Supported directly by customer/system evidence.
Coral Shark Lab · DETERMINE / INTERPRET / DECIDE / ACT
How Can AI Help My Business? AI vs Workflow Automation
Direct answer
AI is useful for reading varied enquiries, documents or messages. Rules are better for exact checks and calculations; people should control prices, promises, disputes and other important decisions.
DETERMINE / INTERPRET / DECIDE / ACT
Classify the step before choosing the technology
A calculation and a judgement are not the same problem. A classification and an external action are not the same risk.
DETERMINE / INTERPRET / DECIDE / ACT
Use AI for the uncertainty you actually need
This pattern is easier to test, replace and control than giving an agent a broad objective plus wide system access.
DETERMINE / INTERPRET / DECIDE / ACT
Where AI adds value
AI is often useful when language or documents vary widely and deterministic parsing would be brittle: classify an enquiry, extract candidate facts, summarise context, identify missing information or draft bounded text.
DETERMINE / INTERPRET / DECIDE / ACT
Where AI adds risk without much value
- Exact arithmetic, lookups or fixed rules.
- Critical business truth that belongs in a system of record.
- High-consequence decisions.
- Actions without a validation/stop boundary.
- Undefined processes where improvisation hides missing rules.
- Tasks already solved by existing features.
- Outputs the business cannot reliably evaluate.
DETERMINE / INTERPRET / DECIDE / ACT
AI output is an evidence class, not a fact replacement
A supplied estimate whose source stays visible.
Derived interpretation or recommendation.
Not established by current evidence.
If a model infers that quote preparation probably contains duplicate entry, that does not rewrite the customer’s evidence into a fact.
DETERMINE / INTERPRET / DECIDE / ACT
An AI agent does not replace workflow design
Greater flexibility increases the need to define tool access, data access, approvals, completion, stop conditions, authoritative state, retries, evidence and fallback. The more flexible the component, the more deliberate the control envelope must be.
DETERMINE / INTERPRET / DECIDE / ACT
Synthetic enquiry-follow-up architecture
- Receive enquiry — deterministic trigger.
- Acknowledge receipt — deterministic template + idempotent ACT.
- Extract candidate facts — bounded INTERPRET; AI may help.
- Check completeness — deterministic DETERMINE.
- Ask bounded missing-information question — controlled ACT.
- Decide fit/price — human DECIDE unless explicit rules truly determine it.
- Send final response/booking — deterministic ACT after approval.
The workflow may use AI in one or two places. It is still mostly workflow automation—deliberately so.
Evidence and claim boundary
What this page is — and is not.
DETERMINE / INTERPRET / DECIDE / ACT is a Coral Shark decision framework, not a claim of scientific or industry consensus. Synthetic examples are illustrative, not customer proof. Where a material external fact is added later, it should be sourced to a current primary source and dated.
Related reading
Need this applied to one real job?
The Owner’s Brief applies the same discipline to one recurring job: what you told us, your estimates, our interpretation and what is still unknown stay separate. The answer can be stop, wait, start smaller or proceed.
Apply this to one job · £249 + VAT