Skip to content

    Business process automation

    Intelligent automation vs AI: the part and the whole.

    Intelligent automation vs artificial intelligence: AI is one capability, intelligent automation is the process built around it. Where each fits, how to choose.

    9 min readBy the Uplift team
    A magnifying glass set among wooden gears that carry a paper strip, AI as one part of intelligent automation

    In June 2017 the IEEE published a standard with an unusual purpose. IEEE 2755-2017 is a guide to the words people use for software that runs business processes, written because, in the standard's own framing, the field was so new that there were no common definitions for its concepts, capabilities or terms. A panel of industry participants had to sit down and agree what "robotic process automation," "digital workforce" and "intelligent process automation" meant.

    Nine years later, the question of intelligent automation vs artificial intelligence is that same vocabulary problem with a larger budget attached. Vendors use the two labels interchangeably, buyers inherit the confusion, and a request for "AI" can end up as a bot that copies fields between screens, or the other way round.

    The short version: artificial intelligence is a capability, and intelligent automation is a process built to use it. The rest of this piece explains why that difference changes what you buy, who you need and what can go wrong.

    What is the difference between intelligent automation and artificial intelligence?

    Artificial intelligence is a family of techniques that let software make a judgment it was not explicitly programmed for: reading a messy document, sorting an email by intent, predicting which invoice will be late. Intelligent automation is a process that combines those judgments with rules, system integrations and human review so that a piece of work finishes end to end.

    AI on its own produces an answer. Intelligent automation decides what happens to that answer: which system it is written into, who checks it when the model is unsure, and what the process does next. You can use AI with no automation around it, which is what anyone pasting a contract into a chat window does. You can also run automation with no AI in it, which describes most of the software robots companies bought in the last decade.

    A plain example

    Take supplier delivery notes arriving by email. An AI model can read a scanned note and pull out the supplier, the items and the quantities. That is the AI part, and it takes a few seconds.

    The intelligent automation is everything around those seconds. Code fetches the email, the extracted lines are matched against the purchase order in the ERP, a mismatch above a set tolerance goes to the buyer who owns that supplier, and a clean match is posted and filed. Five or six steps, and only one of them needed a model.

    Intelligent automation is an arrangement, and AI is one of its parts.

    The standard description of intelligent automation is RPA plus AI plus some orchestration, usually with process mining for finding candidates. As a parts list it is accurate, though the parts are far from equal and most of the work sits in the joints between them.

    The parts, in the order a process meets them:

    • A trigger and a route. A schedule, a new email or a changed record starts the work, and a workflow decides the order of steps.
    • Integrations. APIs or, where none exist, screen-level bots that read from and write into the systems the work touches. This is where RPA lives.
    • Judgment steps. The AI: reading free text, classifying, extracting, drafting.
    • Rules. Tolerances, approvals, thresholds, the business logic someone in finance or operations can state in a sentence.
    • An exception queue. A named person who sees what the process could not decide, with enough context to decide it in one look.

    Deloitte's 2022 Global Intelligent Automation survey, which asked 479 executives in 35 countries how they implement and scale these programs, shows how lopsided the mix was. Seventy-four percent were already implementing RPA and 50% were implementing OCR, while only 46% planned to implement AI at all in the following three years (Deloitte, 2022). For a large share of companies, intelligent automation was a program name for bots and document scanning, with the intelligence still on the roadmap.

    Is RPA a form of artificial intelligence?

    No. Robotic process automation follows recorded steps: click here, copy this field, paste it there. It makes no judgment, which is why it breaks when a screen layout changes and why it is reliable when nothing does.

    RPA becomes part of intelligent automation when a judgment step is added next to it, for example a model that reads an attachment before the bot types its contents into a system. The bot itself still only repeats steps; the judgment comes from the model beside it. We cover how the two behave in production, including why scripted bots keep breaking, in AI agents vs RPA.

    Why does the distinction matter when you are buying?

    It matters because the two labels fail in different places, and a buyer who asks for the wrong one funds the wrong fix. If the hard part of a process is a judgment, you need a better model step. If the hard part is the handoff between systems and people, a better model changes nothing.

    Deloitte's data points at the handoffs. The hardest barriers its respondents named were difficulty integrating solutions (62%), a lack of skills and experience (55%) and an inability to change business processes or ways of working (52%). The quality of the AI does not appear on that list. The same survey found that 41% of organizations had no enterprise-wide intelligent automation strategy, and that more than half had not calculated the cost reduction they expected.

    That figure deserves a careful read. It is self-reported, and it counts only organizations that got beyond piloting, in the areas they targeted. It shows that intelligent automation pays once it reaches production, and says nothing about how many programs get there.

    Where the labels get stretched

    McKinsey's State of AI 2025, based on 1,993 participants surveyed in mid-2025, found that 88% of organizations use AI in at least one business function, while nearly two-thirds have not yet begun scaling it across the enterprise (McKinsey, 2025). Those two numbers do not conflict. The first counts any regular use, a team with a chat assistant included. The second counts AI built into how the company runs. The gap between them is roughly the gap between artificial intelligence and intelligent automation.

    Vendors know buyers search for both words. Gartner's June 2025 forecast, which expects more than two in five agentic AI projects to be scrapped before 2028 (Gartner, June 2025), also described "agent washing," the relabeling of chatbots and assistants as agents, and estimated that only about 130 of the thousands of vendors claiming agentic features actually offer them. The same stretch happens with "intelligent." Ask which step in the process makes a judgment, and what happens when it is wrong.

    How much of an intelligent automation should be AI?

    Less than most demos suggest. Every model step is a place where the output can be plausibly wrong, and errors compound: ten chained steps that are each 95% reliable finish correctly 0.95^10 = 59.9% of the time. A process that runs a few hundred times a month cannot carry that.

    The design that survives puts AI only where code cannot do the job, such as reading a note no template covers, and writes everything else as tested code with rules a person can check. At Uplift we build to a ratio of roughly 95% tested code and 5% AI, and the how we build page explains why. We go deeper on the reliability math in intelligent workflow automation.

    Who builds it, and what that costs you

    Someone has to design those steps, connect the systems, write the rules and keep the whole thing working after an ERP update renames a field. Most intelligent automation suites hand that job to your team, which is how a company buys a shelf of fishing rods and finds most of them still in the closet a year later.

    Uplift takes the other route. A team describes a routine in plain language, and we build the automation, run it and maintain it as the apps and APIs around it change. Pricing is flat and set by results, so there are no tokens, no meters and no per-seat licenses. For teams that cannot say where to begin, our Brainstormer looks at each role in turn and proposes candidate routines, learned from over 14,000 real-world automations.

    That model has a real cost, and a buyer should weigh it. There is no codebase handed over at the end; the automations run while the plan runs and stop when it ends. If your company wants an asset it owns outright, with its own engineers to maintain it, a development shop or an in-house build is the better purchase. If you weigh a managed service of any kind, ask the supplier plainly what happens to your processes the day you leave, and get the answer in writing.

    A decision rule you can apply without any vendor.

    Before you buy anything labeled AI or intelligent automation, write the process down as a list of steps and mark each one with one of three letters.

    • J for judgment. A person reads, interprets or decides something that a written rule could not cover. These are candidates for AI.
    • R for rule. A person applies a threshold, a policy or a lookup they could explain in one sentence. These should be code, never a model.
    • H for handoff. Work moves from one system or person to another. These are integration work, and they are where Deloitte's respondents got stuck.

    Count the letters. A process that is mostly J needs a good model and a review queue; buy for model quality and for how errors reach a person. A process that is mostly R and H needs automation with little or no AI in it, and anyone selling you intelligence for it is selling the wrong part. Most real processes come out as many R's and H's with one or two J's, which is the shape intelligent automation was named for.

    The rule has a limit. It assumes you can see the steps, and in many teams the work lives in one person's habits and inbox, invisible until that person is on leave. If you cannot write the list, the first job is finding the work. Our comparison of agentic AI and traditional automation picks up where the J steps start to multiply.

    Frequently asked questions

    Is intelligent automation the same as artificial intelligence?

    No. Artificial intelligence is the capability to make a judgment, such as reading a document or classifying a request. Intelligent automation is a full process that uses that capability alongside rules, system integrations and human review so a piece of work finishes end to end.

    What are the components of intelligent automation?

    Usually a trigger and workflow, integrations through APIs or RPA bots, one or more AI judgment steps, business rules, and an exception queue where a named person handles what the process cannot decide. Process mining is often added to find candidate processes.

    Is RPA considered AI?

    No. RPA repeats recorded steps on screens and makes no judgment of its own. It becomes part of intelligent automation when an AI step is added beside it, such as a model that reads an attachment before the bot enters the data.

    Is ChatGPT an example of intelligent automation?

    On its own, no. A chat assistant is AI that a person uses one request at a time. It becomes part of intelligent automation when a model is called inside a process that fetches the input, checks the output against rules and writes the result into a business system.

    Should I start with AI or with automation?

    Start by listing the steps of one process and marking each as judgment, rule or handoff. If most steps are rules and handoffs, automation with little AI is the right purchase. AI earns its place only on the judgment steps.

    Stop being the middleman. Get an agent that does it for you.

    Tell us the routine. We'll plan it, build it, and run it.

    Questions? Read the FAQ on /pricing, or talk to us.