Deterministic First
Use AI to help build a tool. Let the tool follow tested rules for repeated work.
Team guide · 6 October 2026, ladder added 7 October 2026 · These examples explain ideas. They do not describe systems already in use.
The main idea: if we know the correct rule for a task, we should not need to ask AI to decide the answer every time.
We want reusable tools that do routine work correctly, at a lower cost. We keep AI or a person involved when the work needs judgment.
1. Three ways to do the work
AI is software that can interpret information, write answers, and suggest decisions. It can also make mistakes.
These three names tell us what decides the result.
Deterministic: follow a fixed rule
“Deterministic” means agreed, tested rules decide the result for the cases the tool can handle. The tool does not ask AI each time.
Think of a spreadsheet adding a list of numbers. It follows a formula. It does not ask AI what the total should be.
Hybrid: rules run the process; AI helps at one step
“Hybrid” means a mix. Rules control the work, and AI has a specific job within it. We define when AI helps, what it may do, and how its answer will be checked.
AI-driven: AI makes the main decision
“AI-driven” means AI supplies the main answer or judgment, or decides what normally happens next. For example, AI may recommend a content plan after reading a client's website.
This can be useful when a task needs interpretation or new ideas.
2. One SEO task, three ways
Task: check whether a business listing shows the opening hours the client approved.
Deterministic
The listing and the approved record both have separate fields for days and times. The tool compares them using agreed rules. It returns match, mismatch, or needs review. AI does not decide the answer.
Hybrid
The same rules compare the hours. If a listing uses a sentence such as “Monday to Friday, nine to five,” AI helps put those words into day and time fields. A person confirms any uncertain meaning before the rules compare the hours. AI has a limited role.
AI-driven
We send both records to AI and ask, “Do these opening hours agree?” AI makes the main decision each time.
Look at the final result. A tool may have many steps that follow rules. If AI still makes the main decision, those extra steps do not make the whole task deterministic.
A check and a fix can be different. Rules may find a mismatch. A person may need to confirm which hours are correct before a tool updates them.
Automatic work and approval are separate questions. An AI-driven job can start automatically every morning. A deterministic tool can still need a manager's approval before making changes.
3. Why this helps our team
- Less dependence on the most capable AI tools. We can save them for difficult work.
- More consistent results. The same information and rules lead to the same decision. Accuracy still depends on correct information and correct rules.
- Lower costs. We reduce repeated AI use and repeated manual work.
- Tools we can keep using. We can use our Codex, Claude, and ChatGPT subscriptions to help build reusable tools, sometimes called processors. A processor takes information and performs a specific job.
4. Building a tool and running a tool
Suppose AI helps us write and test the opening-hours checker. That is the building stage. Using AI during this stage uses our subscription allowance.
Next week, we give the finished checker another list of businesses. It follows its rules. That is a run.
A normal run that does not ask AI for anything does not use AI allowance. AI helping to build the tool does not mean AI must help every time it runs.
If the checker sends each listing to AI for a decision, it still uses AI during every run. Count AI used through other connected services too.
Reusable tools still need a computer or hosting service to run. Those services may have costs and limits. Tools also need maintenance when websites, information, or business rules change.
5. When the tool should stop
We define which cases the tool can handle. We do not need to predict every possible future situation.
For example, our checker may handle regular weekly hours. Special holiday hours may need review until we agree and test a rule for them.
If information is missing, conflicting, or outside the agreed cases, the tool should stop or flag it for review. It should not guess.
Correct formatting does not prove an AI answer is true. An opening time can look correct but belong to a different business.
Rules can also produce wrong answers repeatedly if the approved record is outdated. Keep the source information current. Keep AI or human judgment where rules cannot reliably produce the result we need.
6. What the team should do
Choose one repeated task and work through these questions:
- What is a correct result? Name the approved source we should use, such as the client's confirmed business details.
- Who decides today? Is it a person, AI, or a fixed rule? If we do not know, say so.
- Can we make the task simpler? Write a clear rule. Consider a better form: separate day and time fields are easier to check than a message.
- What should the tool stop on? List missing information and unusual cases. The tool stops or flags each one. Name the person who turns each flagged case into a new rule, so the tool handles it next time.
- Does it work on approved examples? Test ordinary cases and exceptions against confirmed answers. An old AI answer is not proof.
- Can it go live? Get the required approval before changing live records or processes.
- Did it help? Track correct results, mistakes, cases needing review, costs, and how often the tool still asks AI.
Keep these stages clear: how we work today, a proposed change, a tested change, and a change in live use. An idea is not yet a working tool.
7. Moving a task up the ladder
Section 6 tells us where a task stands today. This section shows how to move it closer to fixed rules, one step at a time.
The goal: no person needed in normal runs. Ben, 7 October 2026: “These deterministic systems have to work without human involvement. The only time a human would be involved is if we bump up against an edge case and we need to fix that and add another rule.”
So a person does not check or approve every case. The tool handles every case it was built for. When it flags a case it cannot handle, a person fixes or adds a rule, and the tool handles that kind of case from then on. The same kind of case should never need a person twice.
Some steps stay with a person because a company rule says so. For example, a person always sends messages to clients. We record those steps as required by policy. We do not count them as a gap. Every other routine step a person does is a gap to close.
Five things we measure
There is no single score. A task moves up by improving these five things, step by step.
| What we measure | The question | Moving up means |
|---|---|---|
| Foundation | Is a correct result written down? Do we have an approved source and approved examples? Does someone own each unclear policy? | Fill each gap first. Without these, nothing can be tested. |
| What decides | Does a rule, AI, or a person make the main decision? Or do we not know yet? | The main decision moves toward tested rules. |
| Rule coverage | Out of all real cases in a set period, how many did rules finish with no AI and no person deciding? | That share grows. A case the tool safely stopped on is handled correctly, but it does not count as finished. |
| Person's role in a run | Does a person decide, approve, or do the work by hand during normal runs? | Their role becomes none, except steps a company rule requires. Their job becomes fixing rules for flagged cases. |
| Proof | Is the change only an idea, tested on past cases, tested quietly alongside live work, or proven in live use? | We only claim what has been proven. |
Replacing a person with AI is not a step up. It can make the work faster, but it does not make the result follow fixed rules. Where we can, we go straight from a person deciding to a written rule.
Where to start
Nothing is set up yet
Write down the correct result and its approved source. Collect real cases with confirmed answers. Ask the owner to decide any unclear policy. Then try rules first. An unclear policy is a question for its owner, not a reason to use AI.
A person decides today
Record each decision in a simple form: what the person looked at, what they decided, and why. Find the decisions that repeat. Write those as a rule table. The tool stops on anything the table does not cover yet, and each stopped case becomes a new rule. That list of stopped cases should shrink every week.
AI decides today
Read what the AI decided on real cases, and why. Split them into “a rule could decide this” and “this really needs judgment”. Move collecting, filtering, and sorting into rules. Shrink what is left to the smallest possible AI job, with clear limits.
Rules run it, and AI helps at one step
Count how often the AI step is used, and why. Some “unusual” cases turn out to be ordinary ones. Turn those into rules using the steps below.
Rules were tried, and they keep failing
Before changing the rules, check two things. Is the tool missing the information it needs? Has nobody decided the policy? Rewording a rule fixes neither.
We do not know how it works today
Say so. List the evidence we need to find out. Do not guess.
Turning repeated answers into rules
When AI or a person keeps giving the same answer to the same kind of case, that answer can become a rule. It must earn its place first:
- Collect. Record each decision, what it was based on, and why it was made.
- Spot the pattern. Group the repeats into a proposed rule.
- Test it fairly. Check it on confirmed examples that were not used to write it.
- Replay it. Run it on past cases and compare with confirmed answers.
- Run it quietly. Let it run beside live work without changing anything.
- Get approval. The owner approves it. A rule that can make a big or permanent change always needs a person's approval.
- Switch it on. Make sure it can be switched off quickly if it goes wrong.
This never happens automatically. One AI answer never becomes a rule by itself. Nothing changes the rules on its own.
When can we call it deterministic?
Only when all four are true:
- We have looked at how it actually works, including everything it depends on.
- We know where every piece of information comes from.
- It has run correctly with AI switched off.
- Its answers match confirmed correct answers, not just the old AI's answers.
Giving AI a smaller job is real progress. But if AI still makes a required decision, the task is still hybrid. We report the progress as more cases decided by rules.
The opening-hours checker, moving up
- Today a person compares each listing with the approved hours by eye.
- We record their decisions for two weeks. Most cases are simple weekly hours.
- We write a rule table for weekly hours. The tool stops on anything else, such as holiday hours.
- Holiday hours keep coming up. The owner agrees a holiday rule. We test it on confirmed past cases, then run it quietly, then switch it on.
- The tool now handles weekly and holiday hours with no person and no AI. A person only steps in when the tool flags something new.
When to stop
Stop moving a task up when a test fails, not when it gets hard. For example, there may be no reliable way to check the answers, or the cost may outweigh the benefit. Write down why, and what would change the decision, such as new information, an owner's decision, or a cheaper method.