A request reaches an approver’s queue and waits. The approver is in meetings, traveling, or handling something urgent. The person who raised the request follows up, then waits again. The decision itself is often routine, so the delay comes from the step, not from the judgment. This is one of the most common reasons workflows slow down, and it’s a good place to start using AI agents in workflows. Teams that already use process optimization methods will recognize the problem: the bottleneck is the handoff, not the work.
Adoption is growing quickly, though definitions vary. KPMG’s AI Quarterly Pulse Survey for Q1 2026 found that 54% of large US organizations are actively deploying AI agents, according to a summary of the survey published by Second Talent. Gartner’s 2026 Hype Cycle for Agentic AI reports a very different figure: only 17% of organizations have deployed AI agents so far, though more than 60% expect to within two years, as summarized by AYAutomate. The gap comes from how each survey defines “using agents.” When you read these numbers, check the definition before you quote them.
The market is also moving. On October 21, 2026, Oracle announced Oracle AI Agent Studio for Industry Applications, a no-code and low-code environment for creating, testing, orchestrating, deploying, and managing AI agents and workflows for industry-specific business processes, as reported by Unite.AI. Agents are configured through natural-language instructions and visual workflow editing. Large platforms are moving from fixed workflows toward agents that act within a process.
AI agents in Quixy follow the same direction. An agent sits inside the workflow and takes the decision a person would otherwise make at that step.
How an AI agents in workflows replaces an approver
Consider an approval process where X raises a request and Y approves it. With an AI agent assigned to that step, the agent takes Y’s place. It decides based on the data in X’s request, so the request moves forward without waiting for Y to act.
Each decision comes with a confidence score. How that score is produced depends on how the agent is configured, so the setup you choose shapes how the agent weighs the information in a request.
Because the agent runs as part of the workflow, it performs its tasks as the process runs. That removes the delays that build up when a request waits for a person at an approval step. For more on how approval steps are designed, see our complete guide to approval process and workflow and our guide to enterprise approval workflow design.
Setup is configuration, not training
You don’t train an agent. You set it up by describing what it should do.
When you select Configure New Agent, two fields appear. The Decision Prompt defines how the agent evaluates input and makes its decision. The Response Format defines the JSON structure the agent returns. When you click Save, the agent is created and attached to the workflow.
The Decision Prompt does most of the work, so it deserves the most attention. Describe what the agent should look at, what each possible outcome means, and what the response should contain. A vague prompt produces vague decisions, and a specific one makes the agent’s behavior easier to review.
Agents you create are stored in the AI Agents Directory. When you configure an agent step in another workflow, you can select an agent that already exists instead of creating a new one. Because the agent is managed in one place, its decision logic stays consistent across every workflow that uses it. Reusing agents also fits with the consistency goals in workflow compliance management.
Configuring the process matters as much as the agent
An agent is only one part of the change. Deloitte’s 2026 research, as summarized by Second Talent, found that only 30% of organizations are redesigning key processes around AI, and 37% report using AI at a surface level with little or no change to underlying processes. Placing an agent inside a process that was built for a human approver may improve speed a little. Getting the full benefit means checking the steps around the agent too: what data the request captures, what happens after the decision, and who owns the outcome.
This is a good moment to review the process itself. Our guide to no-code automation tools covers the wider set of options when you’re planning changes across several steps.
What agents do and don’t do
Agents focus on reasoning, decisions, and actions within the workflow. Quixy’s documentation states that they don’t handle data capture, UI interactions, or manual configuration steps. Those still belong to the app’s forms and interface, which you build separately.
Routing, approvals, and automations continue to work as you configured them. The agent works on top of that process logic and uses application data to make its decisions while the workflow runs. It helps route records, trigger actions, and influence the path a request takes, based on context rather than fixed conditions.
Governance comes first
Many organizations are slowing down AI rollouts for governance reasons, not technical ones. AvePoint’s State of AI 2026 report found that 86.9% of companies have delayed AI deployments because data security and governance weren’t ready, not because of budget or buy-in, as reported in its Artificial Intelligence Report 2026. AvePoint sells governance software, so treat the figure as directional, but the pattern is worth planning for.
Before an agent goes live, decide three things. What data can it read? Which decisions can it make on its own, and which must go to a person? Who owns the outcome? Our guide to low-code governance and security covers the wider controls that apply to any custom app.
Where agents fit best
Agents suit steps where the decision depends on reading the details of a request. A purchase request might need a judgment about whether it fits an approved category. An intake form might need to be sorted by urgency and topic. A case might need to be routed based on the information it contains.
Start with an approval step that causes real delays and where the criteria are clear enough to write down. Steps with unclear criteria are harder to describe in a Decision Prompt, and their decisions are harder to check.
Getting started
- Pick one approval step. Choose a step where requests routinely wait and where the decision follows clear criteria.
- Write the Decision Prompt. Describe what the agent should look at, what each outcome means, and what the response should include, in plain language.
- Define the Response Format. Set the JSON structure so the workflow can act on the result.
- Save and attach the agent. Click Save to create the agent and attach it to the workflow step.
- Test on past requests. Run the agent on requests your team has already decided, and compare its decisions with the outcomes.
- Save it to the directory. Once the agent works well for one step, reuse it in other workflows.
Industry commentary on low-code AI workflows makes a similar point. Teams should test how a workflow handles duplicate events, unavailable dependencies, invalid model output, and recovery before rollout. Our workflow compliance management guide can help you build those checks into your rollout plan.
What to keep in mind
AI Agents are an add-on feature and require a separate subscription, so confirm the cost before planning a rollout.
The agent is only as good as the data it receives. If a decision depends on a field, make sure the form captures that field. Review the agent’s decisions regularly, especially early on, and adjust the Decision Prompt where its results differ from your team’s decisions. Keep a person accountable for the outcome of each workflow, even when an agent makes the decision.
Conclusion
AI agents let workflows make decisions at the step where people used to wait. You describe the decision in writing, the agent evaluates each request as it arrives, and the process keeps moving. For approvals that cause delays, the result can be faster decisions and less time spent chasing people for a sign-off. The teams that get the most from agents will be the ones that also review the process around them.
Frequently Asked Questions(FAQs)
Does an AI agent replace the approver?
Yes, at the step where you assign the agent. It takes the decision the approver would otherwise make, based on the data in the request. Routing and other steps in the workflow keep working as you configured them.
Do I need to train an AI model?
No. You set up an agent by writing a Decision Prompt, which explains how it should evaluate a request, and a Response Format, which defines the structure of its output. The Decision Prompt matters most, so describe the decision clearly.
Does every decision come with a confidence score?
Yes. Each decision includes a confidence score, and the score depends on how the agent is configured. Writing a clear Decision Prompt is the main way to shape how the agent reaches its decisions.
Can agents create documents?
Yes. Agents generate the documents and other material that later steps of the workflow need, wherever the workflow requires them. This means the next step can start with what it needs, without someone preparing it by hand.
Can I reuse an agent in several workflows?
Yes. Agents you create are stored in the AI Agents Directory, and you can select an existing agent when you configure an agent step in another workflow. Using one agent in several places helps keep decision logic consistent.
Recent Posts
- Download the App
