The Cognitive Engine learns the everyday decisions your business makes, from past examples you have confirmed. Before it is trusted, it must pass a test on cases it has never seen. Then it decides instantly, the same way every time, and tells you why. Anything new goes to AI or to one of your people.
Facts from your system
Rule, in plain words
A reading in its normal range, from a sensor that is working properly, is OK.
Test
1,453 of 1,453 right
Cases it had never seen
Independent checks
6 of 6 agree
From Factory Floor, a simulated test business with made-up data. Measured in October 2026.
Nx·Studio designs your digital workers. Nx IDE helps your developers write software with AI. The Cognitive Engine makes routine decisions once it has proved it gets them right, so AI doesn't have to work them out again every time. All three are built by Nexgile.
Many AI tools ask an AI model about every decision, even ones they have seen many times. That is fine in a demo. In daily work it is slow, every question costs money, and the answer can change from one day to the next.
Each new case is checked against the rules the engine has proved. If one fits, the case is decided instantly, with no AI. If none fits, the engine never guesses: depending on the mode you choose, the case goes to AI within your budget, or to a person.
New cases
Proven rules
Checked first
Decided by
These numbers are from the animation, not from real tests. You'll find measured results lower down the page.
No business lets a new hire make decisions alone on day one. New staff learn from past cases and prove themselves first. Only then do they work on their own, and they still ask for help with anything unfamiliar. The Cognitive Engine works the same way.
Owns the systems and knows what really happened. The engine learns only from the outcomes your business has confirmed.
Makes suggestions while the engine is learning and, if you allow it, handles cases the engine hasn't learned yet. It never decides what counts as right.
Learns only from confirmed outcomes, and must pass a test before it may work on its own.
Confirm outcomes and settle cases where past answers disagree. In the strictest mode, they also get anything the engine hasn't proved yet.
The same five steps apply to every kind of decision. It takes over one kind at a time, and only once the evidence shows it is ready.
It studies past cases where your business has confirmed the right answer, from your systems or your reviewers. AI may make suggestions, but they never count as proof.
People at the factory confirm, case by case, whether each old sensor reading was normal or a problem.
Each rule is tested on past cases it didn't see 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.
On a simulated factory with made-up data, one rule got 1 case in 515 wrong, so it was not allowed to act. With more examples, it later got 1,453 of 1,453 right.
A proven rule answers in a fraction of a second, with no AI. The same facts always get the same answer. Six separate checks run the rule, and all six must agree, or the rule is withdrawn.
When a healthy sensor sends a normal reading, the engine marks it OK and shows which rule decided.
It never forces an unfamiliar case into a rule. Depending on the mode, it reuses your own past decision for the exact same facts, asks AI within your budget, or passes the case to a person. Where your past answers disagree, a person decides.
A sensor reading outside every range the engine has seen is sent to AI or to a person instead of being guessed.
Rules expire after 30 days, or sooner if you choose. A rule is withdrawn the moment a confirmed outcome proves it wrong, and must pass a new test before it acts again.
If the factory changes what “normal” means, the first confirmed case that disagrees stops the old rule. A replacement must pass its own test.
Much of what an AI tool decides each day is routine. The engine takes over a routine decision only once it has proved it gets it right. Anything else is answered from your own past decisions, by AI within your budget, or by your people.
Nine sensor readings from the shift, waiting to be decided.
Passed its test. Instant, no AI used.
Matches the facts to a decision you made before, or asks AI while staying within budget.
Sent to a person, or your past answers disagree.
Illustration based on Factory Floor, a simulated business with made-up data, where the engine has proved itself on routine sensor readings. Your mix will depend on which decisions your engine has proved and the mode you choose. What never changes: proven rules never ask AI, AI stays within your budget, and anything handed off reaches a person.
Proven routine decisions never use AI, so they add nothing to your AI bill. Any AI that is still needed stays within the budgets and optional daily limits you set, and you are warned before a limit is reached.
How our pricing worksThe same facts always get the same answer, even on days when an AI-only tool's answers change. Re-run a shift and every proven decision comes back exactly the same.
Every decision names the rule behind it in plain words, and six separate checks must agree on it. Every decision is kept in a record that can't be edited, and anything handed off goes to a person.
Read the Trust CenterYou choose how much the engine decides, for each application and each run. In Shadow mode it only practices: it decides alongside your current system and is scored, but changes nothing.
Your current system keeps deciding. The engine decides alongside it and its answers are only scored, so you see how it would have done before it changes anything.
Who decides, in order
The engine decides first. If it hasn't proved a case yet, it reuses your past decision for exactly the same situation. If your past decisions differ, a person decides. Only brand-new situations go to AI, within your budget.
Who decides, in order
Only proven rules decide. Everything else goes to your system's safe backup route, which is usually a person.
Who decides, in order
The Studio is where your team collects examples, trains and tests the engine, compares it with an AI-only tool, and approves it before it goes live. Everyone gets the role they need, from viewers and reviewers to trainers and administrators.
Each application shows its AI tools, their limits, what must never go wrong, and which decisions the engine has earned so far.
On our simulated businesses, one click compares the engine with a tool that asks AI every time: accuracy, decisions left to people, speed and AI use. On your own system, it compares the engine with your system's decisions.
Read each learned rule in plain words, the test it passed, and how each of the six checks answered.
Your people confirm cases from a simple list, including blind reviews that hide the suggested answer. The Studio tracks how well reviews are done.
AI drafts each decision's facts, limits and policy from your documents and systems, with a source for each. A person must approve every draft, and the sources are re-checked weekly by default.
A launch checklist for each application, sign-off by its owner, a security review, stress and recovery drills, and a hand-over package.
User roles, optional sign-in with your company login, optional two-person approval for setting changes, masked names and IDs, and an activity record that can't be edited or deleted.
AI budgets per run and optional daily spending limits per application, with alerts at 80% and 100%. A switch decides whether AI may see case facts at all.
Warnings for expiring or withdrawn rules, slow decisions, AI spend and low storage space, checked every 15 seconds.
We test the engine on twenty simulated businesses with made-up (synthetic) data, from insurance claims and payments to factories and energy markets. We compare it with an AI tool built the usual way, which asks an AI model about every decision. Here is what we measured.
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.
The engine works next to your existing AI tools and systems instead of replacing them. Your systems stay in charge of what actually happens.
Any AI tool can ask the engine for a decision through a standard web connection (an API), using its own access key.
Send in decisions your system has already made (a shadow feed). The engine decides alongside without acting, and is scored against yours. New situations become cases for your people to confirm.
AI assistants can consult the engine through a local MCP server, the open standard AI tools use to connect to other systems. It only advises and never acts. A Python toolkit for developers (SDK) is also available.
Comes ready to run on your own servers, packaged with Docker Compose. After an upgrade, every rule is checked again before it acts.
Careful with your systems. Before the engine does anything through your tools, it records what it is about to do. It never tries a failed action again on its own, and if the result is unclear, it stops and waits for your team.
Visit the Trust CenterEverything below is built and available. The extras stay off until you switch them on.
No, the two work side by side. The engine handles the routine decisions it has already got right in testing, and anything new goes to AI (within budget) or to your staff.
No. Only rules that passed their test may decide. Depending on the mode you choose, 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 choose. A rule is withdrawn as soon as a confirmed outcome proves it wrong. A replacement acts only after it passes a new test.
Yes. Every decision names the rule behind it in plain words. In the Studio you can also see the test that rule passed and how each of the six checks answered.
By default, yes: AI may see case facts when it is asked to help, with names and other identifying details from your data sources masked first. You can switch this off for each application or data source. AI also drafts from your documents, unless you turn drafting off.
Not so far. Every result on this page was measured on simulated businesses with made-up data. In Shadow mode, the engine decides alongside your own system without acting, so you can measure it there before it acts.
