How it works

Build once, then make specific tools everywhere they're needed.

Decidient first builds a production-line for a whole family of decisions. The production-line then makes a separate decision-support tool for each place, institution or other unit the decision spans, such as a university, a farming zone or a country, and each tool gives every person the decision support that fits their own situation.

Three levels

Built once

A production-line

The design for a family of decisions: what to research, which facts matter, what to ask people and how to rank the choices.

One per family of decisions

Made many times

Decision-support tools

Each order to the production-line makes one tool, researched and built for one place or institution, with that place's own facts.

Hundreds or thousands per production-line

Used by everyone

Each person's decision

Each person answers a few questions in the tool and gets their own recommendation, with the reasons behind it.

Any number of people per tool

Two examples

Choosing which crops to plant

The production-lineBuilt to help smallholder farmers in the developing world choose the best crops to plant. Choosing crops is the family of decisions.

The tools it makesOne for each farming zone, each researched for that zone's soil, rainfall and markets, in the local language. For example, in the Indian state of Odisha:

a crop advisor for the East and South Eastern Coastal Plain;
another for the North Western Plateau;
and so on, zone by zone, state by state and country by country.

The peopleEach farmer answers questions about their land, water and plans, and gets crops ranked for their own farm.

Choosing a university residence

The production-lineBuilt to help students choose the university residence that suits them. Choosing a residence is the family of decisions.

The tools it makesOne for each university, each researched from that university's own residence information. For example:

a residence finder for Queen's University;
another for McGill University;
and so on, one for every university that has residences.

The peopleEach student answers questions about what matters to them, such as cost, room type and quiet, and gets the residences ranked for them.

Because the production-line is built once, each new tool costs a fraction of the first. That is how one production-line can support thousands.

The steps

Once, for the family of decisions

Describe the family of decisionsWrite one paragraph: who is deciding, what they are choosing between, and what matters to them. Decidient asks about anything it cannot work out, and you approve its understanding before work begins.
Build the production-lineDecidient designs the steps every tool will need, from finding sources to writing the report, and tests them on a few real places before it is used. It lists anything it cannot yet do.

Once for each tool

Research the placeFor one place or institution, Decidient finds authoritative sources and records each fact with the exact words, the page and the date. Code checks every fact against its source. Unknowns stay unknown.
Build the toolIt turns those facts into the tool: the rules that rule choices out, the things people weigh, the questions, the ranking and the report.

For example

Once a production-line is built, you can order tools in bulk. You could ask the crop production-line to "build a crop advisor for every agro-climatic zone in India, Thailand and Malaysia." Or you could ask the university residence production-line to "build a residence finder for every major university in Canada that is not a commuter school."

For every person

DecideEach person gives their details, answers a few questions, and gets the decision support they need: software ranks the choices from their answers, explains the reasons, and says what would change the answer.

Inside a production-line

Workstations, machines and checks.

A production-line is a sequence of workstations. Each workstation does one job, such as finding sources or building the questionnaire, and every workstation is built the same way. That sameness is what lets one design make thousands of tools and lets every step be inspected.

The orderWhat this tool must be: the place or institution, its language, who it is for.

Earlier workThe checked results of the workstations before this one.

InstructionsThe written instructions each AI step follows, stored as data and versioned.

Reference dataCatalogues and datasets the step is allowed to read.

RulesWhat each step may read, which AI models it may use, what it may spend and who must approve.

One workstation

AI stepThe instructions are filled in with this order's inputs and sent to an approved AI model.

Contract checkSoftware checks the reply against a written output contract: every required item present, in the right form.

RedoOnly the items that failed are asked again, with the reasons, up to twice. The better result is kept.

Code stepTested software does what software should: calculations, rankings, joins, formats.

Quality reviewThe whole workstation's output is reviewed against written criteria: first by code, then by an independent AI auditor where judgement is needed.

Governance checkEvery step is checked against the rules, and the verdict is recorded and scored apart from quality.

OutputThe workstation's checked result, passed to the next workstation only when the supervisor allows it.

RecordEverything the workstation did, exactly: each instruction as sent, each reply, each check, its cost and time.

Order supervisorWatches each order as it runs and decides: continue, run again, hold or stop.

Learning across ordersAdjustments that work on one order are tested on others before they are adopted for later orders, never mid-order.

The stages in detail

What happens at each stage.

Understand the request

Once per family of decisions

An AI step reads the one-paragraph description and answers a fixed set of design questions about it: who decides, what they choose between, what one tool covers (one university, one farming zone), what the result must show, and how people will use it.

Every answer is typed, with a confidence level and a reason, and checked by code against the question's allowed form.
The description is read more than once; answers that differ between readings are listed for the person to settle.
A short summary names the main assumptions. The person approves it, or corrects it in a chat and the step runs again.

Compose the production-line

Once per family

Decidient turns the approved request into the capabilities the tools must have, then chooses workstations from its library that provide them. The order of work follows from what each workstation needs from the others.

Anything the request needs that no workstation provides is listed as a gap, marked blocking or not.
Every choice is recorded with the part of the request it serves.

Test before release

Once per family

The production-line is tried on a small design sample of real places, then judged on a separate set it has never seen. It is released only if the unseen set passes, so a design cannot be tuned to its own test cases.

Find sources

For each tool

For one place or institution, Decidient plans its research, searches for authoritative sources, and ranks them: official sources first, then recognised secondary sources.

Each page read is stored as a dated copy with a content hash, so every fact can be traced to exactly what was read.
Every search is logged, including searches that found nothing.
Sources can be web pages, documents, datasets, maps or measurements; each kind has its own reader and its own way of being checked.

Extract and check facts

For each tool

AI steps extract facts from the stored sources. Each fact is stored as a typed value, such as an amount in a currency, a distance to a named place, a date or a level on a scale, together with the exact words it rests on.

Code checks every fact against its stored source: the quoted words must be on the page, in order, and the value must be inside the quote. Facts that fail are dropped.
A gap is a gap: what a source does not say is recorded as unknown with its reason, searched for again, and never counted against an option.
Contradictions are paired and kept. Plausibility checks catch values that cannot be right, such as a whole university's total given as one building's size.
Samples of what was set aside are checked to measure what the research may have missed.

Build the decision model

For each tool

The checked facts become a decision model in the international standard for decision logic (DMN): the rules that rule options out, the characteristics people weigh, and how each characteristic is scored.

The model can be exported and run outside Decidient, so its logic can be inspected by anyone.
Each rule and score points back to the facts it rests on.

Build and review the tool

For each tool

From the model, Decidient writes the questions, the pages, the result explanations and the report. Every text is checked against the facts and against the house style.

Before release, the tool is reviewed against criteria drawn from the international standard for decision aids: balanced presentation, sources shown, an upside and a downside for every option, no pressure to choose.
Test personas with known needs are run through the tool; each must get the result its needs imply.
A governance record for the tool is assembled by code, signed, and kept.

Recommend

For every person

When a person uses the tool, no AI model decides anything. Software applies the person's must-haves, weighs the rest by the priorities their answers set, and ranks the options.

The ranking is tested against changes in the person's priorities; close calls and the answer that would change the result are shown.
The same answers always give the same result, and the record of how it was reached is kept.

How a recommendation is made

Must-haves first

Options that break a person's must-haves are set aside first, with the reason shown. A preference never overrules a must-have.

Then what matters most

People choose between pairs of real options ("this or that"). Their choices set how much each thing counts for them.

Close calls are named

When two options are nearly tied, the tool says so, rather than presenting a narrow win as a clear one.

What could change the answer

The tool shows which answer, if changed, would move a different option to the top.

AI answers and decisions

Why a typical AI answer is not enough for a decision.

Most AI tools that work from documents use a method called retrieval-augmented generation, or RAG. When you ask a question, the tool finds a few passages in its documents and an AI model writes an answer from them. That works well for questions. When someone will act on the answer, six things can go wrong.

Decidient also works from sources, so it belongs to the same family. The difference is when and how. It does its research ahead of time, checks every fact, and builds a decision model. When a person uses the tool, ordinary software works out the answer from that model.

1. The source may not say what the answer says

A typical tool cites a page but does not check that the page says it. Decidient keeps the exact words each fact rests on, and software checks that those words are on the page and that the value is in them. A fact that fails is dropped.

2. Silence is taken as no

When a page says nothing about something, a typical answer guesses or leaves it out, and an option can lose because its page was silent. Decidient records "unknown" with the reason, searches again, and never counts unknown against an option.

3. Disagreements are blended away

Asked to summarise, a model turns two different numbers into one smooth sentence, and the reader never knows there was a disagreement. Decidient keeps both values and shows them when it matters.

4. Facts stay as sentences

A fact stored as a sentence is re-read by a model every time, and each reading can differ. Decidient stores each fact as a typed value, such as an amount in a currency, a distance or a date, so software can compare and calculate with it exactly.

5. The AI makes the call

In a typical tool the model decides the ranking, and studies show such rankings shift with small changes in wording. Decidient uses AI to research and to write; software ranks the options from each person's own answers.

6. No record is left

A typical answer leaves a log for the operator and nothing for the person who relied on it. Decidient keeps a signed record of the sources, facts, checks, AI models and approvals behind every recommendation.

In short

A typical AI answer	A Decidient recommendation
Cites some sources, sometimes	Every fact is tied to its source, and software checks it
Fills gaps with a guess	Says plainly what is unknown
Blends disagreements into one answer	Shows both values when it matters
May give a different answer each time	The same answers always give the same result
The AI writes the conclusion	Software computes the ranking from the person's answers
Leaves no record	Keeps a signed record of how the result was made

What we trade for it

Researching every source in full and checking every fact takes more work up front than pulling a few passages at the moment of a question. Each tool answers the decision it was built for, not any question at all. We make that trade because people act on these recommendations.

Under the hood

Built for organisations and for audit.

Decision logic in a standard

Decision models are kept in DMN, the international standard for decision logic, and can be exported and run elsewhere.

Many AI models, one gateway

AI calls go through one gateway to approved models from several providers, with limits per step on model, spend and retries.

Every AI call recorded

Each instruction exactly as sent, each reply, the model and its settings, the cost and the time are stored for every call.

Instructions as data

Instructions, contracts, criteria and rules are stored as versioned data, so a change is reviewed, tested and released, never patched into code.

Step-through debugging

A development run can be paused before any step, stepped forward or back, and inspected part by part, from the page, an API or an AI assistant.

API and AI-assistant access

Production-lines can be designed and orders run through a versioned API or through an AI assistant connected over the Model Context Protocol.

Made for concurrent work

Many organisations, many production-lines and many people at once. Each organisation sees only its own data; every write is checked so two people cannot overwrite each other.

Personal information stays close

Personal information goes only to AI models hosted in Canada or the United States, and only when the work needs it.

Prompt caching and cost limits

Fixed instructions are sent first so providers can cache them, and every order has a hard spending limit.

One design, many local versions

A design is built once and used many times. A crop advisor, for example, gets a separate version for each farming region, with that region's sources and in the local language. Each version is checked the same way.