Automation-versus-Machine-Learning Diagnostic
Classify workflows before choosing automation or predictive AI
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 97%
The Automation-versus-Machine-Learning Diagnostic begins by mapping a workflow’s inputs, decisions, actions, and outputs. If the correct response can be expressed as a stable, predefined sequence or rule, use task automation. If the workflow must ingest historical or live data to predict an outcome, rank alternatives, personalize a choice, or adapt behavior, evaluate machine learning. Many effective systems combine both: a model selects the next-best action, while deterministic automation sends the message, changes the page, or routes the record. The diagnostic also checks whether enough representative data exists and whether uncertain decisions require human review. By naming the mechanism accurately, teams avoid buying unnecessarily complex tools, overstating AI capabilities, or forcing decision-tree chatbots into situations that require genuine prediction and conversational adaptation.
Origin
Kieran Flanagan distinguished task automation from machine learning while discussing an AI-driven demand-generation funnel. Extracted from Marketing Against The Grain.
Core principles
- 01Task automation follows predefined steps triggered by known inputs.
- 02Machine learning ingests data to make predictions or choices.
- 03The two mechanisms solve different classes of problems.
- 04Clear classification prevents ordinary automation from being mislabeled as AI.
- 05Hybrid workflows can combine deterministic execution with learned decisions.
How to run it
- 1
Map the workflow
Document the inputs, decisions, actions, and desired outputs from beginning to end.
Pro tip Use one concrete example record or customer journey to expose hidden decisions.
Watch out Do not classify the technology before understanding the actual process.
- 2
Find deterministic steps
Identify actions whose correct sequence can be specified in advance from known inputs.
Pro tip Express each candidate as an if-then rule or fixed sequence.
Watch out A large rule set is not necessarily machine learning.
- 3
Find predictive choices
Identify decisions that require learning from data, estimating probabilities, ranking options, or adapting to an individual.
Pro tip Specify the prediction target and the decision it will influence.
Watch out Machine learning adds little when the correct action is already known with certainty.
- 4
Test data readiness
Check whether relevant, representative, and sufficiently clean data exists for the predictive component.
Pro tip Confirm that historical success measures align with the outcome currently desired.
Watch out Biased or obsolete data can optimize the wrong behavior.
- 5
Assign the mechanism
Use automation for predefined execution and machine learning for evidence-based prediction or choice.
Pro tip Combine the two when a learned decision should trigger a reliable automated action.
Watch out Do not permit model uncertainty to propagate invisibly through downstream automation.
- 6
Set oversight
Define confidence thresholds, monitoring, and human review for decisions with material customer or business impact.
Pro tip Escalate ambiguous cases rather than forcing a prediction.
Watch out Automated scale can amplify small model errors quickly.
In the wild
A demand-generation team uses machine learning to predict which message and send time best fit each prospect based on firmographics, demographics, and engagement. Once the model makes that choice, deterministic automation schedules and sends the approved email, records the action, and applies frequency limits.
→ The workflow personalizes decisions while keeping execution controlled and auditable.
A company maps its support bot and finds that navigation and account routing follow stable rules, while understanding free-form questions requires learned language capabilities. It keeps deterministic routing but uses machine learning only for interpretation and response selection, with human escalation for low-confidence cases.
→ Customers receive more useful conversations without discarding reliable workflow rules.
Common mistakes
Calling every workflow AI
Pre-programmed steps are task automation even when they are numerous or sophisticated.
Using machine learning without a prediction
If no uncertain choice, ranking, or forecast exists, a deterministic solution may be clearer and cheaper.
Automating uncertain outputs blindly
A predictive error can trigger many downstream actions unless confidence thresholds and review paths exist.
Is it for you?
Best for
It is best for marketers and operators evaluating funnel, personalization, forecasting, and workflow technology.
Not ideal for
It is not ideal for deciding whether a use case is ethical, strategically differentiated, or legally permissible.
From the transcript
“We should delineate between task automation and machine learning.”
“Like task automation is somewhat of a pre-programmed set of steps that you would take based upon inputs, whereas machine learning is like ingest data…”
From the episode
The Impact of AI in Marketing (Friend or Foe?)