Skip to content
Hakan KaracaENGINEERING JOURNAL
AI & Intelligent AutomationEki 8, 20265 min read

How to Start an Industrial AI Project: A 90-Day Engineering Plan

A useful industrial AI project can start small. It needs a narrow question, measurable value and disciplined execution.

An industrial AI project should begin with a process problem, a baseline and a person who owns the resulting decision. Starting with a vendor demonstration and then hunting for a use case is a reliable way to spend six months producing a dashboard that nobody opens.

The plan below is a pragmatic 90-day route for a bounded pilot. It is not a universal schedule: data access, validation requirements, procurement and production-change windows can extend the work. Its purpose is to make the decision gates explicit.

Days 1–15: define the decision

Choose one problem that has a measurable operational consequence: shorten fault investigation, prioritise maintenance, identify a quality risk earlier or improve a specific planning decision. Avoid goals such as “use generative AI” or “make production smarter”; they cannot be tested without further definition.

Interview the people who actually perform the work. Document the current process, decision authority, delay, cost of an incorrect recommendation and how success is measured. Define both benefit and risk. If no existing baseline is available, the first task may be to collect it rather than to train a model.

Gate 1: a written problem statement, process owner, baseline metric, expected action and a defined stop condition.

Days 16–30: audit the data and interfaces

Map the required signals, records and context. Confirm data ownership, access permissions, retention, units, timestamps, sampling rate, missing values and equipment identifiers. For MES-oriented work, link observations to order, batch or product context. For a control-system use case, confirm that reading the data will not affect controller performance or violate site policy.

Review data quantity and label quality, not only record count. One million samples from a steady-state machine may contain less useful information than a smaller dataset covering normal operation, changeovers, maintenance and known faults. Check whether the target label is observed reliably and whether timestamps are aligned to the actual event.

Gate 2: a data-quality report, approved access path, known limitations and a decision on whether the available evidence is adequate.

Days 31–45: establish the simplest baseline

Implement the existing rule, statistical baseline or manual process in a reproducible form. Split training and evaluation data according to time, asset or batch so that near-duplicate samples do not leak across the boundary. Define acceptance metrics that reflect consequences: false-accept and false-reject rates, lead time, precision/recall, prediction error or expert review time.

Only then compare candidate methods. The winning model is not necessarily the most complex; it is the method that improves the decision within the operational constraints. Record where it fails, how it behaves on incomplete data and when it should refuse to provide an estimate.

Gate 3: benchmark results on a fixed evaluation set, with documented limitations and a review by domain experts.

Days 46–60: design the integration and controls

Place the analytical service on a clearly defined boundary. Specify input schema, output schema, timeouts, authentication, logging, data retention and software/model version. Decide whether the initial deployment will be offline, in shadow mode or visible to a limited user group. Keep the existing control and quality authority unchanged until a separately approved change says otherwise.

Plan failure behaviour before implementing the happy path. What happens if the model is unavailable, the input is stale, the service times out or a result refers to a configuration that has changed? Define a fallback and make it visible to the user. A silent fallback to a default score is worse than a clear “no result”.

Gate 4: an approved interface design, threat/risk review, rollback plan and an agreed test protocol.

Days 61–75: test under realistic conditions

Use historical replay and controlled tests. Include normal cases, edge cases, data gaps, unexpected operating modes and cases where there is insufficient information. For a language model, test unsupported questions and source citations. For anomaly detection, test legitimate mode changes and maintenance activity. For a predictive model, evaluate error at the required lead time, not only at an arbitrary horizon that produces nicer numbers.

Run the system in shadow mode against live or representative data where permitted. Compare results with the current workflow without allowing the model to trigger unauthorised actions. Review disagreement cases with the engineers responsible for the process.

Gate 5: documented test evidence, resolved high-priority defects, trained users and explicit release approval.

Days 76–90: run a bounded pilot and decide

Limit the pilot to an asset family, shift, line or user group. Monitor model performance and operational outcome, not just whether the service remains online. Track user overrides, ignored recommendations, time saved, false alerts, latency, data freshness and support effort. Assign an owner for incident response and model maintenance.

At the end of the pilot, choose among three outcomes: scale with a controlled deployment plan, revise the use case and repeat a specific test, or stop because the benefit does not justify the cost and risk. Stopping a weak pilot is a successful engineering decision, not a failed AI project.

What must exist before production

  • An accountable process and system owner.
  • Versioned data and model/prompt artefacts.
  • Repeatable tests and acceptance criteria.
  • Defined security, access and audit controls.
  • Monitoring for drift, stale data and service failures.
  • A rollback, fallback and change-control procedure.
  • Training and a support route for users.

The point of a 90-day plan is not to force every project into three calendar months. It is to prevent enthusiasm from bypassing evidence. A modest model with a clear owner, a stable data contract and a tested fallback is a stronger result than a spectacular prototype whose only known operating mode is the demo.

References and further reading

  1. NIST AI Risk Management Framework
  2. NIST SP 800-82 Rev. 3: Guide to Operational Technology Security
  3. ISA-95 Series: Enterprise-Control System Integration