What Is AI Automation? Examples, Costs and How It Differs from RPA

What is AI automation? It is software that handles work involving judgement, such as reading an email or a document and deciding what happens next. Rules and people around it keep it safe. Ordinary automation only follows the steps it was given. AI automation copes with the untidy inputs that most office work arrives in. That is why jobs like sorting a shared inbox or checking an incoming document have only recently become practical to automate.

This article gives a plain answer for a business owner. It covers how AI automation differs from RPA, what it is good and poor at, seven examples and what it costs. Our AI agents start from £5,000. It closes with the risks to plan for and how to start with a single process.

Key takeaways

• AI automation handles work that needs judgement, such as reading emails and documents, with rules and a person around it. If a process runs the same way every time, ordinary automation is cheaper and steadier.
• RPA copies fixed, rule-based steps. AI automation can deal with input that varies, and most working systems combine both.
• Off-the-shelf tools start free, then from US$19.99 a month (Zapier, annual billing). With us, an AI readiness assessment is £3,000 and a first agent starts from £5,000.
• The main risks are wrong answers, untrusted input and personal data. Each has a design answer: review points, narrow permissions and a clear view of what the system may decide alone.
• Start with one frequent, text-heavy process that has a clear right answer, and run the new system beside the old one before switching.

What is AI automation, in plain terms?

AI automation puts an AI model inside an automated process, so the process can read, classify, extract, draft or decide instead of only moving data from one place to another. Definitions vary by vendor. AWS defines it this way: “AI automation is the process of using artificial intelligence (AI) technologies to expand the capabilities of robotic process automation (RPA) technologies” (AWS).

Salesforce words it differently: “AI automation uses advanced technology to manage tasks and processes by programming computer systems to review data, recognise patterns, and make logical choices” (Salesforce UK). Both are vendor definitions, written to fit what each company sells. They agree on the shape: a machine does work that needs some interpretation of data.

Our working definition for a business owner is software that does a job involving judgement or untidy input, inside rules you set. It hands a person whatever it is unsure about. Three things mark it out:

• The input is untidy. Emails, PDFs, photos and free text, rather than a neat form field.
• The software makes a judgement call. It classifies, extracts, matches, drafts or decides.
• A rule or a person checks the result when the stakes or the uncertainty are high.

If one of the three is missing, you may not need AI. A job with tidy inputs and fixed rules is ordinary automation, and it is usually the cheaper and sturdier answer.

You are not early. In the ONS’s Business Insights and Conditions Survey, 30% of businesses reported use of at least one type of AI technology in September 2026 (ONS). The ONS puts it at 38% for businesses with 10 or more employees and 29% for those with fewer than 10. The figure covers any AI technology, so it says little about automation specifically, but it shows how ordinary AI use has become.

How is AI automation different from RPA?

RPA follows instructions written in advance, while AI automation can interpret input that varies. UiPath, an RPA vendor, describes the older approach like this: “Robotic process automation (RPA) uses software robots to automate repetitive, rule-based tasks like data entry and system integration” (UiPath).

Rule-based workflow automation, the “when this happens, do that” kind, works the same way on structured events such as a form submission or a status change. Both are quick and predictable while the input matches what was expected. When a supplier changes an invoice layout, or a customer puts three questions in one email, they stop or give a wrong result.

Approach	Works on	Struggles when
Rule-based workflow	Forms, dates, status changes	Input needs reading
RPA	Repetitive, rule-based tasks	The situation leaves the script
AI automation	Emails, documents, messages	A wrong answer goes unchecked

The three overlap in practice. AWS’s definition above frames AI automation as an extension of RPA. AI workflow automation is the usual meeting point: a rule-based chain with an AI step where judgement is needed. Rules handle dates, sync records and send messages. The AI reads and classifies, and a person approves what the system is unsure about.

For the rule-based side in depth, see our workflow automation guide. An AI agent is the more autonomous end of AI automation, free to choose its own next step, and what an AI agent is has its own explainer.

How does AI automation work?

A working AI automation has four parts.

• A trigger. Something starts the work: an email arrives, a form is submitted, a contract end date comes close.
• An AI step. A model reads the input and does the judgement work. It classifies the message, pulls out the fields, drafts the reply or matches the record.
• Rules and checks. Ordinary code validates what the model produced, limits what it may touch and keeps a log of what happened.
• A person. Anything the system is unsure about, or that is costly to get wrong, goes to someone for approval.

Our work for a UK energy broker, SBU, shows all four. A headless automation layer sits across Salesforce, an LLM (Anthropic Claude) and the client’s email. Outreach is triggered by each customer’s contract end date. Replies are classified and responses are drafted for approval, and an 85% confidence threshold decides what runs automatically and what waits for a person.

Within three months the team was generating five times the qualified leads, without adding headcount. The threshold is the key design choice in a system like this, because it sets how much runs alone and how much a person sees.

What is AI automation good at, and where does it fall short?

It is good at work built on language and documents, and a poor fit for work that needs an exact answer every time.

Good at:

• Reading and classifying emails, tickets and messages by topic, urgency or intent.
• Pulling fields out of documents that vary in layout, such as invoices and forms.
• Drafting first versions of replies, quotes and summaries for a person to approve.
• Matching records that do not line up exactly, such as two spellings of the same company.

A poor fit:

• Exact arithmetic and lookups, where ordinary code is cheaper and always right.
• Decisions where a wrong answer is expensive and nobody reviews it.
• Jobs with no agreed definition of “correct”, so nobody can tell whether it worked.
• Rare tasks, where the build cost never pays back.

Models can give a confident wrong answer. That is why our agents run with clear limits on what they decide alone, and with human sign-off where it matters. If your process runs the same way almost every time, we will tell you to use ordinary automation, and sometimes to buy a tool.

What are examples of AI automation?

Here are seven, by function. Four come from systems we have built. Three are common patterns, labelled as such.

1. Sales follow-up. The energy broker system above triggers outreach from contract end dates, classifies the replies and drafts responses for approval.
2. Quoting. ChatQuote is an AI quoting assistant we built for service businesses. It drafts a quote from a website enquiry for the team to review.
3. Finance and contracts. For an international events group running 20+ exhibitions, we connected Pipedrive (CRM), SmartDocs (contracts) and Xero without replacing any of them. When a signature is detected the CRM updates, multi-event agreements split into separate deals, and the invoice is raised in Xero automatically. That means around 500 sales invoices a month no longer get typed by hand. This events group project is rule-based automation rather than AI, and it belongs here because most real processes need both kinds.
4. Claims handling. Our claims platform work replaced legacy case management, shared inboxes, spreadsheets and supplier portals with one platform covering notification, assessment, decision, settlement and repair. It includes AI matching and AI document review, with full audit trails.
5. Inbox triage (a common pattern). An AI step reads each message to a shared inbox, tags its topic and urgency, and routes it to the right person with a suggested reply.
6. Document handling (a common pattern). Supplier invoices, forms and contracts arrive in different layouts. An AI step extracts the fields, rules check the totals, and a person approves the exceptions.
7. Customer replies (a common pattern). An AI step drafts a first reply from your help pages and past answers, and a person reviews it before it is sent, at least at the start.

How much does AI automation cost?

It depends on the route, and the cheapest route is sometimes the right one. At the time of writing (October 2026), Zapier’s pricing page lists a Free plan with 100 tasks a month. Its Professional plan starts from US$19.99 a month when paid yearly, or US$29.99 a month when paid monthly (Zapier). For a simple, low-volume workflow, such as copying form entries into a sheet and drafting a reply, a tool like that is often the right answer. We would say so.

A custom build earns its cost when the work is higher in volume, touches several of your systems, or needs the AI step tuned to your own data. These are our published prices.

Route	Cost	Fits when
Off-the-shelf tool	Free, or from US$19.99 a month	Simple, low-volume jobs
AI readiness assessment	£3,000	You need to know where to start
First agent	From £5,000	One job, scoped
Agent rollout	From £12,000	Several workflows

The AI readiness assessment is a one-week programme at a fixed £3,000, with no obligation to build, and it includes an honest verdict where AI is not worth it. A first agent is one agent doing one named job inside your existing tools, with clear limits on what it decides alone and human sign-off where it matters. An agent rollout covers several workflows, with escalation to a person when confidence is low.

On a build, scope and price are fixed and agreed before work starts. You pay by milestone as each stage is delivered, and all code and IP are yours once the work is delivered and paid for. If you cancel within the first 21 days of a build, you get your money back in full.

After launch, our maintenance and support start from £400 a month. Running costs sit apart from the build: hosting, platform subscriptions and usage charges from the AI model provider, which vary with volume. Ask for an estimate of those before you sign, whoever you work with. If connecting systems and moving data matters more than reading, our business process automation service is the better starting point.

What are the risks of AI automation?

There are three to plan for.

Wrong answers. Set a confidence threshold, send low-confidence cases to a person, check the model’s output with ordinary rules, and compare error rates with the manual process before you switch over.

Untrusted input. An automation that reads inbound emails and documents is reading text written by strangers. The National Cyber Security Centre says prompt injection arises when developers put their own instructions and untrusted content in one prompt, then trust the model to keep them apart. It adds that “The best we can hope for is reducing the likelihood or impact of attacks” (NCSC). In practice, treat every message the system reads as untrusted. Give it the narrowest permissions that do the job, and keep a person in front of anything that cannot be undone.

Personal data. If the automation handles personal data, UK data protection law applies. The ICO’s guidance on AI and data protection says: “Due to changes made by the Data (Use and Access) Act, this guidance is under review and may be subject to change”. Its page dates the last update to 15 March 2023, so check it for the current position.

The ICO’s separate guidance defines automated decision-making as “where you use personal information to make a significant decision about someone using solely automated processing, including profiling” (ICO). That page is draft guidance, updated on 31 March 2026 for the Data (Use and Access) Act 2025.

Sorting, extracting and drafting for a person to approve is a different thing from deciding about an individual with nobody involved. If your automation could make a decision with legal or similarly significant effects on a person, take advice before you build. This is not legal advice.

Ask any supplier, us included, where your data is processed, who can access it, how long it is kept and whether the AI provider trains on it. You should also be able to leave: all code and IP are yours once the work is delivered and paid for.

How do you start with one process?

Start with a single process. AI automation for a small business works the same way as for a larger firm, with a smaller first job. There are five steps:

1. Pick a job with volume and untidy input. Look for something that happens often and arrives as text or documents, such as enquiries, invoices or onboarding forms. Our post on where automation pays off shows how to count the hours before you commit.
2. Write down the right answer. If two people would disagree about what “correct” looks like, settle that first.
3. Decide what the system may do alone. List what it may decide itself, what it drafts for approval and what it must never touch.
4. Run it beside the current process. Compare results for a few weeks before you switch.
5. Measure the result. Track hours saved, errors and turnaround time against the old way.

If you are not sure which process to pick, the AI readiness assessment maps your workflows over one week for £3,000. It tells you where AI will and will not pay off. If you already know the job, a first AI agent starts from £5,000, with scope and price agreed before work starts.