Frequency
Does this step repeat often enough for automation to remove meaningful work?
All Guides
01 02 03 04
Guides
What should be automated - and what should stay manual?
A practical framework for choosing between deterministic automation, AI-assisted steps, and human decisions in a service-business workflow.
Guide by Ian Kirs · ARCHKIRS
Short answer
Automate work that is repetitive, rule-based, and expensive to repeat manually. Keep judgement, exceptions, commercial commitments, and high-consequence decisions under human control.
Use deterministic logic when the rule can be stated exactly. Use AI only where the input is genuinely unstructured - language, conversations, or documents - and where the output can be checked or bounded. AI is not a reason to automate a process.
Choose the simplest reliable control
| Task | Best default | Why |
|---|---|---|
| Store submitted form fields | Deterministic automation | The data and destination are known. |
| Send routine notifications | Deterministic automation | The trigger and recipient can be defined exactly. |
| Calculate known prices or quantities | Deterministic automation | Repeatable arithmetic should not depend on a language model. |
| Approve scope, price, or contract terms | Human decision | These are commercial commitments and may require judgement. |
| Summarize a long project discussion | AI-assisted | Natural-language compression can save time if the original remains available. |
| Translate project messages | AI-assisted | Language is unstructured; preserving the original keeps the result auditable. |
| Enforce project access and isolation | Deterministic automation | Security boundaries must be explicit and testable. |
| Handle an unusual exception | Human with system support | Exceptions are where rigid automation can create the wrong outcome. |
Four useful tests before automating
Rule clarity
Can you explain the trigger, inputs, output, and exceptions clearly?
Error cost
What happens if the system makes the wrong decision or sends the wrong result?
Auditability
Can a person later understand what happened and correct it?
Good deterministic automation candidates
- Moving known fields between an enquiry and a client or project record.
- Sending a notification after a defined event.
- Creating a document from approved structured data.
- Calculating values from explicit formulas.
- Updating a status after a verified action.
- Checking required fields, permissions, or fixed workflow rules.
Decisions that usually deserve a person
- Whether a project is a good fit.
- Final scope, price, schedule, and contract commitments.
- Sensitive approvals and exceptions.
- Changes that affect a client's rights, access, payment, or acceptance.
- Ambiguous situations where the cost of a wrong automated action is high.
Where AI can add real value
- Summarizing long conversations while keeping the source available.
- Translating messages while preserving the original.
- Extracting a draft structure from unstructured text for later review.
- Answering factual questions from a bounded and verified evidence source.
- Helping a person review or prepare information without making the final commercial decision.
Where ordinary logic is usually better than AI
- Permissions and security boundaries.
- Known calculations and pricing formulas.
- Required-field validation.
- Fixed notifications and deterministic routing.
- Document version identifiers and immutable records.
- Any rule where the correct result can already be written precisely.
Related proof
ARCHKIRS examples: deterministic core, bounded AI
ARCHKIRS.SITE uses deterministic structures for enquiries, client and project records, commercial calculations, document versions, and protected private data.
Project Workspace uses server-side AI for tasks such as translation, summaries, transcription, and repository-grounded questions, while authorization, project isolation, evidence boundaries, and final owner decisions remain explicit system controls.
Map one real process before choosing tools
Take one repeated workflow and mark each step as data transfer, fixed rule, human judgement, or unstructured interpretation.
That usually makes the right split between ordinary automation, AI-assisted work, and human control much clearer.