AI AUTOMATION · PRIORITISATION

AI business automation: how to choose a workflow and measure impact

The best first workflow is not necessarily the most visible. It recurs, has understandable inputs and an owner, allows its output to be checked, and has a manageable error cost. Impact is measured against a recorded baseline process.

GUIDE · UPDATED August 5, 2026

01 · DIRECT ANSWER

Direct answer

For a first AI automation, choose a frequent and sufficiently stable workflow with an observable outcome, accessible data, an accountable owner, and reversible errors. Compare AI with conventional automation first, then run a narrow pilot and measure time, manual touches, quality, exceptions, and total cost—without promising savings before evidence exists.

02 · CRITERIA

Criteria before development

Apply these criteria to the specific workflow, data, and cost of error. They are not a universal readiness checklist.

01

Repeatability

The workflow recurs and has recognisable inputs, steps, and outputs. A rare or constantly changing exception is a poor first pilot.

02

Observable outcome

Correct completion can be defined in advance and the new flow can be compared with the current one.

03

System and data access

APIs, documents, roles, security constraints, and source quality are understood; the process owner can authorise a pilot.

04

Manageable error cost

An output can be reviewed, reversed, or sent to a person before an irreversible action. The higher the consequence, the narrower the AI role should be.

05

Economic lever

The workflow has measurable labour, delay, queue, quality loss, or another baseline that can be recorded before the change.

03 · PROCESS

Validation sequence

  1. 01

    Map the current flow

    Record the trigger, inputs, actors, systems, decisions, exceptions, output, and next action before designing the AI solution.

  2. 02

    Measure the baseline

    Measure selected indicators for comparable current cases and document data quality together with the collection method.

  3. 03

    Compare solution types

    For each step, consider removal, simplification, conventional rules, and integration before adding a probabilistic AI component.

  4. 04

    Run a narrow pilot

    Limit inputs, users, and permissions; retain an action log, human review, and a fast path back to the original process.

  5. 05

    Compare impact and cost

    Compare not only speed but quality, exceptions, human review, support, infrastructure, and newly introduced risk with the baseline.

04 · DECISION

Decision table

Priority depends on value, measurability, and manageable risk together—not solely on the amount of work that might be automated.

Workflow characteristicDecisionBasis
Rules are stable and inputs are structuredConventional automation or integrationA predictable mechanism is easier to verify and maintain.
Inputs are unstructured and output is easy to reviewAI prepares a draft or classificationAI reduces manual work while a person controls the error.
The action is irreversible or high-impactAI only assembles contextThe final decision and send action remain with an authorised person.
The workflow is rare, unstable, and has no ownerDo not automate it firstFormalisation and support cost may exceed the available benefit.
The workflow is frequent, measurable, and errors are reversibleCandidate for a bounded pilotA baseline exists and real evidence can be gathered safely.

05 · MEASUREMENT

What to measure

Record the baseline workflow and collection method first. Then compare equivalent scenarios without presenting a planned target as measured impact.

01

Cycle time

Measure from the same trigger to the same useful output, including waiting and hand-offs between systems.

02

Manual touches

Count active human time for preparation, review, correction, and exception handling—not merely the number of automated steps.

03

Quality and exceptions

Record correct outcomes, error types and severity, abstentions, escalations, and repeated work.

04

Throughput and queue

Compare completed equivalent cases and waiting time without assuming that higher volume automatically means greater value.

05

Total cost

Include model, infrastructure, integrations, human control, support, incidents, and process change.

06 · BOUNDARIES

Risks

01

Automating a poor process

AI can accelerate an unnecessary step. Review the process map before translating the current sequence into code.

02

Displaced accountability

Without explicit roles and an audit trail, staff may rely on system output without authority or review.

03

Integration and access risk

Broad permissions, tokens, and automated actions increase failure impact; access should be minimal and testable.

04

Local optimisation

Speeding up one step can move the queue downstream, increase review work, or reduce final quality.

05

Invalid comparison

Different case types, seasonality, or team changes can create a false effect when pilot and baseline are not comparable.

WHAT CANNOT BE CLAIMED

Limitations of the conclusion

  • Before measuring the baseline, no savings percentage, return, or payback period can be promised honestly.
  • Evidence from one narrow pilot does not automatically transfer to other teams, channels, or input types.
  • AI automation requires ongoing exception handling, integration maintenance, and repeat quality evaluation.
  • Legal, financial, medical, and other high-impact actions require qualified human oversight.
  • Greater speed is not an impact by itself if quality, safety, or the final outcome deteriorates.

07 · PORTFOLIO

Related projects and honest stage

Marketing Optimisation is shown as a technical MVP using synthetic data. It demonstrates a measurement workflow, not validated client impact or a production replacement for a CVM function.

Current stage: Technical MVP

Marketing Optimisation

The technical MVP implements one end-to-end CVM reactivation workflow and has been tested on synthetic data. The next step is a design-partner pilot with real CRM data; a production replacement for the department and commercial impact have not yet been validated.

Open project review

08 · METHODOLOGY

Primary sources

These sources support the methodology and definitions. They do not validate AI SaaS Solution project outcomes.

  1. Official publication · NIST

    Artificial Intelligence Risk Management Framework (AI RMF 1.0)

    Open source
  2. Official publication · NIST

    Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

    Open source
  3. Research · Microsoft Research / CHI

    Guidelines for Human-AI Interaction

    Open source