Skip to content

Why Deterministic Systems Still Matter in the Age of AI

AI is powerful partly because it can interpret information that is difficult to encode as fixed rules.

That does not mean every problem should be handed to a model.

A trustworthy AI product should first ask a simpler question:

Can this be established directly?

When the answer is yes, deterministic software is usually the better tool.

Do not ask AI to rediscover stable facts

A language model may recognize that a URL belongs to a survey platform, that an attachment is a PDF, or that a calendar invitation contains a video-conference link.

But those facts can often be established through parsing and lookup.

Using a model for them introduces avoidable problems:

  • the same input may produce different answers;
  • the model may overlook an obvious structural clue;
  • the result is harder to test precisely;
  • the explanation is less direct;
  • processing costs more time and energy;
  • private content may be exposed unnecessarily when a local rule would suffice.

Spillway's principle is simple:

If something can be known reliably, know it directly. Use AI for the ambiguity that remains.

Deterministic does not mean simplistic

A deterministic system can do far more than match a few keywords.

It can combine:

  • structured message headers;
  • exact sender and recipient relationships;
  • URL hosts, subdomains, paths, query-key names, and identifiers;
  • attachment filenames, extensions, MIME types, and metadata;
  • calendar organizers, attendees, recurrence, locations, and conferencing fields;
  • exact and normalized dates;
  • thread direction and response state;
  • reviewed source catalogs;
  • prior explicit user decisions.

These facts can also be related.

A calendar invitation containing a Zoom link is more informative than either fact alone. A PDF named invoice.pdf from a known utility is different from an arbitrary PDF. A date near the phrase "applications close" has a different role from a date in a quoted news article.

Deterministic observation can establish the components and relationships. AI can then interpret the ambiguous meaning.

Rules are strongest when their boundaries are honest

Fixed rules fail when they try to answer questions beyond what they can support.

A domain lookup may establish that a link belongs to a bank. It should not automatically conclude that the message is urgent, legitimate, or personally important.

A date parser may identify September 15. It should not automatically declare that date a deadline.

An attachment parser may identify a spreadsheet. It should not assume the spreadsheet is a budget.

Good deterministic systems emit bounded evidence:

  • what matched;
  • at what level of specificity;
  • from which source;
  • with what confidence or support;
  • under which version of the rule or knowledge package.

Interpretation happens later.

Why rules alone are not enough

Some questions resist exhaustive rules because their meaning depends on language, relationships, history, and context.

For example:

  • Is this request polite background information or something that requires action?
  • Is this deadline consequential to the user?
  • Does the sender expect a response?
  • Is an attachment the central work item or merely supporting material?
  • Are several messages part of the same project?
  • Which of several plausible categories best fits the user's way of working?

A fixed-rule system can cover common patterns, but each new exception multiplies complexity. Eventually the rules become brittle, opaque, and difficult to maintain.

AI is valuable precisely because it can synthesize these interacting signals without requiring every possible combination to be encoded in advance.

A layered division of labor

The strongest design is not deterministic software versus AI. It is a division of labor.

Deterministic systems

Use them to:

  • extract stable structure;
  • normalize equivalent forms;
  • resolve reviewed facts;
  • enforce permissions and safety rules;
  • preserve identity and provenance;
  • detect exact state changes;
  • apply explicit user preferences.

AI systems

Use them to:

  • interpret ambiguous language;
  • combine incomplete evidence;
  • recognize semantic similarity;
  • propose classifications and summaries;
  • identify possible requests, commitments, and relationships;
  • reason where exhaustive fixed rules are impractical.

Humans

Rely on them to:

  • make consequential decisions;
  • correct facts and interpretations;
  • define personal meaning and importance;
  • resolve conflicts the software cannot settle responsibly.

Each layer should contribute without pretending to be another.

Deterministic context can make smaller AI better

Local models are attractive because they can keep sensitive information on the user's device. But smaller local models have less capacity than the largest cloud systems.

Deterministic context helps close that gap.

Instead of asking a local model to infer everything from raw text, Spillway can provide compact reviewed context:

Known facts:
- Sender organization: University of Michigan
- Link platform: Qualtrics
- Resource type: survey form
- Attachment: none
- Thread state: unanswered request

The model can spend its limited capacity answering the hard question—what the message means and what the user may need to do—rather than rediscovering structural facts.

That can improve accuracy, reduce prompt size, and make the model's contribution easier to understand.

Deterministic systems improve explainability

A suggestion supported by explicit facts can be explained more honestly:

Spillway recognized a survey link, found request language in the message, and proposed that a response may be needed.

That is stronger than:

The AI thought this looked important.

Deterministic evidence creates anchors. It does not make the final judgment infallible, but it shows which parts were observed directly and which parts were inferred.

Deterministic systems improve correction

Different errors require different fixes.

  • If the URL was parsed incorrectly, fix the parser.
  • If a source catalog contains the wrong organization, correct the catalog.
  • If the model misunderstood the request, improve the inference stage.
  • If the system misunderstood personal importance, preserve a user-specific correction.

A system that collapses all of these into one model output cannot learn precisely from failure.

A layered system can.

The goal is not maximal automation

The goal is dependable assistance.

That means using deterministic mechanisms where they provide certainty, AI where it provides flexible interpretation, and human judgment where meaning and consequences remain personal.

AI does not make traditional software obsolete.

It makes careful software architecture more important.

Read how Spillway separates facts, evidence, and knowledge · Read the approach to explainable AI