AI or deterministic automation: a decision matrix
Which steps of a process need rules, which need bounded AI, and which need an accountable person? A simple matrix, with failure modes.
“Should we use AI for this?” is usually the wrong question. A process is a chain of steps, and different steps deserve different logic. Some are fully specified and should behave identically every time. Some involve messy input that needs interpretation. A few carry consequences that a named person must own.
This article gives process owners and security stakeholders a way to sort the steps of one process into three lanes: deterministic rules, bounded AI, and accountable humans. It is about the logic choice. Where the capability comes from, whether you configure, buy, integrate or build, is a different decision, covered in configure, buy, integrate or build. What coordinates the lanes is the subject of what an agentic operating layer is.
The three lanes
| Lane | Use when | Behavior you can expect | Typical examples |
|---|---|---|---|
| Deterministic rules | Inputs are structured, the logic can be written down, and the same input must give the same output. | Repeatable, testable, cheap to run, easy to audit. | Required-field checks, calculations against agreed terms, routing by known status, state transitions, duplicate checks on exact identifiers. |
| Bounded AI | Inputs are unstructured or ambiguous, and some variation in output is acceptable if it is checked or reviewed. | Useful but variable; can be wrong with confidence; needs evaluation and limits. | Reading free-text correspondence, classifying a message, comparing two document versions, finding relevant context, drafting a recommendation. |
| Accountable human | The decision is consequential, contested, irreversible, or legally or commercially binding. | Slower, but with clear ownership and judgment. | Credit exceptions, concessions, write-offs, engineering sign-off, approval of a system update with material effect. |
Microsoft’s agent guidance, which is vendor documentation, supports the first lane directly: where steps are clear, repeatable and follow strict rules, it recommends regular code or non-generative models, describing them as faster, cheaper and more reliable than an agent for fixed workflows. Its strategy guidance draws the same contrast in another way: generative AI can produce different outputs for the same input and suits unstructured inputs, while non-generative AI gives consistent, repeatable outputs from structured inputs and suits defined workflows. Adding a model to a step that a rule handles exactly adds cost and uncertainty without adding value.
Why the boundary matters
NIST’s generative AI profile (voluntary guidance) names a risk that is easy to forget in a demonstration: confabulation, meaning confidently stated but wrong or false content. It also notes that generative systems enlarge the attack surface, for example through prompt injection, where crafted input changes how the system behaves. Neither risk exists in a rule that checks whether a field is present. Both are reasons to keep AI away from steps where an unnoticed error would be costly, unless a check, limit or review stands behind it.
The same profile expects post-deployment monitoring that includes appeal and override, incident response and recovery. Those mechanisms are easiest to design when each step’s lane is explicit.
A decision matrix for one step
Take each step in the process and answer these questions. Where the answers disagree, the safer lane wins until you have evidence.
| Question | If yes, lean towards | If no, lean towards |
|---|---|---|
| Can the logic be written as explicit conditions that cover the cases you see? | Rules | AI, with a rule or person behind it |
| Must the same input always give the same output? | Rules | AI may be acceptable |
| Is the input structured (known fields, codes, formats)? | Rules | AI for extraction or interpretation |
| Does the step require reading or comparing free text or documents? | AI (bounded) | Rules |
| Would a wrong output be costly and hard to detect? | Rule or person decides; AI only prepares | AI may act within limits |
| Is the action reversible, and is there a rollback path? | AI may propose or act within limits | Person approves |
| Is a legal, financial or engineering authority involved? | Accountable person | Rule or bounded AI |
| Can we measure accuracy on real examples, including exceptions? | AI may be tested and monitored | Do not deploy AI on this step yet |
| Does a stable pattern now appear in what AI handles? | Encode it as a rule; keep AI for the remainder | Keep AI under review |
Use the matrix per step, not per process. A single process often needs all three lanes.
Five patterns that work
- Rule first, AI on the remainder. Apply exact rules. Send only unmatched items to AI for classification or suggestion.
- AI proposes, rule validates. AI extracts values from a document; rules check them against authoritative data before anything is written.
- AI drafts, person approves. AI prepares a recommendation with its evidence; a named person decides.
- Rule gates the action. Whatever AI suggests, a deterministic permission check and limit decide whether an action may run.
- Exception queue with owner. Anything outside the confidence or rule boundary goes to a queue with a named owner, an age limit and a record of resolution.
A synthetic example
The following is illustrative and invented; it describes no client.
A customer disputes part of an invoice by email. Treat the handling as steps.
| Step | Lane | Why |
|---|---|---|
| Capture the email and link it to the customer and invoice by exact reference | Rules | Structured identifiers; must be exact. |
| Check whether the dispute falls within the contractual window | Rules | Calculation against agreed terms. |
| Read the message and classify the reason (pricing, delivery, quantity, other) | Bounded AI | Free text; category list is fixed and results can be reviewed. |
| Compare the claim against delivery evidence and price terms | Rules for lookups, AI for document comparison | Lookups are exact; documents differ in layout. |
| Draft a recommended resolution with its evidence | Bounded AI | Preparation only. |
| Approve a credit note above an agreed threshold | Accountable person | Financial consequence. |
| Update the ledger and notify the customer | Rules | State change after approval, with audit entry. |
| Anything unclassified or low confidence | Exception queue with owner | Outside the boundary. |
AI appears in three of the eight steps, and always in a bounded role with a rule or a person behind it. That is typical, and it is a design outcome, not a target to maximize.
Failure modes
| Failure mode | How it shows up | Safeguard |
|---|---|---|
| AI used where a rule would do | Higher cost, inconsistent results on a step that should be exact. | Default to rules for structured, specified steps. |
| Rules stretched over ambiguity | A growing pile of special cases and brittle conditions. | Move the ambiguous portion to bounded AI with review. |
| Confident wrong output | A plausible classification or figure is accepted unchecked. | Evaluate on real exceptions; validate against authoritative data before action. |
| Over-trust in the recommendation | Reviewers click approve without reading. | Show evidence with the recommendation; sample-review approvals; measure overrides. |
| Hidden write access | A model-driven step can change records without a gate. | Permission checks, action limits and audit on every write. |
| Prompt-injected input | Text inside a document tries to steer the system. | Treat document content as data, not instructions; restrict tools and permissions. |
| No owner for exceptions | The queue grows and nobody is accountable. | Named owner, age limit and escalation rule. |
| Drift | Accuracy slowly falls as inputs change. | Monitor on a schedule, keep a sample set, define retraining or rule-update triggers. |
What to ask before approving an AI step
- What is the input, the output and the action it may trigger?
- What is the accuracy on real examples, including the awkward ones, and how was it measured?
- What happens when it is wrong, and who notices?
- What can it read and write, and is that limited per action?
- Can a person override it, and is every override recorded?
- Who owns monitoring, incidents and rollback?
What this does not prove
The matrix is a decision aid, not a measurement. It does not establish that any given step will reach a particular accuracy, cost or speed, and its yes/no questions simplify judgments that are often graded. The cited material is vendor and voluntary-standards guidance, and the example is synthetic. Any real classification of steps needs your data, your exceptions and a test against your own acceptance criteria. Nothing here is legal advice on regulatory obligations such as those under the EU AI Act, which depend on the system, its use and your role.
Where Sembium fits
Sembium designs whole operating processes: deterministic rules for known paths, bounded AI where interpretation helps, and accountable people for consequential decisions, with permissions, failure handling, audit and handover. Producing the lane assignment for each step is part of the redesign phase; see the approach for how the work is staged.
Got a bottleneck that costs you $1M or more, and you want it solved? Book the 30-minute bottleneck call. Bring one process. You leave with one practical recommendation and a clear answer on fit.
Questions buyers ask
When should a process step use rules instead of AI?
When the inputs are structured, the logic can be written down and the same input must always give the same result. Validations, calculations, routing by known fields and state changes usually belong here.
Where does AI add value in operations?
Where the input is unstructured or ambiguous: reading correspondence, comparing documents, classifying free text, retrieving relevant context or drafting a bounded recommendation for a person to review.
Does AI have to be the decision-maker?
No. In a well-designed process AI prepares, classifies or recommends, and a rule or an accountable person decides. Consequential financial, commercial or engineering decisions need an explicit human owner.
Can a step start with AI and move to rules later?
Often yes. If AI handling reveals a stable pattern, encode it as a rule, test it and keep AI for what remains ambiguous. That usually reduces variability and cost.
Sources
Sources give context. Dates show when each was last checked.