'll be the ones that know when to use AI, when to follow rules, and how to interact with each system reliably. For distributors, suppliers, and other businesses trying to make better use of automation, understanding those distinctions is a good place to start.

    Eric Granata

    Eric Granata

    Managing Director PromoPilot, LLC
    https://www.linkedin.com/in/eric-granata/

    Eric Granata is the founder of PromoPilot, helping print and promo distributors automate workflows, streamline e-commerce, and maximize efficiency using no-code tools like Zapier. With over a decade of distributor experience, Eric shares insights on automation, tech, and scaling smarter.