Decision framework · AI automation
What business processes should you automate with AI?
The best AI automation opportunity is rarely the most ambitious process. Start with work that happens frequently, uses accessible digital inputs, has an observable definition of done, and can fail safely while the system earns trust.
Cognate Automation Scorecard v1.0
Score value and feasibility from one to five, then treat risk as a gate rather than subtracting it from an attractive total. The anchors below give reviewers a common basis for comparing their reasoning; they are not evidence of statistical inter-rater reliability.
| Dimension | 1 · low | 3 · moderate | 5 · high |
|---|---|---|---|
| Value | Rare task; little delay or rework | Weekly work with measurable handling time | Frequent bottleneck tied to cost, revenue, or service |
| Feasibility | Inputs inaccessible; outcome subjective | Most inputs available; exceptions need review | Digital inputs, representative cases, testable output, available integration |
| Risk | Internal, reversible, no sensitive effect | External communication or limited sensitive data; approval can contain harm | Money, rights, safety, deletion, privacy, or irreversible commitment |
Use 2 when the evidence falls between the 1 and 3 anchors, and 4 when it falls between 3 and 5. Record the observations behind every score; a number without evidence is not an assessment.
| Illustrative candidate | Value | Feasibility | Risk | Decision |
|---|---|---|---|---|
| Inbound supplier-invoice intake | 4 · daily handling and rework | 4 · digital files and checkable fields | 3 · financial data, but posting stays behind approval | Pilot extraction and duplicate checks; require approval before accounting entry |
The example is hypothetical, not a client result. The scorecard supports a go/no-go discussion; it is not a business case or a safety certification. Observe real work and sample actual inputs before accepting any score. The NIST AI RMF Manage 1.1 guidance likewise recommends weighing benefits and risks when deciding whether AI is appropriate for a task.
What makes a strong AI automation candidate?
A strong candidate contains a specific bottleneck that conventional software cannot resolve cleanly. Documents arrive in inconsistent formats. A person must interpret a request before it can be routed. A report requires synthesis across several sources. The AI capability has a named job inside the process rather than a vague instruction to “make the workflow intelligent.”
| Signal | Stronger candidate | Weaker candidate |
|---|---|---|
| Frequency | Occurs daily or weekly | Rare, exceptional work |
| Inputs | Digital and accessible | Scattered, undocumented, or mostly offline |
| Variability | Language or documents vary within a known domain | Every case requires novel strategy |
| Outcome | Correctness can be checked | Success is subjective or delayed for months |
| Examples | Accepted and rejected cases exist | No record of past decisions |
| Failure | Detected and recoverable | Silent, irreversible, or safety-critical |
High volume alone is not enough. A process with inaccessible data, no owner, and no agreement about the correct outcome will remain unreliable after AI is added. Automation often exposes operational ambiguity rather than solving it.
Examples of useful starting workflows
| Workflow | AI contribution | Deterministic controls | Human boundary |
|---|---|---|---|
| Inbound request triage | Classify intent and extract details | Required-field validation and routing | Approve sensitive or ambiguous cases |
| Document intake | Extract and normalise variable fields | Schema checks and duplicate detection | Review low-confidence fields |
| Weekly reporting | Explain changes and draft narrative | Scheduled collection and arithmetic | Approve conclusions and actions |
| Knowledge response drafting | Retrieve and compose an answer | Access control and source allow-list | Send externally or revise |
| Creative quality review | Assess an asset against stated criteria | File validation and version tracking | Make the final brand decision |
AI does not own every step in these examples. Deterministic code controls stable operations; the model handles a named interpretation task; and a person retains the consequential judgment. Cognate Labs prefers this separation because each boundary has an observable input, output, and owner.
Processes to avoid as a first project
- Undefined executive judgment: “Make the right strategic decision” has no testable boundary.
- Broken source processes: AI will not repair missing ownership, inconsistent policy, or inaccessible records.
- Rare, catastrophic decisions: there may be too little evidence to test and too much downside to learn in production.
- Unreviewable output: if a qualified person cannot verify the result, errors can accumulate silently.
- Automation for appearance: a demo without a workflow owner or adoption path becomes another abandoned tool.
Sometimes the correct recommendation is conventional automation, a better form, a database cleanup, clearer policy, or no automation at all. The comparison of automation, chatbots, agents, and AI employees can help identify the smallest appropriate system.
Measure the baseline before building
A credible business case begins with the current process. Sample enough cases to capture normal work and exceptions. Record active handling time, waiting time, throughput, correction rate, escalation rate, and any direct external cost. Separate estimates from observed measurements.
- How many cases arrive in a normal week, and how uneven is the volume?
- How long does a person actively work on each case?
- Where does work wait for information, approval, or handoff?
- What percentage requires correction or rework?
- Which exceptions consume disproportionate attention?
After release, compare like with like. Include review time, model and infrastructure cost, maintenance, and the cost of mistakes. Faster output is not automatically valuable if it creates more review or moves a bottleneck downstream.
Move from candidate to controlled pilot
01
Name one outcome
Define the output, customer, and current baseline.
02
Map the workflow
Record triggers, inputs, systems, decisions, exceptions, and owners.
03
Assemble an evaluation set
Include routine cases, difficult cases, and cases the system must refuse or escalate.
04
Run without authority
Compare proposed outputs with accepted outcomes before changing production records.
05
Release one bounded action
Automate only after the evaluation target and rollback path are explicit.
This is the core of a practical AI consulting engagement: turn a broad ambition into a viable workflow, then make the evidence—not enthusiasm—decide whether it should progress.
Common questions
Which department should automate with AI first?
Choose by workflow rather than department. Look for frequent digital work with variable language or documents, accessible examples, observable outcomes, and recoverable mistakes. The best first opportunity may cross several teams.
Should we automate the entire process?
Usually not at first. Isolate the step where AI adds value and keep stable steps deterministic. Begin with recommendation, drafting, or approval modes before granting autonomous authority.
How do we calculate whether AI automation is worthwhile?
Measure the current volume, active handling time, delay, correction rate, and external cost. Compare those with the operated system’s review time, model cost, maintenance, error cost, and resulting business outcome. Label estimates clearly until production evidence exists.