ON-DEMAND WEBINAR
How to find & evaluate agentic AI use cases worth building
If you're leading an automation program and fielding pressure to show agentic AI ROI, this webinar is for you. Get a practical, lightweight framework for spotting high-value AI use cases with real business impact and learn how to generate your best agentic AI and automation ideas.
When to Automate? And When to Automate with Agentic AI?
An automated solution is worth building as an agent when it clears three gates in order. First, it produces measurable value, hard or experiential. Second, it has the volume, frequency, and data readiness to justify automating at all. Third, it scores well against five agent-specific criteria: logic variability, range of acceptable outcomes, human oversight, input adaptability, and integration complexity.
Most processes that pass the first two gates should still be built with rules-based automation. That is the point of scoring them.
Agentic AI has a hype problem, and everyone evaluating it right now knows it. The useful question is no longer whether agents work. It is which of your processes actually deserve one, and which will run cheaper, faster, and safer with the automation you already have.
Ashling's Head of Advisory, Filip Swat, and Head of Delivery, Xavier Hansen, walked through the light weight framework Ashling uses with clients on a recent webinar, including live scoring of real processes with the audience. Here is the framework and how to apply it.
What Counts as an Agent, and What Doesn't
The terms get used interchangeably, so start with definitions.
TYPE |
WHAT IT DOES |
BEST FIT |
| Rules-based automation (RPA, scripting) | Follows predetermined paths. Fixed inputs produce fixed outputs. | High-volume, well-defined processes with predictable inputs |
| AI-based automation | Takes unstructured input and returns an answer, a summary, or a classification. It does not act. | Document understanding, classification, drafting, chat |
| AI agents | Combines the reasoning of a model with the ability to take action through tools and systems. | Variable processes with many decision paths and unstructured inputs |
| Orchestration | Platforms like UiPath Maestro coordinate across all three, plus humans and more. | Every agentic solution. It is implicit, not optional. |
Within AI itself, four categories do different jobs. Specialized AI handles document understanding and extraction. Predictive AI runs regression-based models on historical data, and has done so for decades. Conversational AI answers questions. Agentic AI adds action to reasoning.
Rules-based automation is not going away. Ashling has yet to see a client environment that does not still depend on it. The strongest programs mix all four and orchestrate across them, because agents rarely run a process end to end. Control and cost both argue for injecting agents at specific steps instead.
Where Agents Fit in a Process
Agents work through four capabilities, and most solutions use two or three rather than all four.
- Perceive. Ingest unstructured or semi-structured input and give it a usable structure. Free-text emails, handwritten forms, and mixed-format documents all qualify.
- Reason. Weigh multiple inputs against context and choose a path.
- Act. Trigger the next step through tools and system access, with governance around what the agent is allowed to touch.
- Learn. Catch its own mistakes and feed them back so prompts and context improve on the next pass.
From the Field
A useful pattern from the field: agents earn their keep at the front and the back of a process. At the front, consolidating omnichannel information into something a consistent process can handle. At the back, summarizing and generating output for people.
Order processing shows all four. An agent picks orders out of an inbox in whatever format they arrive, structures them, and hands off. A second agent applies customer history and SKU logic, searching the product catalog by number or close name match when a product is missing. A third routes the order to fulfillment. Each step is a separate agent with narrow tool access, which makes them easier to control, govern, and check.
Email triage follows the same shape. An agent reads intent and urgency from a shared inbox, prioritizes and routes, then spots patterns worth flagging. Fifty password resets in an hour is a signal that something upstream is broken, and the agent can warn the service desk before the calls land.
Start With the Work You Already Automated Around
The fastest place to find agent candidates is not a blank page. It is the backlog of things that hit the cutting room floor.
For the last decade, automation teams have built around sticky steps. A human drops data into a spreadsheet. A file gets handed off through SharePoint. A process with a hundred business decisions gets scoped down because coding all that logic in traditional automation kills the value case.
Those are now buildable. An insurance client, for example, ran claims auditing where every extraction from an explanation of benefits landed in a spreadsheet for a human to eyeball and re-upload. An agent can handle the extraction and the system entry, with the human reviewing instead of transcribing.
This is also the safest place to start. You already know those business units. You already understand the requirements. Introducing new technology into a process you have mapped beats introducing it into one you have not.
Software testing is the other example worth revisiting. Automated testing has been a solved problem for decades. An agent solves it better, because it recognizes that a submit button is still a submit button when it moves from the top right to the bottom right of a screen. Something working today does not mean it cannot work considerably better.
The Three Gates
Score every candidate in this order. Each gate can stop the process from moving forward, and the most valuable output of the framework is a defensible no.
Is There Value?
Value comes in two forms, and programs that only count one of them undersell themselves.
Hard value is what a CFO can put in a budget line: cost takeout, cost avoidance, revenue impact through faster turnaround or speed to market, total cost of ownership reduction, and risk budget reduction.
Soft value covers operational and experiential outcomes: regulatory compliance, audit readiness, error reduction from removing swivel chair work, customer satisfaction, and turnaround times.
Both count. Consider an agent that monitors patient vitals and flags the pattern that suggests sepsis, combining elevated heart rate with high blood pressure and rising temperature to prioritize a nurse review. Measure that in nursing hours saved and you have missed the point entirely. The value is lives. If your business case ignores experiential value or makes no effort to quantify it, you are telling a weaker story than the work deserves.
Should You Automate It at All?
Two criteria, both familiar to anyone who has run an intelligent automation program.
Process volume and frequency. How often does it happen, multiplied by how long it takes. Something that runs once a quarter rarely justifies the build. There is a real trade-off for low-volume, high-handle-time work: loan origination that takes 90 minutes per file, or insurance investigations that stretch across days. Those usually involve document gathering, checks, outreach, and multiple system touches, which is exactly where agents help. Low volume alone is not a disqualifier.
Data availability and process readiness. An agent is only as good as the data it can reach. Partial or incomplete data can produce a worse result than no automation at all. If the inputs live in fifty databases owned by different business units, the first project is re-engineering the process, not automating it. The old adage holds. Do not automate a broken process.
A low score here does not mean a bad candidate. It means it is not a good candidate right now.
Should an Agent Do It?
Five weighted criteria, each scored 1 to 100 in multiples of 10, then multiplied by its weight and summed.
CRITERION |
QUESTION IT ANSWERS |
SCORE HIGH WHEN |
| Logic (25%) | How many variants can the process take? | Many branches, heavy conditional complexity |
| Outcomes (20%) | How many acceptable answers exist? | Nuanced responses are valid, not just yes or no |
| Humans in the loop | How much human supervision does it need? | Human review is required and frequent |
| Adaptability (highest weighted) | How variable are the inputs? | Unstructured or shifting inputs, changing rules |
| Integration complexity | How many systems must it reach? | Many sources, where more context means better answers |
Two notes on how to read this. Human involvement raises the score rather than lowering it, because agents interpret what people give them without the rigid input structure that rules-based automation demands. And integration complexity works differently for agents than for RPA, where every connection is more work. For an agent, more integrations means more context, and more context means better answers.
Scoring Two Real Processes
Example 1: B2B Order Entry Scored 21.75 out of 100
A parts manufacturer receives orders from industrial clients through pre-populated forms in a client portal and via EDI. Every order is checked for product existence, stock, customer match, and pricing, then passed to the ERP or handed to customer care.
It passed Gate 2 comfortably. High volume, accessible structured data. Then the agent criteria: logic scored 25 because the rules are well defined and the flow is essentially singular. Outcomes scored 20 because an order is approved or escalated. Humans in the loop scored 40, limited to exceptions like a customer that does not exist. Adaptability scored 10, since EDI arrives pre-populated. Integration complexity scored 20 to 30 across two systems.
Total: 21.75. Strong automation candidate. Weak agent candidate. Build it with rules-based automation and IDP, and spend the agent budget somewhere it changes the outcome.
Example 2: Foreclosure Minimum Bid Calculation Scored 76.5 out of 100
A mortgage servicer calculates the minimum bid for foreclosure sales and communicates it to the lawyers handling the sale. The calculation depends on the number of liens, the property value, and the location, which triggers state and in some cases county regulations. That is 50-plus variations of one calculation.
Only a couple of people run it, so headcount value is small. The risk value is not. A decimal in the wrong place or the wrong state rule can turn a $100,000 exposure into a $1 million one. The stress drives high turnover, which is its own cost.
Logic scored 100, because the core math is simple and the matrix of state and county variants is enormous. Outcomes scored 10 to 20, since exactly one answer is correct. Humans in the loop scored 100, driven by risk. Adaptability scored 100, because states and counties change their rules more than once a year and publish them on public websites. Integration complexity scored 80 to 90, with structured data from a core mortgage system and unstructured rules from county sites.
Total: 76.5. This process was attempted with traditional automation years ago. It is a much better fit for an agent, and the reasoning step is exactly why.
How to Turn Scores Into a Pipeline
A single score in isolation tells you very little. There is no threshold where 30 means build an agent. Value comes from relativity. Score four or five processes and the pipeline sorts itself, and once you have built one of them, that result anchors everything scored after it.
Where to start depends on where you are:
- Little or no automation in place. Build capability first. Take small steps and show the business what is possible before scoping anything end to end.
- Established automation program, new to agentic. Go back to the cutting room floor. Enhance what exists and plug the gaps you built around.
- Agents already in production. Move to reimagination. Redesign processes around what agents make possible, rather than fitting agents into processes designed for people.
Two things hold across all three. Define your guardrails and governance before you scale, including where you will and will not let agents act. And scale the way you find opportunities, through operational discovery that surfaces large swaths of processes at once and runs them through the framework, so the pipeline builds faster than you can build the agents.
The framework earns its value in what it rules out. As Filip put it, quoting advice he got years ago, the most important part of a strategy is what you say no to. Every no in this framework is specific: no because the value is not there, no because it should not be automated yet, or no because an agent is the wrong tool for work that rules handle perfectly well.
.png?width=200&height=60&name=Ashling%20Logo%20(2).png "Ashling Logo (2)")
Frequently Asked Questions
A process where an AI agent both reasons over variable inputs and takes action in downstream systems. The distinguishing feature against conversational AI is action. A chatbot recommends the next step. An agent triggers it.
Score it on logic variability, range of acceptable outcomes, human oversight, input adaptability, and integration complexity. Processes with well-defined logic, structured inputs, and binary outcomes belong with rules-based automation, even when they qualify strongly for automation overall.
There is no universal threshold. Scores are comparative. Run four or five candidates through the same criteria and use the ranking to prioritize, then let real build results anchor future scoring.
No. The underlying models are trained by their providers. What agents can do is identify their own errors and feed that back to the people maintaining them, so prompts and context improve over time. Learning happens in the loop, not in the model.
Yes. Agents rarely run a process end to end, for both control and cost reasons. Most solutions place agents at specific steps alongside rules-based automation and human review, which means something has to coordinate across all of it.
Map every process and sub-process in the domain, identify the intersection points and the teams that own them, then apply the framework to each one separately. An exception path handled by a customer care team is its own process, and it scores on its own merits.