Receive
Take the incoming text as data. Do not execute instructions inside it.
← Back to Work
Demonstration Study • AI & Workflow Automation
Automate the right work. Keep people in control.
A workflow opportunity assessment and a working inbox-screening model. Explore what is worth changing, what the evidence supports and where a person should make the final decision.
Context
A busy inbox is not a reason to automate every decision.
Start with the work, measure the errors, then design the handoff.
The business question is where AI can remove repetitive handling without losing a legitimate enquiry or creating another review burden. This demonstration separates an editable business case from a measured text-classification experiment.
The working example uses historical, publicly available SMS messages to screen incoming text for spam. It is a small machine-learning model, not a generative chatbot. It proposes a queue; a person can override that proposal and record the decision.
Demonstration note: This is not a client deployment. Test results describe this English SMS corpus only. Planning inputs are illustrative assumptions, and the source cannot establish productivity gains in a modern business inbox.
01 · Opportunity assessment
Make the business case visible before building.
Adjust the illustrative scenario below. Account for assisted handling, adoption and running cost, then check whether the workflow is ready for a bounded test.
Workflow and assumptions
Planning scenario · editable assumptions
02 · Working automation
A proposal, a review, a visible decision.
The trained model runs in your browser. Choose a held-out public example or enter short English text. Your typed message is processed locally and is not sent to a server.
Screen
Apply the trained model and a vocabulary-support check.
Review
Keep uncertain inputs in manual triage. Let a person override any route.
Record
Record the proposal and final decision. Do not automatically delete a message.
Try the screening workflow
Examples retain the source wording and published labels.
This historical SMS model has not been validated on business email, WhatsApp enquiries, Nigerian languages or current spam campaigns. Do not use this demo to make a production decision.
Loading the model…
Human review desk
Every route is a proposal. Nothing is connected to a live inbox.
Manual triage
Recognised words with the largest contribution to this score:
These are model contributions, not proof of intent or a causal explanation.
The demonstration log stays in memory until this page closes. Its CSV excludes message content. Confirming a decision does not send, delete or update anything outside this page.
03 · Measured evidence
Accuracy is a start. The errors decide the workflow.
The model was trained on 3,560 messages and checked on a separate validation set before reporting the untouched 763-message test set. Normalized exact duplicates were removed before splitting. The test contains 675 legitimate messages and 88 spam messages.
What the default classifier gets wrong
| Published label | Predicted legitimate | Predicted spam |
|---|---|---|
| Legitimate | 664 | 11 |
| Spam | 9 | 79 |
Eleven legitimate messages were flagged as spam; nine spam messages were missed. That makes silent deletion an unsuitable default, even with high headline accuracy.
Test accuracy is within this corpus. Near-duplicate templates may remain; this random split does not measure future performance or transfer to another channel.
Change the review policy
Scores at or above this level go to spam review. Scores below 20% go to the normal queue. The remaining messages go to manual triage.
- Normal queue 649
- Manual triage 40
- Spam review 74
- Legitimate messages flagged for spam review 2
- Spam entering the normal queue 6
This threshold experiment uses held-out model scores. The interactive text tool adds a vocabulary-support guard. Scores are not independently calibrated certainty. Changing the threshold here explores the test set; it must not be presented as independent tuning evidence.
04 · Operating design
Move from a useful example to a controlled implementation.
Map the real workflow
Identify the owner, inputs, handoffs and exception types. Measure actual handling time, rework and missed messages before setting a productivity target.
Validate on the intended channel
Build a permissioned, appropriately handled dataset of real business messages. Compare a simple rules baseline with the trained model and any proposed language-model alternative. Measure errors by language, message type and business consequence.
Run in shadow mode
Generate proposals without changing the live workflow. Review disagreements and record overrides. Set acceptable error rates and escalation rules before connecting any action.
Connect a bounded action
Start with reversible tagging or queue assignment. Keep approvals for outward messages and consequential actions. Monitor drift, review volume, incident rates and actual time released; retain a rollback path.
Method and reproducibility
A complete model, with its limits left visible.
The downloaded file contains 5,574 lines. After normalized-text deduplication, 5,086 messages remain. A seeded, stratified 70/15/15 split separates training, validation and test data. TF–IDF text features feed a logistic-regression classifier; only the training split fits the vocabulary and model. The same coefficients run in the browser and are checked against Python predictions.
Reproduce the methodDownload validation resultsDownload test predictionsRead the implementation controlsDownload workflow assessment template
Source: Almeida, T. & Hidalgo, J. (2011), SMS Spam Collection, UCI Machine Learning Repository, DOI 10.24432/C5CC84. UCI licenses the dataset under CC BY 4.0. ONE Light Analytics performed deduplication, modelling, evaluation and workflow design. The reported test metrics are model results, not client outcomes.
Have a repetitive information workflow?
Bring us the work. We will help identify what is worth changing.
ONE Light Analytics can assess the opportunity, test the model and design the handoffs before connecting automation to everyday operations.