Many organisations begin an assistant project by choosing a language model. That is often too early. The more useful first question is: what does the user need, and how much variation is acceptable in the answer?
If the correct response can be derived from known facts and business rules, clear logic may be more dependable, faster and less expensive than generation. This is the principle behind the mccaigs philosophy: "Don't pay for intelligence when logic is enough."
What deterministic means in an assistant
Deterministic does not mean that every screen must be rigid or that the assistant cannot understand natural language. It means that important outcomes are controlled.
The assistant can recognise a user's intent, gather the information needed and select an approved response path. If a price, eligibility rule, policy or process is already known, the system can retrieve or calculate the answer directly. The same valid input produces the same business outcome.
This makes the assistant easier to test. A team can create expected inputs, expected answers and expected fallbacks. When the underlying information changes, the approved source can be updated without rewriting the whole conversation.
Why businesses need more control than a general chatbot
A general chatbot is designed to be flexible. Flexibility is useful when a person wants to explore ideas, rewrite text or discuss an open question. It becomes risky when the answer commits a business to a price, gives regulated guidance, interprets a policy or directs a customer through an important process.
Common failure points include:
- answering beyond the available evidence;
- combining two policies incorrectly;
- presenting an old answer as current;
- failing to distinguish guidance from a binding decision;
- giving different answers to equivalent questions;
- hiding uncertainty behind confident language.
A deterministic design reduces these risks by limiting what the assistant can claim and by making the response path visible to the people responsible for the service.
A practical deterministic assistant architecture
A useful architecture can be understood as five layers.
1. Intent
The system identifies what the person is trying to achieve. This can use keywords, a decision tree, a lightweight classifier or a language model with a constrained list of allowed intents.
2. Knowledge
Approved facts, policies, services and answers are stored in a structured form. Each important claim has an owner, source and review date.
3. Rules
Explicit logic controls decisions that can be defined. Rules can check eligibility, ask required follow-up questions, calculate a result or select the next step.
4. Response
The system returns an approved answer, a structured result or a generated explanation grounded in the controlled result.
5. Fallback
When the intent is unclear, the evidence is missing or the request is outside scope, the assistant explains the limit and directs the person to a safe next step.
This architecture can still feel conversational. The difference is that the conversation sits on top of a controlled service design.
Where deterministic assistants work well
This approach is particularly useful when questions repeat and the underlying answers can be maintained.
Strong use cases include:
- service and eligibility guidance;
- product selection with defined criteria;
- internal policy and process support;
- visitor information and structured directories;
- quotation or triage flows;
- onboarding and document collection;
- support assistants that must cite approved sources;
- knowledge tools for staff who need consistent answers.
The value is not limited to risk reduction. A controlled assistant can also be quicker to operate because simple answers do not need a large model call. Costs are easier to forecast, and the system can continue to work when a generative service is unavailable.
Where generative AI still adds value
Deterministic design is not an argument against generative AI. It is a way to use it deliberately.
A model can help classify a varied question, summarise a long document, explain a structured result in plain English or draft content for human review. It can search across approved information when the answer cannot be represented as a simple rule.
The important boundary is that the model should not silently become the source of truth. The system should preserve the approved facts and rules underneath the generated language.
A hybrid design is often strongest: logic handles what the business knows, retrieval supplies evidence, and generative AI handles the parts that truly benefit from flexible language or interpretation.
Frequently asked questions
Is a deterministic assistant the same as a decision tree?
Not necessarily. A decision tree can be one component, but a deterministic assistant can also use structured data, calculations, retrieval and constrained AI classification.
Can a deterministic assistant understand natural language?
Yes. Natural language can be mapped to a controlled intent or response path. The important outcomes remain governed by approved knowledge and rules.
When should a business use a general AI chatbot instead?
Use a more generative experience for open-ended exploration, drafting or analysis where variation is useful and a person remains responsible for checking the result.