AI Automation5 min read
Which Business Processes Should You Automate With AI? A Scoring Framework
Picking the wrong first process is the most expensive mistake in AI automation. A five-factor scoring method, worked examples and the red flags that say 'not yet'.
Contents
Every AI automation program starts with a list of candidate processes and a hunch about which is best. The hunch is usually drawn toward the impressive one, the process where a demo would make executives lean forward. That is also, more often than not, the wrong choice. This framework replaces the hunch with a score you can defend.
The five factors
| Factor | Question | 5 looks like | 1 looks like |
|---|---|---|---|
| Volume and frequency | How often does it happen, and how much time does it consume? | Hundreds of items a week, hours of staff time | A few times a quarter |
| Clarity of a correct output | Can an expert say whether the result is right? | Clear rules, examples and a definition of done | Depends on taste or negotiation |
| Cost of an error | What happens when the AI is wrong? | Caught in review or easily reversed, small impact | Irreversible, regulated or reputationally damaging |
| Data readiness | Can the system read what it needs, and is it reliable? | Structured access, current, permissioned | Scattered, stale, on paper or in someone's head |
| Measurability | Can you capture a baseline and see the change? | Existing metrics for time, cost, errors | No records, benefits are subjective |
Add the five scores for a total out of 25. Anything under 15 is not ready. Above 20 is a strong candidate. Between, look at what would raise the weak factor: often it is data readiness, which is cheaper to fix than it looks.
Then estimate value
A high score tells you a process is suitable, not that it is worth it. Estimate value with three numbers: monthly volume, minutes saved per item (net of review time) and loaded hourly cost. Multiply, then subtract usage and tooling costs and amortize build cost over 12 months. Be conservative; a process that only pays under optimistic assumptions is not a first project.
Scoring three candidates
| Process | Volume | Clarity | Error cost | Data | Measurable | Total |
|---|---|---|---|---|---|---|
| Support ticket triage and routing | 5 | 4 | 4 | 4 | 5 | 22 |
| Contract clause review for legal | 2 | 3 | 1 | 3 | 3 | 12 |
| Weekly sales report summaries | 3 | 4 | 5 | 4 | 3 | 19 |
Ticket triage wins: high volume, clear categories, mistakes are caught quickly. Contract review is important but rare, subjective and expensive to get wrong, so not a first project. Report summaries score decently but the value is modest. This is how a scoring table can turn an argument about enthusiasm into a comparison of facts.
Good first candidates by department
- Customer support: classification, routing, reply drafting, knowledge suggestions. See the support playbook.
- Sales: inbound lead enrichment, scoring and routing, call notes to CRM. See AI sales automation.
- Marketing: content repurposing, tagging, reporting summaries. See AI marketing automation.
- Finance: invoice and receipt extraction, expense categorization, reconciliation suggestions.
- HR and IT: request classification, policy Q and A, onboarding checklists.
- Engineering: test generation, code review assistance. See AI coding agents.
Red flags that mean not yet
- Nobody can define what a correct output is, or experts disagree.
- The process changes every month, so any workflow is obsolete on arrival.
- The relevant data is not accessible programmatically and cannot be made so cheaply.
- An error would be irreversible, regulated or affect individuals' rights. See the EU AI Act overview.
- The process is rare enough that savings cannot cover build cost.
- No one on the business side will own it and give time for review.
Rules of thumb that hold up
- 01Automate the boring before the impressive.
- 02Prefer assistive automation, where AI prepares and a person decides, for the first release.
- 03Choose a process whose result you can check within a day.
- 04Pick one, finish it, measure it, then re-score the rest with what you learned.
Once you have a winner, the next step is the 90-day roadmap. For background on why this discipline matters, see why agentic AI projects fail and the AI automation guide.
Sources and further reading
Frequently asked questions
How do I decide what to automate with AI?
List candidate processes and score each on volume, clarity of correct output, cost of errors, data readiness and how easily you can measure the result. Choose the one with high volume, clear correctness, low error cost and available data. Then estimate value in hours and money before committing.
What are good first AI automation projects?
Common first wins include support ticket triage and drafting, invoice and document data extraction, lead enrichment and routing, meeting summaries and follow-ups, internal knowledge search and IT request classification. They are frequent, text-based and easy to check.
What should I not automate with AI?
Avoid, at first, tasks that are rare, high-stakes and irreversible, such as legal commitments, medical or financial decisions affecting individuals, and processes with no documented right answer. Also avoid anything where your data is missing, scattered or unreliable.
How do I estimate ROI before building?
Multiply monthly volume by minutes saved per item and the loaded hourly cost, subtract expected review time and tooling and usage costs, then compare with build and maintenance cost over 12 months. Use conservative assumptions. Our ROI guide has a worksheet.
Should the first project be customer-facing?
Usually internal or assistive first. Automations that prepare work for staff let you learn about quality and failure modes before customers see the output. Move to customer-facing steps once you have data on accuracy and a solid escalation path.
Work with Navelo Software
Published by Navelo Software
An independent software and AI studio building AI systems, web platforms, mobile apps and SaaS for clients worldwide.
