Automation should start where work is repeated, rule-based enough to structure and painful enough to matter. In many companies, the best first opportunity is not advanced AI. It is the copy-paste work that connects disconnected systems.
People export files, re-enter data, reconcile spreadsheets, copy customer details, update status fields, paste messages into templates and assemble reports by hand. These tasks are often visible symptoms of missing integration.
Find the manual data movement
Start by asking teams where information moves by hand.
Examples:
- CRM data copied into project tools;
- orders copied into spreadsheets;
- product information updated in several places;
- support emails classified manually;
- invoices checked against records;
- reports assembled from exports;
- partner data uploaded through portals;
- approval status tracked in email.
These tasks are good candidates because they are concrete and measurable.
Separate automation from AI
Not every automation needs AI. Some workflows are better handled with deterministic rules, API integrations, validation and user interfaces. AI becomes useful when the input is unstructured, ambiguous or language-heavy: emails, documents, descriptions, notes, contracts, support messages or reports.
A practical automation roadmap may include:
- direct system integration;
- validation rules;
- workflow status tracking;
- document extraction;
- AI-assisted classification;
- human review;
- reporting automation.
The point is to choose the simplest reliable method for each step.
Map the workflow
Before building, map:
- trigger;
- input data;
- current manual steps;
- systems involved;
- business rules;
- exceptions;
- reviewer;
- output;
- destination system;
- success measure.
This turns “we need automation” into a buildable workflow.
Integration first
NNIT’s systems integration material is relevant because it frames integration as essential to modernisation and improvement. If systems cannot exchange data, automation has to work around the gap. That usually means scripts, spreadsheets or manual uploads.
A better foundation is:
Source system → API / data layer → workflow logic → review → destination systemOnce this exists, AI can be added where it creates real value.
Product data example
IMPACT Commerce’s AI-powered PIM article shows how AI can support product data workflows such as enrichment, localisation and data-quality management. The broader lesson is that AI works best when there is a clear data domain, structured process and review model.
The same applies to many non-commerce workflows.
Start small, but not carelessly
The first automation should be narrow enough to ship and valuable enough to matter. Good candidates are:
- frequent;
- rule-based or reviewable;
- connected to known systems;
- measurable;
- low to moderate risk;
- painful for users today.
Avoid starting with a workflow that is rare, politically complex, poorly understood or impossible to measure.
Memory(One) perspective
Memory(One)’s Integration & Platforms and AI, Data & Automation services meet in this topic. The work is to replace manual data movement with connected workflows, then apply AI where it improves speed, consistency or decision support. Automation should reduce operational friction without creating a new fragile black box.
Sources and inspiration
- NNIT — Importance of systems integration: https://www.nnit.com/insights/articles/importance-systems-integration
- IMPACT Commerce — AI-powered PIM benefits: https://impactcommerce.com/insights/12-advanced-benefits-of-ai-powered-pim/
- Novicell — What is composable architecture?: https://www.novicell.com/uk/latest-thinking/what-is-composable-architecture/
- NoA Ignite — How we plan for GenAI task by task: https://noaignite.com/insights/how-we-plan-for-genaitask-by-task/