In claims triage, credit checks and maintenance alerts, most decisions repeat. With the Cognitive Engine, rules come only from outcomes your team confirmed, and a rule may decide only after answering 400 or more unseen past cases correctly. Proven rules answer in milliseconds with no AI call and a reason your team can read. Anything unfamiliar lands with your reviewers, or with AI if you allow it.
Facts submitted
Stated rule
Classify as OK when the reading is inside tolerance and the sensor is responding without gaps.
Exam result
1,453 of 1,453 right
Unseen during learning
Six-way check
6 of 6 agree
From the Factory Floor simulated business. Measured in October 2026 on synthetic data.
Nx·Studio turns a defined business process into a digital worker that your company owns. Nx IDE adds AI help for the engineers who work through your business backlog. The Cognitive Engine is separate from both. Once its rules pass their exam, it settles the repeat decisions inside your processes without asking an AI model each time. Nxgyle builds all three.
In a claims or credit process, an agent that consults a large AI model on every case repeats work it has done many times. A pilot rarely shows the effect. At full volume, each case waits for the model and adds a usage charge, and identical cases may get different answers from one week to the next.
The engine checks every incoming case, such as a claim, a payment or a sensor alert, against its proven rules. A match is answered within milliseconds and without any model, and identical facts receive identical answers. A case with no proven rule is never guessed: the mode you select routes it to your reviewer, or to AI on a budget where you allow it.
Cases received
Rule lookup
Runs first
Resolution
Illustrative animation. These counters are not test results; the measured figures appear further down.
New employees do not approve decisions in their first week. They learn from past cases with known outcomes, prove themselves on unfamiliar work, and gain independence gradually, still asking for help when unsure. Evaluate the Cognitive Engine against the same standard.
Operates the systems that record what happened, and has the final say on it. The engine treats nothing else as evidence.
Proposes answers during the training period and, with your permission, takes cases the engine has not covered. None of its proposals count as evidence.
Learns its rules from verified results alone, and works without supervision only once its exam is passed.
They verify outcomes and resolve cases where earlier decisions disagree. Under the strictest mode, every case lacking a proven rule is theirs as well.
Each type of decision faces the same five tests, in order, before it touches a customer, a supplier or your ledger. The engine takes on decision types individually, and only when the measured results support it.
It trains only on past cases whose result was verified, either read from your own systems or confirmed by a person you name. AI proposals can shape a draft rule; they are not evidence.
Maintenance logs establish which earlier sensor readings were acceptable and which triggered work orders.
Every candidate rule sits an exam drawn from past cases it did not learn from. To qualify it needs 400 or more right answers and not one error. Your team may tighten that bar, but it cannot be loosened.
Measured on a simulated factory with synthetic data: one rule was wrong on 1 of 515 cases, so it could not act. With additional confirmed cases it reached 1,453 correct out of 1,453.
Once qualified, a rule returns its decision in a few milliseconds, and no model is called. Identical facts always produce an identical decision. Six built-in checking methods must all agree with a rule before it qualifies, and one dissent blocks it.
A healthy sensor's in-tolerance reading is classified OK; the decision record cites the rule applied.
Cases outside its training are not stretched to fit a rule. The mode you select sets the referral: an earlier decision your team made on the same facts, AI within budget, or a reviewer. Where earlier decisions disagree, a reviewer decides.
A value beyond anything in its training is referred, not estimated, and lands with a reviewer, or with AI where you have allowed that.
Each rule's approval runs for 30 days unless you choose a shorter period. A single confirmed contradiction withdraws it immediately, and the rule stays inactive until a new exam is passed.
Change the plant's definition of normal, and the first verified case against the old rule takes it out of service until a new rule qualifies.
A monitoring shift is mostly repeat decisions. The engine handles a decision type once its rule has qualified on unseen cases. Where no rule has qualified yet, it reuses what your team decided before on identical facts, or asks AI within budget. Hand-offs always reach a person.
The shift's nine readings are queued for scoring.
Rule answers directly, in milliseconds, with no AI.
Past answer reused if the facts match; if not, budgeted AI.
Anything handed off or decided inconsistently before.
An illustrative scenario from the Factory Floor simulated business (synthetic data), whose routine sensor readings are the decision type that has qualified. Your own proportions will reflect the decision types that pass and the mode you pick. Whatever the mode, a qualified rule never calls AI, AI spending respects your budgets, and hand-offs are routed to a person.
Decision types that qualify stop generating AI charges. Remaining AI calls are bounded by a budget for each run, plus daily caps per application if you set them, with alerts at 80% and 100%.
Pricing, explainedIdentical facts under an unchanged rule set get identical answers, whatever an AI-only agent does that day. Run the same shift twice to confirm every qualified decision matches.
Every engine decision carries its rule in everyday words, and all six built-in checks must agree with that rule. Studio activity goes into a log nobody can alter, and every hand-off ends with a person.
What the Trust Center coversDecide the engine's reach application by application, run by run. On our simulated businesses, you can compare it side by side with an AI-for-everything agent.
Your current process keeps deciding. The engine runs alongside, and its answers are scored but never applied, so you can measure its accuracy while nothing changes for customers.
Decision sequence
Proven rules answer first. Next comes your organization's own earlier decision on identical facts, used only when those past decisions agree; if they diverge, a reviewer takes it. AI sees only situations nobody has met before, within your budget.
Decision sequence
Decisions come only from qualified rules. Anything else drops to your system's safe default, normally a person.
Decision sequence
In the Studio, your team builds the evidence base, trains the engine, scores it against AI-only agents and approves the move to live use. Access has three roles: viewers read results, trainers run training and administrators manage settings. Reviewers confirm cases through a queue.
Per application, see the agents involved, the limits none of them may cross, the outcomes you rule out entirely, and how many decision types have qualified.
On each of our simulated businesses, a comparison run sets the engine beside an agent that asks AI every time. The report covers accuracy, AI calls, response time and the share left to people. Connected to your system, it compares the engine's answers with your existing decisions.
Open any learned rule to read it in everyday language, with its exam result and each checking method's verdict.
Reviewers confirm outcomes from a single queue. In blind reviews no proposed answer is shown, and the Studio reports on review quality.
AI reads your documents and systems to draft the facts, limits and policy behind each decision, with every point traced to its source. Each draft waits for a person's acceptance, and sources are rechecked weekly by default.
A readiness checklist per application, owner approval, a security review, load and restore tests, and a handover package.
Access by role, masking of personal identifiers, and an audit log that no one can alter or erase.
Each run carries its own AI budget, and each application can add a daily spending cap. Warnings go out as spend reaches 80% and then 100% of the cap, and one switch controls whether AI sees case facts.
Alerts for expiring or withdrawn rules, slow responses, AI spend and disk capacity, refreshed every 15 seconds.
Nxgyle evaluates the engine on twenty simulated businesses built on synthetic data, in domains such as insurance claims, payments, manufacturing and energy markets. In every run, the comparison is a conventional AI agent that sends each decision to a model. These are the measured results.
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 the same simulated factory decision, an AI-only agent 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 Nxgyle in October 2026 on simulated businesses with 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.
Your IT team connects the engine alongside your agents and systems instead of replacing them. Your systems still determine what actually happened.
Through the Cognitive Studio, your agent systems request decisions over the web; every system is issued a separate access key.
Your system streams the decisions it already makes. The engine answers in parallel without acting and is scored against them; anything unfamiliar becomes a case your reviewers confirm.
Through a local assistant server, your AI assistants can draw on the engine's advice. It advises but never acts. Your IT team also gets a developer toolkit.
Packaged in containers for your IT team to run on your servers. After any upgrade, existing rules stop deciding until each passes a new exam.
Controlled tool use. The engine logs what it intends to do before using any of your tools, and it does not repeat any action by itself. If the result of an action is unclear, the case waits for your team.
Trust Center controlsCapabilities in the current build, plus options that stay off until you enable them.
No. They remain essential for new situations. The engine settles only those routine decision types where a rule has passed its exam. Anything unfamiliar is passed to your reviewers, and to budget-limited AI only with your permission.
No. A rule acts only after its exam is passed. Until then, and depending on the mode, cases are answered from your earlier decisions, by budget-limited AI or by a reviewer.
A rule's approval expires at 30 days, or earlier if you prefer. One verified case that proves a rule wrong removes it straight away, and nothing replaces it until a new rule passes its own exam.
Yes. Every engine decision comes with its rule written in everyday language, and the Transparency screen adds its exam history and each checking method's verdict.
By default, AI may see the facts of a case when it is asked to help, but names and other identifiers from your data sources are masked first. That access can be disabled application by application or source by source, and document drafting can be turned off too.
Not yet. The figures on this page were all produced on simulated businesses with synthetic records. Shadow mode lets you score the engine on your own decisions before it may act.
