Agentic workflows are processes in which at least one step goes to an AI agent that is given a goal rather than a script. The agent plans its steps, picks its data and tools, checks its results and changes course when the evidence surprises it. Rule-based automation runs steps fixed in advance, so the difference is who writes the route.
Put an AI agent on work a rule already does and you pay more for a less predictable version of the same result. Leave a rule on work that needs judgement and the process stops at the first exception, where it waits for a person with a full diary. Both mistakes are easy to make when nearly every product pitched to a travel programme claims to run agentic workflows.
The interest is real enough. In a GBTA poll of 591 travel professionals, run in October 2025 and published on 30 October, one third of buyers said their companies were already experimenting with autonomous AI. Half of buyers said they planned to use agentic AI for expense reconciliation.
Reconciliation holds both kinds of work at once. Matching a card transaction to an expense claim has been a rule's job for years; working out why a batch of transactions will not match is agent work, and that kind of work is rarer than the pitches suggest.
In this article
- What is an agentic workflow?
- Agents vs workflows: why does the word workflow mean two things?
- How is an agentic workflow different from rule-based automation?
- When should a process use an AI agent instead of automation?
- What does an agentic workflow look like in a corporate travel programme?
- What are the steps in an agentic workflow, and where does a person approve?
- What is agent washing, and how do you spot it?
- Do agentic workflows replace the automation you already run?
- Frequently asked questions
- What should a travel programme delegate first?
What is an agentic workflow?
An agentic workflow is a business process in which at least one step is done by an AI agent: software that takes a goal, decides its next action, uses tools and data to take it, checks the result and repeats until the goal is met. The agent's steps are chosen while the work runs rather than fixed before it starts.
What makes a workflow agentic?
A workflow is agentic when something inside it chooses what happens next, instead of following an order fixed in advance. Agentic AI workflows need four properties at once, and each can be checked on any product shown to you in a demonstration:
- A goal rather than a script: you say what you want, and leave the how to the agent.
- A plan written at run time: the agent chooses its steps after it has seen the goal and the data.
- Tools: it can query systems, read documents and run calculations as well as write text.
- A loop: it tests its own output and changes the plan when something does not fit.
Take any one away and you have something useful but different. Without the run-time plan it is automation, without tools it is a chat assistant, and without the loop it is a script that happens to call a language model, which is the version most often sold under the name.
Agents vs workflows: why does the word workflow mean two things?
Engineers and buyers use the word workflow differently. In engineering, a workflow is the fixed part of a system, steps written in code in advance, and an agent is the part that chooses its own steps. In business usage, an agentic workflow is any process with an agent inside it; both are correct, and mixing them causes confusion.
The UK Government Digital Service keeps the two senses side by side in its AI Insights guidance on agentic workflow, published on GOV.UK in April 2025. It describes legacy workflow systems as managing business processes through "fixed, sequential steps", and an agentic workflow as one in which AI agents "manage, coordinate, and execute tasks within business processes".
In the engineering vocabulary, agents versus workflows is a design choice, and a well-built product usually contains both: fixed code for the steps that never change, and an agent for the steps that do.
How is an agentic workflow different from rule-based automation?
Rule-based automation runs steps a person wrote in advance, and stops or hands off when an input does not fit. An agentic workflow starts from a goal, chooses its steps after seeing the data and pursues an exception instead of queuing it. Automation is cheaper and more predictable; agents suit work whose route changes from one run to the next.
The deeper difference is when the judgement gets spent: once, by whoever wrote the rules, or on every run, by the model. That is why an agent copes with cases nobody foresaw, and why it costs more to run.
A third option sits between the two and is the one most often mislabelled: the AI assistant, which answers one question at a time while a person drives every step.
The table below compares all three on the questions a programme lead asks before buying. The last column shows how Orchestra, PredictX's AI agent team for travel and expense, handles each row.
When should a process use an AI agent instead of automation?
Use an AI agent when the steps cannot be written down before the work starts, because they depend on what the data shows. Use automation when both the steps and the exceptions can be written in advance. An assistant is enough when the task is a one-off question that a person asks and answers in a single sitting.
Gartner's June 2025 release on agentic AI gave the short version. Anushree Verma, a senior director analyst there, said organisations can start with "AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval".
The split is right and hard to apply, because nearly every process contains some decisions, so the Route Test below turns it into four questions.
The Route Test
The Route Test is PredictX's four-question check for sorting a process by one property: whether its route can be written down before the work begins. Answer each question yes or no.
What does an agentic workflow look like in a corporate travel programme?
In a travel programme, an agentic workflow takes a goal such as explaining why booked hotel stays did not match the contracted rates. It decides which bookings and contract terms to examine, follows up each exception and returns a finished briefing. The rule that compares every stay with its contracted rate keeps running underneath it, unchanged.
Take hotel rate integrity, described here as an example rather than a customer result. The automated check has already compared every booked stay with its contracted rate and flagged the ones that differ, which is exactly the work a rule should do.
Explaining the flags is different. Given the goal "explain the flagged rate exceptions", an agent groups them by property and checks each against the contract's terms, looking for the cause: a room type the agreement does not cover, breakfast priced separately, or a date the hotel treats as outside the deal. Where a clause is ambiguous it stops and asks rather than guessing, and the briefing separates the exceptions worth raising with the hotel from the ones the contract allows.
Those explanations become the evidence for the next negotiation. Talking about negotiations with a travel management company, David Smith, a corporate travel manager, told Business Travel Executive in April 2026 that the hard part is "having the clarity to make the conversation productive".
The table below runs the Route Test over nine common travel processes, putting the agentic workflow examples beside the ones that should stay automated. Your answers will differ wherever your rulebook does.
What are the steps in an agentic workflow, and where does a person approve?
An agentic workflow runs in six steps: take the brief, plan, act, check, revise and deliver. The agent loops through acting, checking and revising until the evidence holds. In a well-designed system a person approves at four points: the plan before it runs, any finding that contradicts the brief, anything that commits money, and the deliverable before it is used.
- Brief: the goal, scope and deadline, plus answers to the agent's clarifying questions.
- Plan: the steps the agent intends to take and the data each one needs.
- Act: queries, documents and tools, often on several steps at once.
- Check: each result tested against the brief, the source data and the other results.
- Revise: when a check fails or a finding surprises, the plan changes and the loop repeats.
- Deliver: the write-up, with evidence attached so each figure can be traced.
How to place those four approval points, and how to stop an approval becoming a formality, is the subject of PredictX's guide to where the human sits in the loop.
What is agent washing, and how do you spot it?
Agent washing, as Gartner uses the term, is the rebranding of an existing product, such as an AI assistant, robotic process automation or a chatbot, as agentic when it lacks substantial agentic capability. You spot it by asking who writes the route. If a person configured every step in advance, the product is automation, however it is described.
In the same June 2025 release, Gartner estimated that only about 130 of the thousands of agentic AI vendors are real. Verma added that many use cases positioned as agentic do not need an agentic implementation at all.
Much of what is sold as AI workflow automation is rules with a language model on the front, and that matters less than it sounds: for routine work it is a sound purchase at a routine price. The harm is paying for judgement you are not getting.
Five questions that expose a rebadged workflow
- Run the same request on two different sets of data. Does the plan change, or only the numbers?
- Feed it a record that breaks the pattern. Does it dig into the record, or open a ticket?
- Ask to see the plan before anything runs. No plan to see means nothing is planning.
- Ask where one figure in the output came from. An agentic system can show the query behind it.
- Ask what the product will not do. A vendor who cannot name a limit has not found one.
The fourth question is the one you keep asking after you buy. In PredictX's own design a separate validation agent audits every deliverable, as set out in the Cogent AI Framework's validation design.
Do agentic workflows replace the automation you already run?
Agentic workflows do not replace existing automation: they sit on top of it and depend on it. The rules that route approvals and match transactions do that work more cheaply and predictably than an agent could. Agents take the work that stalls at exceptions, and the analysis nobody has had time to do.
Gartner's release makes a related point: bolting agents onto legacy systems can disrupt workflows and require costly changes, and rethinking a workflow from the ground up is often the better path. Rethinking keeps the rules. The work is deciding, step by step, which parts are fixed and which need judgement, which is the Route Test applied to a whole process.
Keep every rule that works. Redesign the part of the process where work waits for a person.
Frequently asked questions
What is the difference between AI agents and agentic workflows?
An AI agent is software that takes a goal, chooses its own next step and uses tools to act. An agentic workflow is a business process that contains one or more agents alongside fixed steps. The agent is the component, and the workflow is the process it works inside, with its inputs, rules and hand-offs.
What is an example of an agentic workflow?
A common example of an agentic workflow in a travel programme is following up hotel rate exceptions. A rule flags every booked stay that differs from its contracted rate. Given the goal of explaining those flags, an agent checks each one against the contract's terms, groups the causes and returns a briefing that a person acts on.
Is RPA the same as agentic AI?
Robotic process automation (RPA) is not the same as agentic AI. RPA repeats steps a person recorded, such as copying data between systems, and stops when the screen or the data changes shape. Agentic AI chooses its own steps towards a goal and handles inputs nobody anticipated, and Gartner lists RPA relabelled as agentic among its examples of agent washing.
How do you know whether a process needs an AI agent?
Two questions decide whether a process needs an AI agent: can its steps be written down before the work starts, and does a rulebook cover its exceptions? If both answers are yes, it needs automation, and an agent only adds cost. If its route depends on what the data shows and it draws on several sources, it needs an agent.
What are the risks of agentic workflows?
Agentic workflows cost more per run than rules, and the same question can take a different route next time. The larger risk is a plausible answer built on a wrong assumption. The controls are a plan a person reads before anything runs and a trail from every figure to its source, with each commitment kept by a named person.
Which travel processes should stay automated?
Any process whose steps and exceptions can both be written down in advance should stay automated. In most travel programmes that includes routing pre-trip approval requests, blocking fares above the policy cap, sending scheduled spend reports and the routine matching of card transactions to expense claims. An agent adds cost and variation to that work and nothing a rule lacks.
What should a travel programme delegate first?
Delegate first the recurring analysis whose steps cannot be written down in advance and that has waited longest. For many programmes that is renewal preparation, the follow-up on hotel rate exceptions or the explanation of a budget variance: work with a known goal and a route that depends on the data, which nobody has been free to do properly.
Orchestra is PredictX's AI agent team for travel and expense: an AI project manager plans the work, directs nine specialist agents and returns a finished, cited deliverable. Nothing runs until a person approves the plan. After that, the specialists it needs carry out the parts in parallel and a validation agent audits the result: autonomous in execution, human-gated in decision.
The table's rows for carrier renewal, hotel rate integrity, travel policy, budget variance and the quarterly business review match five of Orchestra's example projects, and the library is open to more. Orchestra launched in June 2026 and is deploying across enterprise travel programmes, and 4 of the 6 largest global T&E programmes run on PredictX.
Related Posts
AI agents for procurement: what they can run, and what the team keeps
