Of the five root causes RAND lists for failed AI projects, one is misunderstanding the problem. That changes who should build your automations, and which process to start with.
AI Automation for Business Processes: Who Should Build It, and What to Automate First
By James Cotton · Last updated · 10 min read
Part of our topic guide on AI Skills for Business.
By James Cotton, Founder, iO-Sphere
What UK businesses use AI for, and whose work it changes
The Office for National Statistics (ONS) does not yet count automation directly: its survey question on purpose offers no automation option, and it says it is exploring new measures of how AI is integrated into business processes. ONS figures for September 2026 show 38% of businesses with 10 or more employees using AI, up from about 35% in June 2026.
In ONS's dataset, the figures on purpose, affected roles and AI skills are shares of businesses with 10 or more employees that use AI (or are not sure whether they do). In September 2026, about 6 in 10 of these gave "improve business operations" as a purpose, and 38% named administrative and clerical roles as the roles AI had affected most, up from about 29% in December 2025.
In the same month, the commonest approach these businesses named to integrating AI skills was training or retraining existing staff, at about 4 in 10, against about 1 in 10 automating or replacing roles with AI. Of businesses with 10 or more employees currently using AI, 6% reported reduced headcount as a result; in ONS's words, "Fewer than 1% reported that their headcount had increased."
Research for the Department for Science, Innovation and Technology (DSIT) does ask about automation, on its own question and base. In its AI Adoption Research, with fieldwork from February to May 2025, 47% of 919 businesses that used or planned to use AI said they used or planned to use it for automation, rising to 58% of mid-sized and 75% of large businesses.
When a firm asks for AI skills, one thing it often wants is staff who can automate the routine parts of their own jobs, and in 2024 most employers interested in AI training told a survey for DSIT they wanted training on using AI to reduce repetitive tasks; our guide to a whole-workforce AI programme has the detail.
Why AI projects stall
RAND's August 2024 report on the root causes of AI project failure cites other authors' estimates that more than 80% of AI projects fail, twice the rate of IT projects that do not involve AI. RAND's own evidence, in this US study, is interviews with 65 experienced data scientists and engineers.
RAND lists five root causes without ranking them. The first is that "industry stakeholders often misunderstand, or miscommunicate, what problem needs to be solved using AI." The others are lacking the data to train the model, a focus on "the latest and greatest technology" over real problems, inadequate infrastructure, and problems too difficult for AI to solve. No platform supplies the first; it sits with whoever knows the work.
MIT NANDA's July 2025 preliminary report, The GenAI Divide, drawn from research between January and June 2025, makes two separate claims. One is that despite $30 to 40 billion of enterprise investment in generative AI, "95% of organizations are getting zero return."
The other is that "Just 5% of integrated AI pilots are extracting millions in value", and the report names learning as the core barrier to scaling. Its evidence is a review of over 300 publicly disclosed AI initiatives, interviews with representatives of 52 organisations and survey responses from 153 senior leaders, and it calls its figures "directionally accurate based on individual interviews".
In DSIT's 2025 research, what it called agentic AI, defined as systems that "autonomously perform tasks, make decisions, and take actions", with robotic process automation among its examples, was the least adopted AI technology, used by 7% of the 700 businesses using AI. On a base of 194, 32% reported significant barriers implementing it, against 18% for natural language processing and text generation (base 774).
Most AI users in that research had "moved past piloting and into live use", and 84% of the 700 reported significant or some input or checking of AI outputs (67% significant; 2% none). Yet "only a few businesses said that they had a specific procedure for checking AI outputs". A check without a set procedure relies on the checker knowing what a wrong output looks like.
The National Audit Office, Parliament's auditor, advises in its May 2026 good practice guide to start with "small, clearly defined, low-risk use cases" and to "treat pilots as learning exercises rather than routes to rapid automation". In CIPD's 2026 survey with YouGov, people professionals and business leaders reported that around 45% of organisations still used AI on an ad hoc or experimental basis.
What these sources share is a dependence on knowledge of the work: what problem needs solving, what a wrong output looks like, what a pilot should teach. Our view follows from that: the people who own a process are best placed to build, or at least specify, the automations that run on it, and AI lands better when they redesign the process than when outsiders do.
The platforms, and what each one charges for
Automation platforms meter different things, and the meter decides what a long workflow costs. n8n prices by workflow execution "regardless of complexity", with unlimited users and workflows on every plan. On Zapier, "every step in a Zap and every external connector call uses tasks, though rates may vary." On Make, each module action counts as one credit. Power Automate's paid plans are priced per user or per bot.
A ten-step workflow run once is therefore one execution on n8n, while on Make it uses a credit for each module action and on Zapier tasks for each step. The table gives each vendor's prices as its pricing page showed them on 8 October 2026, the day we read every price, listing and register entry on this page.
| Platform | What it charges by | Prices as displayed on 8 October 2026 |
|---|---|---|
| n8n | Workflow executions; unlimited users | Starter £20 a month billed annually (£24 monthly), 2,500 executions; Pro £50 (£60 monthly), 10,000; Business, self-hosted only, £569 a month billed annually (UK browser; currency varies by location) |
| Zapier | Tasks, used by each step | Free, 100 tasks a month; Professional from $19.99 and Team from $69 a month billed yearly (£15.30 and £52.81 with pounds selected), excluding taxes |
| Make | Credits, one per module action | Free, up to 1,000 credits a month; Core $9, Pro $16, Teams $29 a month billed annually, each for 10,000 credits (dollars only) |
| Microsoft Power Automate | Users or bots | Premium £11.50 per user a month; Process £115.30 and Hosted Process £165.30 per bot a month; paid yearly, excluding VAT |
Three of the pages carry a caveat of their own: n8n's shows different currencies by location, Zapier says its pound prices may vary with the exchange rate, and Microsoft says its prices are for marketing purposes and may differ from actual list prices.
n8n describes its licence as "fair-code", a term it coined for software that is source-available but restricts commercial use of its source code. It says data on its hosted plans is stored within the EU, on servers in Frankfurt.
Microsoft's licensing guide says basic Power Automate is included at no cost for work or school accounts in a Microsoft Entra tenant, limited to cloud flows without sharing, attended desktop flows run locally, and standard connectors only.
When an automation decides about people or deals with customers
Customer-facing automation brings in consumer law. The Competition and Markets Authority (CMA), in its guidance on using AI agents published 9 March 2026, says: "you are responsible for what an AI agent does in the same way you are responsible for what an employee does. This is true even if someone else designed or provides the AI agent on your behalf".
The uses it names include handling customer queries and processing refunds, and it asks for "a human in the loop, actively checking that the AI agent is making correct decisions and generating expected results." For teams that answer customers all day, see AI skills for customer service teams.
The CMA warns of fines "up to 10% of your worldwide turnover" and possible compensation to consumers, its short form of the ceiling in section 182 of the Digital Markets, Competition and Consumers Act 2024.
UK GDPR Articles 22A to 22D, in force since 5 February 2026, require safeguards when an automation takes a decision with a legal or similarly significant effect on a person and "no meaningful human involvement", including a way to obtain human intervention and contest it. The regulator, now the Information Commission (the ICO), opened a call for evidence on agentic AI on 8 October 2026, closing 20 November 2026.
Our guide to AI in recruitment and HR decisions sets out the automated-decision rules in full. The controls companies put around an agent are in how companies govern agentic AI, the rules for staff using AI tools in our generative AI use policy guide, and the wider UK register in AI regulation in the UK in 2026.
Training the people who run the process to automate it
The work of building these automations has its own apprenticeship standard. Artificial intelligence (AI) and automation practitioner (ST1512) is a Level 4 standard, approved for delivery, for "identifying, designing, and delivering practical improvements in how work gets done across business functions, using AI and automation". Its practitioners are "typically embedded in operational teams", working on "manual tasks, duplicated data entry, unintegrated tools, and inefficient workflows".
The standard expects them to take "ownership of specific projects or tasks that deliver tangible operational value" without being expected to lead teams. Its duties include to "identify opportunities for automation to drive operational improvement and cost savings" and to "monitor and refine automations incorporating feedback from end- users to improve". Find apprenticeship training had 143 providers listed on 8 October 2026.
The people who decide what gets automated have training of their own. Skills England's March 2026 release described apprenticeship units as shorter programmes of 1 to 16 weeks for adults in work. Three Level 5 AI leadership units, approved for delivery on 21 April 2026, cover strategy and opportunity, adoption, procurement and governance and delivery and organisational transformation.
Each requires the learner to be "working in a leadership position within an organisation with autonomy to deliver technological change and inform investment decisions". The wider market for AI transformation training is compared in our guide to AI transformation courses.
The person who will redesign a process end to end draws on a different standard, Business analyst (ST0117, Level 4), whose duties include to "model business processes using relevant techniques", "perform business process analysis and improvement" and "redesign business process models in order to reflect changes in working practice or deliver improvements". It suits that job because those duties are the process knowledge an automation depends on.
We train teams to automate their own processes as a custom team programme, priced per engagement. For a firm that wants the people who run its processes building sound automations now, that is the route we recommend. For the person whose job includes redesigning a process end to end, whatever the title, there is our AI Transformation apprenticeship (Level 4).
Choosing the first process
Our test for a first automation follows from the evidence above. A process makes a good first candidate when each of these holds:
- The work is routine: the same steps, in the same order, nearly every time.
- The team runs it every week, so the automation is exercised often and its errors show up early.
- The people who run it can name the exceptions, and a person keeps handling them.
- No step decides something about a person of the kind Article 22A covers.
- You can put a price on a wrong output, and it is a cost you could carry while the team learns.
A process that fails one of these can come later, once the team has seen how an automation behaves on its own work.
Frequently asked questions
Is n8n free to use?
Not if n8n hosts it for you: the cheapest cloud plan a UK browser displayed on 8 October 2026 was Starter, at £20 a month billed annually, and the page shows different currencies by location. n8n also publishes a self-hosted Community Edition on GitHub, under a licence that allows use "only for your own internal business purposes or for non-commercial or personal use".
Do we need a developer to automate our processes?
Not for a first automation on a routine process. The apprenticeship standard for this work expects practitioners to "configure and adapt low-or no-code tools to solve problems and drive efficiencies", and its record lists no entry requirements. It also expects them to know "mitigation strategies for post-deployment issues such as overreliance and automation bias". Our guide for non-developers covers what to learn first.
Does n8n's licence let a contractor charge us for setting up our workflows?
Yes: n8n's licence FAQ lists "Charge your customers for workflow creation, setup, and maintenance services" among the things allowed. It lists "Host n8n as a service and allow your clients to build workflows" among the things not allowed, and it defines "internal business purpose" as "the use of n8n within your organization or product where people outside of your business can receive or see what your workflows produce."
Upskilling a whole team?
Tailored data and AI training for organisations, from data literacy to technical upskilling, built around your team's real work.