An agentic workflow is an AI system that breaks a goal into steps and works through them using a language model plus tools, memory and retrieval, instead of answering one prompt in isolation. Anthropic draws a useful line: workflows run the model along predefined code paths, while true agents let the model direct its own steps. Gartner expects task-specific AI agents in 40% of enterprise applications by the end of 2026, up from under 5% in 2025.
What does "agentic workflow" actually mean in 2026?
Strip away the marketing and it means one thing: an AI system that does a job in stages. A normal chatbot takes your prompt and returns an answer. An agentic workflow takes a goal, plans the steps, calls tools to gather or change things, checks its own output, and only then finishes.
The building block is what Anthropic calls the "augmented LLM": a language model given three extra powers. Retrieval, so it can look things up. Tools, so it can act (send an email, query a database, run code). Memory, so it can carry context across steps rather than starting fresh each time.
In Anthropic's December 2024 guide to building effective agents, agentic systems split into two camps. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths." Agents are "systems where LLMs dynamically direct their own processes and tool usage." Most of what UK businesses ship in 2026 is the workflow kind, and that is usually the right call.
How is an agentic workflow different from a normal AI workflow or a single agent?
People use the three terms loosely, so here is the clean split. A plain AI workflow might call a model once inside a fixed process. An agentic workflow chains several model-driven steps with tool use and self-checking. A true agent hands the model the keys: it decides the path itself, loops until the job is done, and stops when it judges the goal met.
The trade-off is predictability against flexibility. Workflows are consistent and cheap to reason about, which matters when an auditor or a regulator asks why the system did what it did. Agents cope with messy, open-ended tasks but cost more, run slower, and are harder to explain after the fact.
Dimension | Single LLM call | Agentic workflow | Autonomous agent |
|---|---|---|---|
Who decides the steps | You (one prompt) | Predefined code path | The model, dynamically |
Tool use | None or one | Several, in a set order | Chosen at runtime |
Predictability | High | High | Lower |
Cost and latency | Lowest | Moderate | Highest |
Good for | Q&A, drafting | Defined multi-step jobs | Open-ended, unpredictable jobs |
Auditability | Easy | Easy | Harder |
Anthropic's own advice is blunt: find the simplest solution that works, and only add complexity when the results justify it. Sometimes that means no agentic system at all, just one well-written prompt. The interesting bit is that the people building the most agents say the same thing.
What are the main agentic workflow patterns?
There are five patterns worth knowing. They are the vocabulary that engineering teams and vendor docs use, and they map neatly onto real UK business jobs. You can combine them, but most useful systems are one or two of these, not all five stacked up.
Prompt chaining
The task is split into ordered steps, and each model call works on the previous output. Draft a proposal, then critique it, then rewrite. Good when a job has clear stages and you want a quality gate between them.
Routing
An input is classified, then sent to the right specialist path. A customer email gets tagged as billing, technical, or complaint, and each route uses a tuned prompt. This keeps each step focused and cheaper than one giant do-everything prompt.
Parallelisation
Independent subtasks run at the same time, then the results are combined. Review a contract for legal, commercial and data-protection risk in parallel, then merge the flags. Faster, and each pass stays specialised rather than diluted.
Orchestrator-workers
A central model breaks a task down on the fly and delegates pieces to worker models. Useful when you cannot predict the subtasks in advance, such as researching a topic where the model decides which angles to chase.
Evaluator-optimiser
One model generates, a second scores against criteria, and the loop repeats until the output passes. This is how you get reliably good copy, code, or translations without a human checking every draft.
What does the adoption data actually show?
Interest is enormous and delivery is patchy. McKinsey's State of AI report from November 2025 found that 88% of organisations now use AI in at least one function, but agentic systems are earlier: 39% are experimenting with AI agents and only 23% have started scaling them anywhere in the business.
Gartner predicts that over 40% of agentic AI projects will be scrapped by the end of 2027, citing escalating costs, unclear business value and weak controls. Source: Gartner via Reuters, June 2025.
Here is the bit that really matters for a UK leader. The gap between 39% experimenting and 23% scaling is where most budgets go to die. The projects that survive tend to be narrow agentic workflows with a clear owner and a measurable outcome, not open-ended agents let loose on a vague brief. Start where the job is defined.
There is a sterling case to be made too. On ONS Annual Survey of Hours and Earnings 2024 median earnings, a UK knowledge worker costs an employer roughly £23.80 an hour once National Insurance and pension are added (Tom & Co analysis). An agentic workflow that reliably saves two hours a week per person clears more than £2,400 a year of that cost, before you count the errors it prevents.
Picture a mid-sized UK firm handling supplier invoices. A routing step reads each invoice, a chaining step matches it to a purchase order, and an evaluator step flags anything that does not reconcile. That is a workflow, not an agent. It runs the same way every time, a person approves the exceptions, and the finance team stops keying data by hand. Boring, defined, and it pays back inside a quarter.
Compare that with the projects that stall. They tend to start with "let's build an agent that runs the whole department" rather than one job with a number attached. Gartner's cancellation figure is not really about the technology failing. It is about scope. The narrower the brief, the higher the survival rate.
Where do agentic workflows fit in a UK business's stack?
The honest answer: behind a human, on jobs where the steps are known and the stakes are bounded. Document review, first-draft drafting, data enrichment, ticket triage, and research all suit a workflow. Anything that ends in a decision about a real person needs more care.
That care is a legal one. Under UK GDPR, a "solely automated" decision with a legal or similarly significant effect, such as declining a mortgage or rejecting a job application, gives the individual specific rights. The ICO's guidance on automated decision-making is clear: a human in the loop only counts if they can genuinely change the outcome, not just rubber-stamp it.
This got sharper in 2026. The Data (Use and Access) Act 2025 replaced the old Article 22 with new provisions on automated decisions, and the ICO's work on AI now includes updated draft guidance issued in March 2026. If your agentic workflow touches decisions about customers or staff, that human-review design is not optional, it is the compliance boundary.
What should a UK leader do this quarter?
You do not need a moonshot. You need one workflow that earns its keep and teaches you how these systems behave. Pick the job first, then the pattern.
Pick one repetitive, multi-step job with a clear owner and a number attached to it (hours saved, tickets cleared, drafts produced).
Start with a workflow, not an agent. Predefined steps are easier to test, cheaper to run, and far easier to explain to an auditor.
Match the pattern to the job: routing for triage, prompt chaining for staged drafting, evaluator-optimiser where quality has to be consistent.
Keep a human in the loop on anything that affects a real person's rights, and make sure that human can actually overrule the system.
Measure against the baseline for four weeks before you decide whether to scale, kill, or expand it.
The teams getting value in 2026 are not the ones with the cleverest agents. They are the ones who picked a boring, well-defined job and made a workflow do it reliably. That is where to start.
