AI that earns the right to decide.
The Cognitive Engine learns your routine decisions from outcomes your business confirms. It acts only after passing an exam on cases it has never seen — then makes those qualified decisions using local rules and small models, without calling a large language model (LLM). Unfamiliar cases go for human review, with an optional LLM fallback if you enable it.
What the engine has learned
Sample dataThe learned rule, in plain words
A reading inside its normal band, from a sensor that isn't stuck or dropping out, is normal — no alert.
Illustration of the Cognitive Studio's overview of one business.
Most AI agents ask an AI model about everything.
Yet most business decisions are routine. They follow rules the business already applies — often rules nobody ever wrote down.
Paying for the routine
An AI agent built the usual way asks an AI model about every decision — even the thousandth identical one. Each question costs money and takes seconds.
Answers that drift
The same model, asked the same question on another day, can answer differently. In our own tests, one model's accuracy on the same task moved sharply from one day to the next.
Mistakes nobody sees
An AI model sounds just as sure when it's wrong. Without checking answers against real outcomes, nobody knows which ones to trust.
The Cognitive Engine takes over the routine decisions — one kind at a time, and only after it proves it gets them right.
Think of a new trainee.
A good trainee watches how things are done, gets checked, and only works alone once they've proven themselves. The engine works the same way.
Your business is the boss
It sets the job, owns the tools, and is the only one who can say what really happened.
An AI model is the tutor
It suggests how to handle cases and proposes rules — but it never decides what counts as right.
The engine is the trainee
It learns only from outcomes your business confirmed, and must pass an exam before it works alone.
The license expires
Its license to act runs out after 30 days by default — or sooner if the job, the tools, its code or its lessons change. Then it takes the exam again.
From empty to trusted, one decision at a time.
Six steps, the same for every kind of decision the engine takes on.
Starts empty
It comes with no ready-made knowledge and no built-in business rules. It knows nothing about your business until you teach it.
Learns from confirmed outcomes
An AI tutor may suggest answers, but only outcomes your systems or your people confirm count as lessons. From those, the engine builds small rules and models that run on your own servers.
Passes an exam before it acts
A kind of decision goes live only when the engine gets at least 400 new cases right, with no mistakes. One miss, and it keeps learning instead of acting.
Decides in a fraction of a second
Once qualified, it makes that decision using local rules and small models, without calling a large language model (LLM), and with the same answer for the same facts. Six independent checks confirm each answer before it acts.
Knows what it doesn't know
Anything new, unusual or unproven gets “Review” instead of a guess, and a person decides. If you allow it, the AI model can answer those cases instead.
Stays qualified — or steps back
Its license runs out after 30 days by default, or sooner if its code, data or your rules change. It retakes the exam on what it has already learned before it acts again.
Watch the engine work a shift.
The engine decides only what it has learned and proven. Here it handles routine sensor readings locally, without calling an LLM. Everything else goes to a person or, when enabled, the LLM fallback.
Nine decisions waiting from the plant's agents.
The engine decides
Learned, and passed its exam. No AI needed.
The AI model answers
Not learned yet — the AI model steps in, if you allow it.
A person decides
Approvals, and anything the engine isn't sure about.
Illustrative scenario on a simulated factory floor; the machine names are made up. How much the engine decides depends on what it has learned and proven.
A decision the engine has learned doesn't use AI at all, so it adds nothing to your AI bill. The AI model helps teach the engine, and answers decisions it hasn't learned yet — only where you allow it.
Learned decisions take a fraction of a second, and the same facts get the same answer every time. When the engine isn't sure, it says so and hands the case to a person instead of guessing.
Every learned rule reads in plain words, next to the exam it passed. Nothing acts until it is right on cases it has never seen, and its license to act runs out whenever its code, data or rules change.
One customer message, start to finish.
A shop wants every customer message turned into the right action. Here is how the engine learns that decision — and then makes it.
- 01
The job
A shop allows five actions: reserve, pack, send by the standard carrier, send by the backup carrier, and cancel. The goal is simple: dispatch the parcel — or cancel it, if that's what the customer asked.
- 02
Teach
An AI tutor sorts 1,200 made-up customer messages, then checks its own work a second time.
- 03
Verify
The shop's own records supply the true answer for every message. Only verified answers become lessons — which matters: in this example, the tutor's own answers were right only about 85% of the time.
- 04
Learn
The engine learns to read the messages from 600 examples, and uses 180 more to learn how sure it should be.
- 05
Exam
It must get all 420 new messages right — both what the customer meant and the action to take. It scored 420 out of 420, so this decision went live.
- 06
Act
“Please dispatch my parcel today.” Stock is in and the carrier is free, so the engine plans: reserve, pack, send standard. It checks the live situation before every step — using its local rules and models without calling an LLM. Sent twice by mistake? It recognizes the repeat instead of shipping twice.
- 07
Adapt
The standard carrier goes down mid-order. The engine notices before shipping, changes its plan, and sends with the backup carrier — still without calling an LLM.
- 08
Hand off
A request written in a language it was never taught gets “Review”, not a guess.
Message
“Please dispatch my parcel today.”
Plan
Carrier down before shipping
A workspace to teach, test and trust the engine.
The Cognitive Studio is the web app around the engine. Pick a business, see what its AI agents do, watch the engine take over one decision at a time — and compare it with an agent that asks an AI model about everything.
A library of businesses
Twenty simulated businesses, each run by its own team of AI agents — a factory, a farm, a power market, an insurer, a card issuer, a hotel and more.
Each business, explained
What the business does, what must never go wrong, each agent's job and limits — and every decision marked “Engine decides” or “Not learned yet”.
How it earned each decision
Every learned rule in plain words, next to the exam it passed. The cautious part is visible too: the rules still waiting for enough exam cases.
Side-by-side comparison
Run the same events three ways: the engine alone, the engine with an LLM fallback, or the LLM alone. Then compare accuracy, speed, LLM calls, and how often a person was needed.
Transparency, decision by decision
For any learned decision: its rules, each rule's own exam, and how the six checks agreed on every recent decision.
Human review, built in
Question cards for reviewers, each option shown with its consequence. Reviewers answer exam questions without seeing the engine's answer, and two must agree.
Readiness and sign-off
A readiness report for each business that the owner signs: what can act, the exam it passed, its evidence and when its license runs out.
Some of the businesses in the library
Simulated businesses with made-up data, built for testing.
What the tests show.
Each figure comes from the engine's test bench and says exactly what it measured.
learned decisions matched each business's own policy
Five simulated businesses, one learned kind of decision each, 1,000 events per run · field test, Oct 6, 2026
typical time per decision: the engine vs an AI model on the same task
Same test, typical times; the engine's time includes its six checks
LLM calls for a decision the engine has learned and qualified to handle
By design — and recorded in every example run
unseen cases it must get all right before a decision goes live
The default rule — you can make it stricter, never looser
requests no allowed action could handle went to a person — none was answered wrongly
Simulated order-dispatch test
automated tests passing
The engine's automated tests, Oct 7, 2026
What it does today.
Included as standard
- Learning from outcomes your business confirms — in data records and short messages
- The exam, and a license to act that runs out or can be withdrawn
- Qualified decisions using local rules and models without LLM calls, and “Review” whenever it isn't sure
- Six independent checks on every decision
- Step-by-step action plans that change when the situation changes, and never repeat an action already taken
- Hand-offs: cases your rules can't cover go to a person
- A full record of evidence, decisions and actions
- The Cognitive Studio, plus ways for your software and other AI tools to ask it questions — it never acts for them
Built in, off until you switch it on
- Letting the AI model answer cases the engine hasn't learned
- An extra grammar check for phrasings it has never seen
- Ranking alternative plans, and learning your preferences
- Extra checks for unfamiliar or uncertain cases
- A separate exam for each rule
- Longer plans, up to 20 steps
Part of the Nexgile platform.
The Cognitive Engine runs on its own, with its own Studio. Your software can connect to it directly, and other AI tools can ask it questions and get suggested plans — it never takes action for them.
Questions people ask first.
Does it learn from the AI's guesses?
No. The AI tutor's suggestions are never used as lessons. Only outcomes confirmed by your systems or your people count. That is the whole point: the tutor can be wrong, and in the engine's own tests it often was.
What happens when it isn't sure?
It answers “Review” and a person decides. It does not guess. If you choose, the AI model can answer those cases instead — and the engine keeps learning from the confirmed outcomes.
Where does it run, and what leaves my servers?
The engine, everything it learns and its records run on your own servers — ordinary ones, with no special AI chips needed. It contacts the AI model's provider only while you teach it, and at decision time only if you switch that on. A decision the engine has learned is made entirely on your servers.
Which AI model does it use?
An AI tutor (a large language model, or LLM) helps during learning. Only outcomes your business confirms become lessons. Once qualified, the engine handles supported decisions using rules and small models running on your servers, without calling an LLM. Unfamiliar cases go for human review, with an optional LLM fallback if you enable it.
How do I know it decides correctly?
A kind of decision goes live only after the engine gets at least 400 new cases right, with no mistakes. Six independent checks confirm each decision before it acts, and every decision is recorded with its evidence. Its license to act also runs out after 30 days, or sooner if its job, tools, code or lessons change. Then it must pass the exam again.
How does it handle customer messages?
It learns to read short messages from confirmed examples. Teach it phrases with “don't”, like “don't ship”, as examples of their own, or switch on the optional grammar check. Anything unfamiliar goes to “Review” for a person.
Bring one decision you make every day.
We'll show the engine learning a decision — and talk through how it applies to yours.