"Agentic" is one of the most used words in AI, and one of the most stretched. In its June 2025 forecast, Gartner predicted that 33% of enterprise software applications will include agentic AI by 2028, yet also that over 40% of agentic AI projects will be cancelled by the end of 2027. In the UK, the government's Artificial intelligence adoption research 2026 found that just 2% of businesses with five or more employees were using agentic AI.
This guide sets out what the term actually means, how agentic AI differs from generative AI, the difference between agents and agentic workflows, and how a UK business can use them without losing control of its data or its decisions.
- Agentic AI pursues a goal over several steps and uses tools, rather than answering a single prompt.
- Generative AI creates content; agentic AI uses that ability inside a loop of planning, acting and checking.
- Workflows follow steps you define; agents decide their own steps. Most production systems are mostly workflow.
- Adoption is early: 2% of UK businesses with five or more staff used agentic AI in 2026, against 32% using any AI.
- Guardrails and a human checkpoint for high-risk actions are not optional extras, and UK GDPR still applies.
What does agentic AI mean?
Agentic AI describes systems that take a goal, break it into steps, act through tools and adjust based on what happens. A chatbot answers a question. An agentic system can read an incoming enquiry, look up the customer in your CRM, check availability in your calendar, draft a reply, and ask a person to approve it before sending.
OpenAI's practical guide to building agents puts it simply: "Agents are systems that independently accomplish tasks on your behalf." The UK Information Commissioner's Office, in its Tech Futures report on agentic AI, says agentic AI "combines the capabilities of generative AI with additional tools and new ways of interacting with the world." Both point at the same two ingredients: a language model that can reason about what to do next, and tools that let it act.
- Receive a goal or trigger, such as a new enquiry
- Plan the next step
- Act through a tool: search, read, write or update
- Check the result against the goal
- Repeat, finish, or hand over to a person
What is the difference between agentic AI and generative AI?
Generative AI produces content in response to a prompt; agentic AI uses a generative model to decide and take actions towards a goal. ChatGPT drafting an email is generative AI. A system that reads the enquiry, checks your stock system, drafts the email and queues it for approval is agentic. The two are not rivals: almost every agentic system has a generative model at its centre.
| Generative AI (chatbot) | Agentic workflow | AI agent | |
|---|---|---|---|
| Starts from | A prompt from a person | A trigger, such as a form or email | A goal |
| Who decides the steps | The person, one prompt at a time | You, in code, in advance | The model, as it goes |
| Uses tools and systems | Rarely, or only to look things up | Yes, a fixed set at fixed points | Yes, chosen by the model |
| Predictability | High per answer | High | Lower, varies run to run |
| Typical use | Drafting, summarising, research | Enquiry triage, quotes, invoice processing | Coding, research, messy inboxes |
What is the difference between an AI agent and an agentic workflow?
The difference is who decides the path. Anthropic's engineering guide, Building effective agents (December 2024), calls both "agentic systems" but separates them like this:
- Workflows are "systems where LLMs and tools are orchestrated through predefined code paths." You decide the steps; the model does the reading, writing and judging inside each step.
- Agents are "systems where LLMs dynamically direct their own processes and tool usage." You give it a goal and a set of tools; it decides what to do and when it is finished.
The same guide warns that agentic systems "often trade latency and cost for better task performance" and recommends "finding the simplest solution possible, and only increasing complexity when needed." In practice, the more predictable the task, the more it should be a workflow. Agents earn their place where the steps genuinely cannot be known in advance, such as investigating a bug, researching a supplier or working through a messy inbox. That is also how we approach AI process automation: workflow first, autonomy only where it pays.
What are the common agentic workflow patterns?
Anthropic's guide names five workflow patterns. Between them they describe most business automations that use AI:
| Pattern | How it works | Business example |
|---|---|---|
| Prompt chaining | A task is broken into a fixed sequence of steps, each using the previous output | Turn call notes into a summary, then a quote, then a follow-up email |
| Routing | The input is classified and sent to a specialised follow-up | Sort incoming emails into sales, support, accounts and complaints, each with its own handling |
| Parallelisation | Several model calls work on a task at once and their outputs are combined | Check a contract against several policies at the same time, then merge the findings |
| Orchestrator-workers | A central model breaks a task down and delegates parts to worker models | Plan a website content update, then hand each page to a separate writer and checker |
| Evaluator-optimiser | One model produces a draft while another evaluates it and asks for improvements | Draft a product description, then check it against brand rules and facts until it passes |
None of these needs a fully autonomous agent. That is good news: workflows are easier to test, cheaper to run and easier to explain to a customer or a regulator. Each one also depends on connecting the model to real systems, which is where the Model Context Protocol (MCP) comes in, and why MCP server development is often the first build in an agentic project.
What are multi-agent systems?
A multi-agent system splits work between several agents, each with its own instructions, tools and context, coordinated by code or by a lead agent. The orchestrator-workers pattern above is the most common form. Coding tools use the idea too: Claude Code can hand a task to subagents that each work in their own context window, which keeps the main conversation focused. Our guide to Claude Code for business covers how that works in practice, and Codex vs Claude Code compares the two leading coding agents.
Multi-agent designs help when a task is too big for one context or benefits from specialists. They also multiply cost, latency and the number of places something can go wrong. Use them because the task needs them, not because the architecture diagram looks impressive.
How many UK businesses use agentic AI?
Very few so far. The government's AI adoption research 2026, based on 3,503 interviews with UK businesses between January and April 2026, found that 32% used at least one AI technology, double the 16% of 2025, but only 2% used agentic AI. Text generation tools such as ChatGPT dominate.
Base: 3,503 interviews with UK businesses with 5 or more employees, surveyed January to April 2026
| AI technology | Currently using |
|---|---|
| Any AI technology | |
| Text generation and NLP | |
| Computer vision and images | |
| Machine learning | |
| Hardware related to AI | |
| Agentic AI |
Source: GOV.UK, Artificial intelligence adoption research 2026
The Office for National Statistics measures adoption differently and gets a similar shape. Its analysis of AI in UK businesses from 2023 to 2026 found use among firms with 10 or more employees rose from around 12% to around 35% since late 2023, but among those adopters "only 10% report using AI extensively." The same DSIT report notes that agentic AI was "less widely trusted and more likely to be subject to scrutiny" than simpler tools.
- 2% of UK businesses with five or more employees were using agentic AI, up from 1% in 2025
- 7% of businesses already using AI had adopted agentic AI
- 34% of agentic AI users developed it in-house, up from 12% in 2025
- 81% of businesses reported some human input or oversight of AI outputs
Source: GOV.UK, Artificial intelligence adoption research 2026
Why do so many agentic AI projects fail?
Mostly because they start from the technology rather than a costed problem. Gartner's June 2025 press release blamed "escalating costs, unclear business value or inadequate risk controls" for its forecast that over 40% of agentic AI projects will be cancelled by the end of 2027. Gartner's Anushree Verma described most current projects as "early stage experiments or proof of concepts" driven by hype.
- 40%+ of agentic AI projects will be cancelled by the end of 2027
- 33% of enterprise software applications will include agentic AI by 2028
- 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028
- 130 or so of the thousands of agentic AI vendors are real, Gartner estimates
Gartner also warned of "agent washing": rebranding existing products as agents without substantial agentic capabilities. Its January 2025 poll of 3,412 webinar attendees shows how cautious most organisations were:
Base: 3,412 Gartner webinar attendees
| Investment so far | Share |
|---|---|
| Significant investment | |
| Conservative investment | |
| Wait and see or unsure | |
| No investment |
Gartner's advice is plain: agentic AI should "only be pursued where it delivers clear value or ROI." For a UK business that means naming the hours or errors a project should remove before anyone writes a prompt.
What makes agentic AI risky?
An agent that can act can also act wrongly. The main risks are practical rather than science fiction:
- Wrong actions at speed: a misread instruction becomes a sent email, a changed record or a refund.
- Prompt injection: text inside an email, web page or document tries to give the agent new instructions.
- Over-broad access: an agent given your whole inbox or database when it needed one folder or one table.
- Runaway cost: loops that call a model hundreds of times.
- Unclear accountability: nobody can say why a decision was made.
Prompt injection deserves special attention. The UK's National Cyber Security Centre, in its December 2025 blog Prompt injection is not SQL injection, argues that "prompt injection attacks may never be totally mitigated" because a language model has no inherent distinction between data and instruction. Its advice is to design systems that reduce the likelihood and impact of an attack, not to assume it can be stopped.
OpenAI's guide makes the same point from the build side: "Think of guardrails as a layered defense mechanism," and treat human intervention as "a critical safeguard," especially when an agent exceeds failure thresholds or attempts high-risk actions. The design rule we use follows from that: the agent prepares, a person approves anything that is expensive or hard to undo.
What does UK data protection say about agentic AI?
UK GDPR applies in full, and the ICO has named the issues it expects. Its Tech Futures report lists concerns including controller and processor responsibilities across the agentic AI supply chain, "a larger amount of automated decision-making," purposes set too broadly for general-purpose agents, agents processing more personal information than a task needs, and new cyber security threats.
The rules on automated decisions have also changed. The ICO's overview of the Data (Use and Access) Act 2025 says all of its data protection provisions are now in force, and that it "opens up the full range of reasons, or 'lawful bases'" for significant automated decisions, as long as safeguards are in place. The ICO updated its draft automated decision-making guidance on 31 March 2026 and is consulting on it, and its guidance on AI and data protection is under review.
| Question | What the ICO says now |
|---|---|
| Which lawful basis can a significant automated decision use? | The full range, potentially including legitimate interests, with appropriate safeguards |
| Does that apply to special category data? | No, special category data remains more protected |
| Which safeguards must people get? | Information about the decision, a way to make representations, human intervention and a way to contest it |
| Is the guidance final? | No, the updated ADM guidance was in draft for consultation at the time of checking (November 2026) |
Sources: ICO, The Data (Use and Access) Act 2025; ICO, Automated decision-making guidance (draft)
For a small business, that translates into a short checklist: give each agent the narrowest access that does the job, write down what personal data it touches and why, keep a log of what it did, and make sure a person can review and reverse decisions that affect customers. Firms in regulated sectors, such as solicitors and accountancy and finance, should take advice before automating any decision about a client.
How should a business start with agentic AI?
Start small and specific. A good first project has a clear trigger, a few well-understood steps and a person at the end:
- Pick one repetitive task with a measurable cost, such as triaging enquiries or preparing quotes.
- Map the steps a person follows today, and the systems they touch.
- Build it as a workflow first, with the model doing the reading and drafting.
- Connect only the tools it needs, ideally through MCP with scoped permissions.
- Keep a human approval step for anything sent to customers or involving money.
- Measure time saved and error rates for a month before adding autonomy.
For concrete ideas, see our AI agent examples for UK small businesses. If you are unsure which processes are worth automating, AI integration consulting is a sensible first step; if your interest is software delivery, see our Claude Code agency service.
Circlemount is based at St James House in Telford, Shropshire, so we can meet businesses in Telford and across the West Midlands, including Birmingham, in person to map a process on a whiteboard. Most agentic work is delivered remotely, so firms in Manchester, London and elsewhere in the UK can work with us the same way. Our agentic AI services cover the full journey, and AI agent development takes a mapped workflow to production.
Frequently asked questions
What is agentic AI in simple terms?
Agentic AI is AI that can work towards a goal over several steps instead of answering one question. It plans, uses tools such as email, calendars or databases, checks the results and continues. A chatbot replies to you; an agentic system can complete a task for you, ideally with your approval at key points.
What is the difference between generative AI and agentic AI?
Generative AI produces content, such as text, images or code, in response to a prompt. Agentic AI uses that ability inside a loop: it decides what to do next, calls tools to act, and checks what happened. Most agentic systems are built on generative models, with tools, memory and rules added.
What is an agentic workflow?
An agentic workflow is a process where AI models and tools follow steps you have defined in code, such as classify an email, look up the customer, then draft a reply. The model handles the judgement inside each step. Workflows are more predictable than fully autonomous agents, so they suit most business tasks.
How many UK businesses use agentic AI?
Very few. The government's AI adoption research 2026 found that 2% of UK businesses with five or more employees were using agentic AI, up from 1% in 2025, while 32% used at least one AI technology. Most firms use text generation tools such as ChatGPT rather than agents that take actions.
Are multi-agent systems better than a single agent?
Not automatically. Several agents can help when a task is large or needs different specialists, such as research, writing and checking. They also add cost, delay and more points of failure. Start with one well-designed workflow or agent and split the work only when a single context clearly struggles.
Is agentic AI safe for a small business to use?
It can be, with sensible limits. Give each agent the narrowest access it needs, keep a person approving anything sent to customers or involving money, log what it does and watch for instructions hidden in emails or documents. The NCSC warns prompt injection may never be fully solved, so limit the damage.
Do I need to tell customers I use AI agents?
Transparency is a core UK GDPR principle, so your privacy information should explain how you use personal data, including through AI. If an agent makes significant automated decisions about people, they must be told and be able to make representations, ask for human intervention and contest the decision under the updated rules.
Sources
Checked on 16 November 2026. Product features and figures change quickly, so check the current source before you rely on a detail.
- Anthropic: Building effective agents (December 2024)
- OpenAI: A practical guide to building agents (PDF)
- Gartner: Over 40% of agentic AI projects will be canceled by end of 2027 (June 2025)
- GOV.UK: Artificial intelligence adoption research 2026 (PDF)
- ONS: Artificial intelligence in UK businesses, 2023 to 2026
- NCSC: Prompt injection is not SQL injection (it may be worse)
- ICO: Tech Futures, agentic AI
- ICO: The Data (Use and Access) Act 2025, what it means for organisations
- ICO: Automated decision-making, including profiling (draft guidance)
- ICO: Guidance on AI and data protection
Thinking about an agent for your business?
Circlemount designs and builds AI agents and agentic workflows for UK businesses, with sensible guardrails and human approval where it matters. See our AI agent development service and the wider agentic AI services, or Join the Circle for a free business analysis and proposal.





