Workato
Azure
MuleSoft
Salesforce
AWSReady To Modernize Your Architecture?
Discuss your current infrastructure and scalability goals with our integration experts.
As AI grows in adoption, enterprises are rightfully looking for ways to leverage the new tools that are emerging and constantly evolving. Recently, the term “AI” most commonly refers to agentic AI. While AI agents are the right choice to fulfill certain business needs, one of the challenges is that agents can often be unpredictable in their decision making and processes.
In this article, I will discuss how we can leverage modern AI in business workflows without compromising control, using an emerging type of AI models, which I will refer to as decision models. Then, I’ll walk through a lead qualification recipe I built in Workato to show what that looks like in practice.
01 - What are Decision Models?
In Thinking, Fast and Slow, psychologist Daniel Kahneman describes two modes of thought about how people make judgements and decisions:
System 1 is fast, automatic, and intuitive: recognizing a face, reading the mood in a room, or sensing that an email just feels off.
System 2 is slow, deliberate, and effortful: planning a trip, working through a math problem, weighing a hard decision.
Applying that concept to technology, this split is where we can draw a line about the different types of AI models and which problems they are most suitable to solve:
| System 1: Decision Model | System 2: Agentic AI with an LLM | |
|---|---|---|
| What it does | Answers a specific question about the data in front of it | Plans, reasons, and decides which steps to take |
| Speed | Milliseconds | Seconds to minutes |
| Output | Structured values: probabilities, scores, categories | Actions, plans, free-form text |
| Who controls the flow | The workflow | The AI agent |
| Process Predictability | Same path every time | Path can vary run to run |
| Best for | High-volume triage, classification, scoring | Open-ended problems with no fixed path |
The most significant difference is that rather than outputting free-form content like an LLM, decision models output a pre-determined shape. In other words, the output schema is deterministic, which is what allows us to easily integrate it into business workflows.
02 - Example: Qualifying inbound leads with a decision model
Suppose you have a “Contact sales” form on your website, which sends you potential leads from prospective customers. The data fed into that form could be spam, vendors trying to sell to you, or other submissions you don’t want to manually sift through. This is a problem that is suitable for a decision model. Nobody needs to plan anything or reason through multiple steps. All we need is a judgment of the data that is immediately available.
For this example, I built a custom Workato connector for TypeSafe’s Jev model and fed in the data from the form into the recipe.
Step 1 - The web form
Here is the form submission we’ll follow through the example:
In a typical workflow, these submissions might have some basic filtering and routing and would still end up needing to be manually reviewed and followed-up on.
In our example recipe, we will receive the form content through an API endpoint and create a Salesforce lead, enriched with some additional insights from the AI model.
Step 2 - The recipe
2.1 Efficiency & Speed
First, notice how we fed all of the form data into Jev and asked it to make a determination on 6 different questions at the same time:
| Noul Questions (Yes / No) | Scoring Questions |
|---|---|
| Is this spam? | Score the buying intent |
| Is this person selling us something rather than buying? | Score the fit of the company profile (details of the company profile were provided in the step) |
| Did this person ask for prices? | Urgency |
This is an especially notable benefit for Workato, where task consumption is always a consideration. Additionally, because the response is so fast (~317ms), we are able to integrate this into a real-time API.
One detail that is not shown here is the capability to process batch data. The API can receive the state (the data about which the questions are being asked) as a list in a single API call, and their results will be weighed against each other.
2.2 Deterministic Output Schema
Because we are defining the shape of the input at design-time, the connector is able to dynamically infer the output schema and we can design the workflow around that. In our example, we filtered out some of the form submissions based on the probability that the submission is spam.
Step 3 - The result in Salesforce
The data comes across into custom Salesforce fields, and can be used as inputs to automation Salesforce already provides. A few ideas for that:
| Example Use Case | Description |
|---|---|
| Queue Routing | Lead assignment rules can send high-fit, high-urgency leads straight to an enterprise sales queue, and lower-fit leads to an inside sales or nurture queue. |
| Real-time alerts | A record-triggered Flow can notify the assigned rep, or post to a team channel, the moment an urgent lead is created. |
| Pricing follow-up | If the lead asks for pricing, a Flow can create a follow-up task or send the pricing sheet automatically. |
| Nurture campaigns | Leads with low buying intent can be added to a marketing campaign instead of taking up a rep’s time. |
| Reporting | Dashboards can track lead quality by source over time, so marketing can see which channels bring in real buyers. |
None of this requires new AI logic. Once the judgment is stored as data, the rest is standard Salesforce configuration that admins already know how to build.
03 - Key Takeaways
Making complex capabilities seem simple is a powerful asset. Here’s what makes it work:
- The recipe maintains control of the workflow: Instead of AI being the orchestrator, it is a step in the workflow. This allows us to maintain visibility and control of business processes through a deterministic flow, while still leveraging the intelligence of AI.
- Speed and Efficiency: Because of how fast the responses are, the AI step fits inside a real-time API request.
- Structured output with type safety: Every answer comes back as a typed value: a probability between 0 and 1, or a string value from a fixed scale. Those can be fed predictably into other software, with confidence on type safety.
04 - So, Does this Replace Agents?
Not at all. However, the options for which tool is best suited for a particular problem should be considered before reaching for AI agents as the solution.
Each task requires an honest evaluation of which kind of thinking it calls for. If the answer is a quick judgment about the data in front of you, an agent is more machinery than the problem needs.
05 - Where else decision models fit
Lead qualification is just one example. Once you start looking for System 1 problems, they show up in many workflows. Here are a few more examples; each one is a quick question about a single piece of data asked at high volume, where the answer decides which branch of an otherwise ordinary workflow to take.
| Workflow | Questions to ask |
|---|---|
| Customer feedback and surveys | Is this response positive or negative? Does it mention a specific product or feature? Is the customer at risk of leaving? |
| Support chat routing | Which team should handle this? Is it a billing, technical, or account question? Does it need a human right away? |
| Returns and refunds | What is the reason for the return? |
| Job applications | Does this applicant meet the minimum requirements? How closely does their experience match the role? |
| IT and system alerts | Is this alert noise, or does it look like a real incident? How severe is it? |
| User-generated content | Is this review or comment appropriate to publish? Is it spam or off-topic? |
06 - Wrapping up
AI agents are a powerful tool, but they aren’t the only way to bring AI into a business process. For many everyday problems, all we need is a quick, reliable judgment about the data in front of us.
Decision models give us that judgment in a structured shape that a controlled workflow can act on directly. The recipe stays in control of the process, and the AI step adds intelligence while maintaining predictability. Before reaching for an agent, consider which kind of thinking the problem calls for. If it’s a System 1 problem, a decision model is likely the simpler, faster, and less expensive choice.
The Jev connector used in this example was built by API People. We design and build custom connectors on both Workato and MuleSoft, so teams can bring new APIs, AI models, and core business systems into their workflows without waiting on an off-the-shelf option.
If you are interested in this connector, want to build a new one for a system you use, or would like to hear more about our other Workato/MuleSoft connectors, reach out to us.
FAQs
1. How can you use AI in business workflows without losing control?
Use AI in business workflows as a step, not as the orchestrator. In the lead qualification recipe, a decision model answers questions about the incoming data and the Workato recipe decides what happens next. The recipe maintains control through a deterministic flow, keeping business processes visible and predictable while still leveraging the intelligence of AI.
2. What is a decision model for AI in business workflows?
A decision model is a System 1 type of AI model: fast, automatic, and focused on answering a specific question about the data in front of it. Instead of free-form content like an LLM, it outputs a pre-determined shape, such as probabilities, scores, or categories, in milliseconds. That deterministic output schema is what makes it easy to integrate AI in business workflows.
3. What is the difference between decision models and AI agents in business workflows?
For AI in business workflows, a decision model answers a specific question in milliseconds, returns structured values, and follows the same path every time, with the workflow in control. An AI agent built on an LLM plans, reasons, and decides which steps to take over seconds to minutes, returns actions, plans, or free-form text, and its path can vary from run to run. Decision models fit high-volume triage, classification, and scoring; agents fit open-ended problems with no fixed path.
4. How does a decision model qualify inbound leads in Workato?
The recipe receives the “Contact sales” form through an API endpoint and sends the data to TypeSafe’s Jev model through a custom Workato connector built by API People. Jev answers six questions in a single step, including whether the submission is spam, whether the person is selling rather than buying, and how strong the buying intent, company fit, and urgency are. The step ran in about 317ms, so this use of AI in business workflows fits inside a real-time API. Likely spam is filtered out, and the rest becomes a Salesforce lead with the answers stored in custom fields.
5. Why does a deterministic output schema matter for AI in business workflows?
Because the shape of the input is defined at design time, the connector can infer the output schema, and the workflow can be designed around it, which keeps AI in business workflows predictable. Every answer comes back as a typed value: a probability between 0 and 1, or a string value from a fixed scale. Those values can be fed predictably into other software, such as Salesforce lead assignment rules, record-triggered Flows, and dashboards, with confidence on type safety.
6. Do decision models replace AI agents in business workflows?
Not at all. AI agents remain the right choice for certain business needs, but before adding AI in business workflows, each task requires an honest evaluation of which kind of thinking it calls for. If the answer is a quick judgment about the data in front of you, an agent is more machinery than the problem needs, and a decision model is likely the simpler, faster, and less expensive choice.
7. Where else do decision models fit for AI in business workflows?
AI in business workflows with decision models fits anywhere a quick question about a single piece of data is asked at high volume and the answer decides which branch of the workflow to take. Examples include customer feedback and surveys, support chat routing, returns and refunds, job applications, IT and system alerts, and user-generated content.
Free PDF
Read it offline — get the full PDF
Fill in two fields and the download starts right here.
- Every figure and diagram at full resolution
- All sources, linked and cited
No spam. We respect your inbox and your data.
Contact Us
Talk to an integration expert
About your environment — no pitch deck required.
