Automation and operations

Which business process should you automate first? A practical scorecard

A risk-gated scorecard for selecting a measurable, recoverable first pilot

Modular cube in three-quarter perspective with selected violet blocks and detached fragments around a black, white and grey structure.

Three candidates are competing for attention: route incoming requests, prepare the weekly report, or transfer order data into the operations system. One has more volume, one is simpler, and one appears to consume more time. Which should go first? Do not settle it by adding hours and frequency. First exclude candidates without ownership, a baseline, controlled data or a recovery path. Only then should a score influence priority.

A priority score is not automation approval

For a first automation, M4TIC recommends a repeated process with a clear trigger, inputs, accountable owner and acceptable result; identifiable exceptions; a baseline for current performance; and a practical way to detect, stop, correct or route failures back to a manual path.

Volume is useful evidence, but it is not permission. A high-volume process can be a poor first project when nobody owns the result, rules change informally or an error can move money, alter a sensitive record or affect operations before anyone notices. A moderate-volume process may be the stronger candidate when its inputs, controls and outcomes are visible.

The decision has two stages:

  1. four mandatory gates decide whether a candidate may enter the comparison;
  2. seven 0–3 criteria help rank the candidates that passed.

A failed gate is not a low score. It is a redesign requirement. It cannot be averaged away.

Why volume and frustration are not enough

Automation discussions often begin with “Where do we lose the most hours?” That is a useful question, but not a complete one.

Reported hours may mix active work, waiting, corrections and steps that should not exist. A case may take two days to cross a business without consuming two days of labour. Before treating cycle time as potential savings, observe a representative period and separate:

  • frequency and volume;
  • active work time;
  • waiting and dependencies;
  • rework and errors;
  • exceptions and human intervention;
  • the consequence of a wrong result.

The U.S. federal Eliminate, Optimize, Automate guidance puts current-state mapping and simplification before technology selection. The sequence travels well: automating a duplicate approval, an unused field or a workaround for unclear accountability preserves the defect and makes it run faster.

Describe the candidate as trigger → work → output, not as a department. “Automate finance” is not a use case. “When an invoice arrives, extract its fields, match it to the order and prepare discrepancies for approval” is a boundary that can be examined.

Inventory, eliminate and simplify before scoring

Collect five to ten processes, then prepare a compact record for each. The observations do not need to be perfect; they need to be concrete enough to expose missing knowledge.

  • Process: trigger, main steps and output.
  • Owner: who owns the result and the go/no-go decision.
  • Workload: volume, frequency, active work and waiting.
  • Rules and exceptions: standard path, edge cases and required judgement.
  • Data and systems: sources, quality, permissions, sensitivity and dependencies.
  • Error consequence: people, records, money or operations that may be affected.
  • Baseline: quality, time, rework or service before the change.

For each step, ask whether it creates value, satisfies a necessary control or merely exists because the process grew around a workaround. Elimination or simplification may solve the problem without automation.

Four mandatory gates before any score

A candidate enters the scorecard only when every gate is yes, or when a short, owned and verifiable action can bring it to yes.

Scroll sideways inside the table to compare every column

The four gates a candidate must pass before scoring
GateQuestionIf the answer is no
Accountable ownerIs one person accountable for the result, exceptions and go/no-go decision?Assign ownership and decision rights first.
Result and baselineCan the correct output be defined and at least one current measure obtained?Observe and measure the current process before the pilot.
Containment and recoveryCan an error be detected, stopped, corrected or approved before serious effect?Reduce the scope or keep approval and execution human.
Data and controlsAre the required data available, authorised and governed with proportionate controls?Resolve quality, access, privacy, security and compliance first.

Hard stops include no accountable owner, a high-consequence irreversible action without approval, uncontrolled sensitive data or a result that nobody can define. The candidate may return with a smaller scope or a redesigned process. A high opportunity score cannot rescue it.

These gates are an M4TIC editorial synthesis informed by public guidance on process selection, accountability, measurement, oversight and recovery. They are not a compliance certificate or a scientifically validated universal model.

The M4TIC First-Automation Scorecard

After the gates, score each criterion from 0 to 3. Every number should have an observation or note behind it.

Scroll sideways inside the table to compare scores from 0 to 3

Seven criteria in the M4TIC First-Automation Scorecard
Criterion0123
Repeated workloadRare or negligiblePeriodic but lightFrequent or burdensomeFrequent and materially burdensome
Outcome valueVague benefitLocal convenienceMeasurable quality, time or serviceImportant, measurable operating problem
Process clarity and stabilityDisputed or changing flowPartly describedFairly stable, clear boundaryDocumented trigger, output, rules and handoffs
Input readinessScattered or inaccessibleHighly manual or inconsistentUsable with preparationStructured, accessible, known quality
Exceptions and judgementInvisible exceptions or judgement dominatesMany uncodified exceptionsRecognisable exceptions with escalationStandard path dominates; exceptions measured
Observability and recoveryError hidden or not correctableLate visibility, costly recoveryReasonable logs and correctionClear checks, alerts, override and rollback
Delivery burdenMany fragile or uncertain integrationsSeveral material dependenciesManageable integrationExisting system or simple delivery path

The maximum is 21, with no universal weights and no automatic pass mark. The total supports comparison among admitted candidates. It is not proof of ROI, safety, compliance or economic value.

A candidate below 2 on exceptions and judgement or observability and recovery is not suitable as a first pilot until the weakness is reduced. When totals tie, prefer the lower consequence, lower delivery burden and faster feedback cycle.

Weights depend on context. Response time may dominate in a service queue; error consequence dominates in payments. If an organisation adds weights, it should document them as its own decision and test how they change the ranking. Hiding them inside a formula creates false precision.

Five illustrative candidates

These are illustrations, not client cases or evidence of performance. They show how gates and scores can change the decision.

Inbound request triage with human review

This can reach the shortlist when categories are defined, wrong labels are visible and uncertain cases route to a person. The baseline should include volume, time to assignment, reclassification and escalation. A first scope can propose a category without sending a response or changing operational priority.

Weekly report preparation from known sources

Moderate volume does not make it unimportant. Verifiable inputs, stable rules, a reviewable output and easy recovery may make it a better first pilot than a higher-volume but opaque process. Measure active work, not the full calendar interval between opening and delivery.

Order-data transfer between approved systems

The candidate is promising when fields, identities and validation rules are stable. Duplicate handling, logs, reconciliation and a path for incomplete records are prerequisites. If exceptions are still resolved through oral knowledge, the process is not ready.

Automatic invoice payment

High workload cannot override the recovery gate. The financial effect may be difficult to reverse and needs explicit authority. A proportionate first scope may collect documents, validate completeness and prepare discrepancies while approval and execution remain human.

Payroll changes from free-form email

Sensitive data, ambiguous language, authorisation and high consequence make this a poor first candidate without redesign. A volume score cannot compensate for that boundary.

Choose the minimum sufficient technology

The scorecard selects a process, not a product. Choose technology only after the pilot boundary is defined.

  • Fixed rules or workflow: the path is known, inputs are structured, logic is stable and the same result is expected. This is the default when exact repetition and auditability matter.
  • Generative AI: a bounded step handles unstructured text or content, more than one output can be acceptable and review is practical. What generative AI can and cannot do clarifies the line between generation, retrieval, rules and action.
  • AI agent: the route varies and the system must choose among approved steps or tools. Testing, authority boundaries and recovery become more demanding; first confirm whether an agent is justified.
  • No automation yet: the process is unstable, ownerless, poorly supplied with inputs, missing a baseline or unable to contain failure.

The least complex technology that performs the controlled job is usually the better starting point. A model or agent does not repair undefined rules.

Run a pilot that creates evidence

Before building, collect the current-state map; roles and decision rights; representative samples and edge cases; volume, active time, waiting, errors and rework; rules and policies; data provenance, quality and sensitivity; identities, permissions and system dependencies.

The pilot should state:

  • scope and allowed inputs;
  • acceptance criteria and cases for escalation;
  • approvals that remain human;
  • logs, alerts and measures;
  • manual fallback;
  • rollback or correction path;
  • the condition that stops the test;
  • full cost, including review, maintenance and exceptions.

If the candidate is agentic, the next step is risk-based pre-launch testing, not an immediate expansion of autonomy.

The result is a decision, not a victory announcement

A first pilot succeeds when, within its stated scope, it meets the process-specific acceptance criteria, improves at least one baseline measure without weakening an essential control, exposes errors and exceptions, and remains stoppable and reversible by the owner.

Check where the work went. A reported time reduction is not realised value if review, correction or exception handling moved to another person. Include control, maintenance and integration effort rather than counting only software calls.

The conclusion may be extend, revise or retire. Retiring a pilot can be useful evidence when it prevents a fragile process from being scaled. The first automation does not need to prove that every process can be automated. It needs to make the next decision less dependent on assumptions.

Sources and further reading

Primary guidance behind this article.

  1. RPA Program Playbook v1.1U.S. Federal RPA Community of Practice

    Public guidance on process selection and assessment, including suitability, alignment and impact.

  2. Federal Eliminate, Optimize, Automate Playbook v1.0U.S. General Services Administration

    Supports current-state mapping, performance data, stakeholders, rules, systems and go/no-go decisions before technology selection.

  3. ISO 9001 explainedInternational Organization for Standardization

    Process context covering inputs, outputs, responsibilities, monitoring and performance evaluation.

  4. Artificial Intelligence Risk Management Framework 1.0 (NIST AI 100-1)National Institute of Standards and Technology

    Principles for context, roles, measurement, oversight, monitoring and risk response in AI systems.

  5. AI RMF PlaybookNIST AI Resource Center

    Suggested practices for oversight, metrics, logs, overrides, response and recovery.