Your people answer the same routine questions every day. The Cognitive Engine starts empty and learns those decisions from past cases your team confirmed. It must pass an exam on cases it has never seen before it may decide. Then it answers in milliseconds, the same way every time, and shows its rule. Anything new still goes to a person, or to AI if you allow.
What your system knows
Its rule, in plain words
If the reading is within its normal range and the sensor isn't stuck or dropping out, it's OK.
Exam result
1,453 of 1,453 right
All on cases it had never seen
Cross-check result
All 6 agree
From Factory Floor, a simulated test business with made-up data. Measured in October 2026.
Nx·Studio builds your digital workers. Nx IDE helps your developers code with AI. The Cognitive Engine handles routine decisions it has proven it can get right. Then no person or AI model has to work out the same answer again and again. Nexgile builds all three.
Most AI tools ask an AI model about every decision, even ones they've seen a thousand times. That works in a demo. Day to day, it's slow, every question adds to the bill, and the answer can change from one day to the next.
Each new case is checked against the rules the engine has already proven. If one fits, it's decided on the spot, with no AI model involved. If none fits, the engine won't guess. Depending on the mode you pick, the case goes to AI within your budget, or to one of your people.
Incoming cases
Its proven rules
Always tried first
Who decides
An illustration only: these counts come from the animation. For measured results, scroll further down the page.
Nobody hands a learner the car keys on day one. They practice with an instructor, pass a test, and only then drive alone, and even then they take extra care on roads they don't know. The Cognitive Engine works the same way.
Your systems record what really happened. The engine learns only from outcomes your business has confirmed as right.
Gives tips while the engine learns and, if you allow it, takes the wheel for cases the engine hasn't learned yet. It never decides what counts as right.
Starts with no knowledge of your business, learns only from confirmed outcomes, and must pass its test before it drives alone.
Confirm outcomes and settle cases where past answers disagree. In the strictest mode, they also handle anything the engine hasn't proven yet.
Each kind of decision follows the same path. The engine takes over one kind at a time, and only when the evidence shows it's ready.
It begins knowing nothing about your business. It learns from past cases whose right answer your systems or your reviewers have confirmed. An AI model may suggest answers, but a suggestion never counts as proof.
The factory marks which past sensor readings were truly fine and which needed someone to act.
Each rule is tested on past cases it never saw while learning. It may act only after getting at least 400 right with no mistakes. You can raise this pass mark, but never lower it.
In a simulated factory test with made-up data, one rule missed 1 case in 515, so it wasn't allowed to act. After more examples, it got 1,453 of 1,453 right.
A proven rule answers in a fraction of a second, with no AI model involved, and the same facts always get the same answer. Six separate checks run the rule, and if they ever disagree, the rule is withdrawn.
A normal reading from a working sensor is cleared as OK, and the decision shows which rule made the call.
It never forces an unfamiliar case into a rule. Depending on the mode, it reuses your past decision for exactly the same facts, asks AI within your budget, or hands the case to a person. Where past answers disagree, a person decides.
A reading unlike anything it has seen goes to AI or to a person. It doesn't guess.
Rules expire after 30 days, or sooner if you choose. As soon as a confirmed outcome shows a rule is wrong, the rule is withdrawn and must pass a new exam before it acts again.
If the factory redefines “normal”, the first confirmed case that disagrees switches the old rule off. The new rule has to pass its own exam.
A lot of what gets decided each day is routine. The engine takes a routine decision off your team only after proving it gets it right. Everything else goes to your own past decisions, to AI within your budget, or to your people.
Nine sensor readings from this shift are waiting for a decision.
Passed its exam. Instant, and no AI needed.
Your past decision for the same facts, or else AI within your budget.
Handed over on purpose, or past answers disagree.
An illustration based on Factory Floor, a simulated business with made-up data, where the engine has proven itself on routine sensor readings. Your own mix will depend on which decisions your engine has proven and the mode you pick. Some things never change: proven rules don't ask AI, AI stays inside your budget, and anything handed over reaches a person.
Routine decisions the engine has proven use no AI, so they add nothing to your AI bill. Whatever AI is still needed stays inside the budgets and optional daily limits you set, with a warning before you reach a limit.
See how pricing worksSame facts, same answer, even on days when an AI-only tool's answers wander. Replay a shift and every proven decision comes back exactly as it was.
Each decision names its rule in plain words, and six separate checks have to agree on it. Every decision is kept in a record that can't be edited, and anything handed over goes to a person.
Visit the Trust CenterYou decide how much the engine handles, for each application and each run. In Shadow mode it only rehearses: it answers next to your current system and gets scored, but changes nothing.
Nothing changes for your team yet. Your current system keeps deciding while the engine answers alongside and gets scored, so you see how it would have done.
Who gets the decision, in order
The engine goes first. For a case it hasn't proven, it reuses your past decision for exactly the same situation. Where your past decisions differ, a person decides. Only truly new situations go to AI, within your budget.
Who gets the decision, in order
Only proven rules decide, and AI is never asked. Everything else follows the backup route your system already has, usually a person.
Who gets the decision, in order
The Studio is where your team collects examples, trains and tests the engine, compares it with an AI-only tool, and signs it off before it goes live. Everyone gets the access they need, from viewer and reviewer to trainer and administrator.
For each application: its AI tools and their limits, what must never go wrong, and which decisions the engine has earned so far.
On our simulated businesses, one click puts the engine next to a tool that asks AI every time: accuracy, decisions left to people, speed and AI use. On your system, it compares the engine with the decisions your system actually made.
Read each learned rule in plain words, see the exam it passed, and see what each of the six checks said.
Your people confirm cases from a plain list, including blind reviews that hide the suggested answer. The Studio keeps track of how good each review is.
AI drafts each decision's facts, limits and policy from your documents and systems, with a source for each one. A person has to approve every draft, and sources are re-checked every week by default.
Each application gets a launch checklist, its owner's sign-off, a security review, stress and recovery drills, and a hand-over pack.
Roles for each user and optional sign-in with your company login. Optional two-person approval for setting changes, hidden names and IDs, and an activity record nobody can edit or delete.
A budget for AI on every run and optional daily limits per application, with alerts at 80% and 100%. One switch decides whether AI may see case facts at all.
Warnings when rules expire or are withdrawn, when decisions slow down, and when AI spend climbs or storage runs low, checked every 15 seconds.
We run the engine on twenty simulated businesses with made-up (synthetic) data, from insurance claims and payments to factories and energy markets. Each time, it's compared with an AI tool built the usual way, which asks an AI model about every decision. These are the results.
Field test on five simulated businesses with made-up data, October 6, 2026.
Every decision the engine made itself, on all 20 simulated businesses (made-up data, 1,000 events each, October 9, 2026). The engine currently decides cases in 14 of them.
Typical decision: 17–31 ms (thousandths of a second) for the engine vs 3.3–4.2 seconds for an AI-only tool. Five simulated businesses, made-up data, October 8, 2026.
Same simulated runs, made-up data: per 1,000 decisions, the AI-only tool made 329–1,519 AI requests; Cognitive + AI mode made 0–14. Most AI-only figures are scaled up from runs capped at 300 requests.
On the same simulated factory decision, an AI-only tool using the same AI model and the same instructions scored 97% one day and 20.5% the next. The engine's proven rule decided the same 865 readings on both days, and got none of them wrong.
All figures were measured by Nexgile in October 2026 on simulated businesses with made-up (synthetic) data. “Right” means matching each simulated business's own policy. They show the method, not a customer's results, and do not by themselves prove the engine is ready for live use.
You don't have to replace anything. The engine runs alongside your current AI tools and systems, and your systems stay in charge of what actually happens.
Any AI tool can ask for a decision over a standard web connection (an API), using its own access key.
Feed in decisions your system has already made (a shadow feed). The engine answers alongside, without acting, and is scored against yours. Anything new becomes a case for your people to confirm.
AI assistants can consult the engine through a local MCP server (MCP is the open standard AI tools use to reach other systems). Here it only advises and never acts. Developers also get a Python toolkit (SDK).
It comes packaged to run on your own servers (with Docker Compose). After any upgrade, every rule has to pass its checks again before it acts.
Careful by design. Before the engine acts through your tools, it writes down what it's about to do. It never retries a failed action by itself, and if a result is unclear, it stops and waits for your team.
Visit the Trust CenterEverything here is built and available today. The extras stay switched off until you turn them on.
No. It only takes over the routine decisions it has proven it gets right. New cases still go to an AI model within your budget, or to your people.
No. Only rules that passed their exam are allowed to decide. Depending on the mode you pick, everything else is answered from your own past decisions, by AI within your budget, or by a person.
Rules expire after 30 days, or sooner if you prefer. As soon as a confirmed outcome shows a rule is wrong, it's withdrawn, and its replacement acts only after passing a new exam.
Yes. Each decision names its rule in plain words. In the Studio you can also see the exam that rule passed and what each of the six checks said.
By default, yes. When AI is asked to help, it may see the case facts, with names and other identifying details from your data sources hidden first. You can switch this off for each application or data source. AI also drafts from your documents unless you turn drafting off.
Not yet. Every result on this page comes from simulated businesses with made-up data. In Shadow mode the engine answers alongside your own system without acting, so you can measure it on your own work first.
