Every routine decision your AI agents send to a model is a recurring cost, a delay and an exposure. The Cognitive Engine learns from outcomes your systems or reviewers confirm, and must pass an exam on cases it has never seen. Only then does it decide alone, in milliseconds and at no AI cost, giving a reason your auditors can read. Anything unfamiliar stays with your people, or goes to AI if you permit it.
Facts your systems supply
The rule, plainly stated
A reading within its normal band, from a sensor that isn't stuck or dropping out, is OK.
Exam
1,453 of 1,453 right
Cases it had never seen
Reasoning checks
6 of 6 agree
From the Factory Floor simulated business. Measured in October 2026 on synthetic data.
Nx·Studio gives the enterprise digital workers it owns outright. Nx IDE extends engineering capacity under governance. The Cognitive Engine, a separate product, removes the recurring AI cost of routine, high-volume decisions, but only for the decisions it has proved it gets right. Nxgyle builds all three.
Many AI agents route every decision through a large model, including cases they have settled many times before. In a pilot, that cost goes unnoticed. At enterprise volume it becomes a standing bill, a delay on every case and an answer that can shift from one day to the next.
Each case meets the rules the engine has already proved. A match is settled in milliseconds, at no AI cost, with the same answer every time for the same facts. Anything unproven goes where your policy says: to a person, or to an AI model inside the budget you approve.
Arriving
Proven rules
Applied first
End point
Illustrative only: the animation produces these counts. Measured results follow further down this page.
No enterprise hands a new hire signing authority on day one. Authority follows a record: verified results on real cases first, then a mandate, with anything unfamiliar still escalated. The Cognitive Engine earns its mandate the same way.
Holds the systems of record and decides what actually happened. Only what it confirms is treated as evidence.
Suggests answers while the engine learns and, with your permission, covers cases it has not learned. It never defines what counts as correct.
Learns only from confirmed outcomes, and acts alone only after passing an exam on cases it has never seen.
They confirm outcomes and rule on cases your teams have decided both ways. In the strictest mode, everything the engine has not proved comes to them.
One five-stage path applies to each decision type. Types are taken over one at a time, and only once the evidence says the engine is ready.
It studies settled past cases, checked against your records or by a named reviewer. An AI model may suggest answers, but a suggestion never counts as evidence.
Plant systems confirm which past readings were truly normal and which called for action.
Every rule faces held-out cases it never saw in training. It may act only after getting at least 400 of them right, with no misses. The bar can go up, never down.
On a simulated factory, one rule scored 514 of 515 and was denied authority. With more evidence it later went 1,453 for 1,453. All on synthetic data.
A proven rule gives its answer in milliseconds, and no AI is involved; the same facts get the same answer. Six built-in reasoning checks must all agree before a rule qualifies; one dissent blocks it.
A normal reading from a healthy sensor is marked OK; its record cites the rule responsible.
Nothing outside its evidence is forced into a rule. Depending on the mode, the case gets your own past decision on identical facts, an AI model within budget, or a person. Conflicting past decisions always go to a person.
A reading outside every range it has seen goes to a person, or to AI if allowed. It is never guessed.
Every rule expires after 30 days by default, or sooner if you choose. One confirmed contradiction withdraws it on the spot, and only a new exam restores its authority. Evidence alone moves the engine.
If the plant redefines “normal,” the first confirmed case that disagrees stops the old rule; a replacement has to pass its own exam.
Most of an operational agent's daily decisions are routine, and routine work is where AI spend piles up. The engine takes over a decision only after proving it gets it right. The unproven goes to your own past decision or to AI within budget, and hand-offs go to your people.
Nine sensor readings are waiting for this shift.
Exam passed. Milliseconds, no AI cost.
Your past decision on identical facts, or AI inside your budget.
Handed off, or past decisions conflict.
Illustrative scenario based on the Factory Floor simulated business (synthetic data), where the engine has earned the routine sensor readings. Your own mix depends on which decisions your engine has proved and the mode you choose. In every mode, proven rules make no AI call, AI stays within budget and each hand-off goes to a person.
Proven decisions carry no AI cost at all. The AI calls that remain run inside per-run budgets and optional daily spend limits, with alerts before and at each limit.
How our pricing worksOperations get consistency. Given identical facts and rules, the answer holds, even on days an AI-only agent wanders. Replay any shift and each proven decision returns unchanged.
Every engine decision names its rule in plain words, and that rule had to win agreement from all six reasoning checks. Decisions stay on the record. Hand-offs always end with a person.
Read the Trust CenterYou set how much the engine may decide, per application and per run. One click on our simulated businesses benchmarks it against an agent that asks AI every time.
Your current system keeps every decision. The engine decides alongside it and is only scored, so leadership sees its record before it touches a single outcome.
Chain of authority
The engine goes first. For anything unproven, it reuses the decision your organization made before in an identical situation. Where past decisions conflict, a person decides. Only genuinely new situations reach AI, inside your budget.
Chain of authority
Only proven rules decide. Anything else lands with the backstop you name, usually a person.
Chain of authority
The Studio is where your team gathers evidence, trains and examines the engine, compares it with an AI-only agent and signs off before go-live. Viewers, trainers and administrators each get their own access, and reviewers confirm cases in a dedicated queue.
Each application shows its agents, their hard limits, the outcomes that must never happen and the decisions the engine has earned so far.
On our simulated businesses, one button compares the engine with an agent that asks AI every time, on accuracy, hand-offs, speed and AI calls. On your own system, it measures the engine against your system's decisions.
See each learned rule in plain words, the exam it passed and how each of the six reasoning checks answered.
Your people confirm cases in a single queue, including blind exam cards with no suggested answer, so every verdict is the reviewer's own.
AI drafts each decision's facts, limits and policy from your documents and systems, citing every source. Nothing takes effect until a person accepts it, and cited passages are re-checked on a schedule.
A checklist for every application before launch, owner sign-off, a security review, load and restore drills, and a hand-over pack for operations.
Distinct roles for viewers, trainers and administrators, masked identifiers and an audit trail of activity.
AI budgets per run and optional daily spend limits per application, with alerts before and at each limit. A separate switch decides whether AI may see case facts at all.
The Studio watches for expiring or withdrawn rules, slow decisions, AI spend and low disk space, and raises an alert.
We test the engine on twenty simulated businesses with synthetic data, from insurance claims and payments to factory floors and energy markets. Each run is set against an agent built the usual way, one that puts every decision to an AI model.
Field test on five simulated businesses with synthetic data, October 6, 2026.
Every decision the engine made itself, on all 20 simulated businesses (synthetic data, 1,000 events each, October 9, 2026). The engine currently decides cases in 14 of them.
Typical decision: 17–31 ms for the engine vs 3.3–4.2 s for an AI-only agent. Five simulated businesses, synthetic data, October 8, 2026.
Same simulated runs, synthetic data: per 1,000 decisions, the AI-only agent made 329–1,519 AI calls; Cognitive + AI mode made 0–14. Most AI-only figures are scaled up from runs capped at 300 calls.
On one simulated factory decision, an AI-only agent with 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 without a single error.
Nxgyle measured every figure in October 2026 on simulated businesses with synthetic data. “Right” means matching each simulated business's own policy. The figures show the method, not a customer's results, and do not on their own establish readiness for live use.
Your existing agents and systems stay in place; the engine takes on only the decisions it has proved. Your systems of record still decide what actually happened.
Agent platforms ask for decisions through the Cognitive Studio, which exposes the engine over a web API.
Feed in decisions your system has already made. The engine decides alongside without acting, is scored against yours, and turns new situations into cases for your people to confirm.
AI assistants can consult the engine through a local MCP server, the open standard AI tools use to reach business systems. It offers advice and never acts. A Python SDK is also available.
Packaged with Docker Compose for servers you run. Following any upgrade, each rule sits its exam again before it may act.
Actions under control. When the engine acts through your tools, it records its intent first and never retries on its own. If an outcome is uncertain, it stops and waits for your team.
Visit the Trust CenterThe engine is a correctness-focused prototype. Below: what is built, and what remains switched off unless you choose it.
No. It removes the recurring AI cost only from routine decisions it has proved it gets right. New cases still go to your people, or to an AI model within your budget.
No. Only rules that passed their exam may decide. Depending on the mode you choose, everything else is settled from your own past decisions, by budgeted AI, or by a person.
Every rule expires after 30 days by default, or sooner if you choose. Any confirmed contradiction pulls the rule at once, and its replacement acts only after a fresh exam.
Yes. Each engine decision names its rule in plain words. In the Studio, the Transparency screen shows the exam that rule passed and how each reasoning check answered.
By default, yes. When AI is asked to help, it may see case facts, with identifiers and names masked first. You can switch that off per application or data source, and disable AI drafting from your documents.
Not yet. Every result shown here was measured on simulated businesses with synthetic data. In shadow mode you can measure the engine against your own system's decisions before it acts on anything.
