Every enterprise AI conversation eventually reaches the same question: “Fine, but what would it actually do here?” The answer, across fourteen industries and several dozen operational use cases we have specified, designed or built, is that the work collapses into five shapes. Not five products, and not five industries. Five workflow shapes, each with the same bones underneath.
That matters to a buyer for a practical reason. If your process is one of the five, the platform, the controls and most of the agents already exist; what remains is connecting your systems and loading your policies. If it is not one of the five, we will tell you, and we will probably tell you that a model or a dashboard is the better tool.
Each shape below is described the same way: what it looks like, where it shows up, what it returns, and where we have built it.
1. Reconcile, explain, approve, post
What it looks like. A deterministic match runs first: two systems compared on agreed keys and tolerances, with no model anywhere in the matching path. Agents then explain each exception, citing the policy row and the account mapping behind every explanation. A controller or revenue manager approves the proposed corrections. Only then is anything posted back, through an action that carries its own permission requirement, so nobody outside that authority can release it.
Where it shows up. Financial close and multi-system reconciliation in any industry. Trade-promotion and retailer-deduction settlement in consumer goods. Supplier three-way match in manufacturing. Marketplace, interline and partner settlement in retail, airlines and telecom. Revenue-leakage detection across commerce, settlement and point-of-sale.
What it returns. Days off the close. Exceptions worked in hours instead of aged for weeks. Write-offs that fall because disputing a deduction is now cheaper than accepting it. Every posting traceable to the person who approved it and the evidence they saw.
Where we have built it. Reconciliation and leakage patterns are specified against our platform for a destinations operator and a cruise line. The deterministic reconciliation engine and the permission-carrying actions they depend on are running today.
2. Event, case, decide, act
What it looks like. A signed event arrives from a system that has already made a deterministic call: a fraud score crossed a threshold, a claim was filed, an alert fired, an application landed. Agents assemble the case from the systems that own the facts and check it against policy, clause by clause. A person with the right authority makes one bounded decision. The action that follows, whether hold, pay, release, refer or approve, runs only because that person approved it, and every irreversible action needs a second signature.
Where it shows up. Fraud exception handling in payments and commerce. Claims triage in insurance. Denial management and prior-authorization exceptions in healthcare. Alert adjudication after sanctions or anti-money-laundering screening. Underwriting referrals, loan exceptions, benefit and permit adjudication, billing disputes, change-order approval.
What it returns. Cases decided in hours, not days. Analyst and adjuster time spent on judgment rather than assembly. Recovery that goes up because nothing falls through. A complete dispute or audit file for every case, produced as a by-product of deciding it rather than reconstructed later.
Where we have built it. Governed Fraud Exception Handling and Auto Claim Triage and Settlement are running on the platform as reference solutions, each with its workflow, its controls and its tests published.
3. Disruption command
What it looks like. A trigger, whether weather, a port closure, an outage or a supplier failure, starts a planner that frames the decision and the deadline. Several researchers work in parallel on the candidate options. A gate holds until all of them have returned. A risk check and a reviewer rank the options on evidence, and a decision brief is produced. Two separate authorities then approve two separate things: the operational decision, and the money and the message that follow from it. Nothing reaches a customer, a vendor or a ledger until both have decided.
Where it shows up. Irregular operations in airlines and cruise. Storm and outage response in utilities. Supply-chain disruption in manufacturing. Major-incident command in technology. Surge response in insurance and healthcare. Any situation where more than one function must move at once, under time pressure, with liability attached to every commitment.
What it returns. A decided, costed plan inside the decision window instead of at the end of it. Compensation exposure known before it is promised. Vendor and regulatory deadlines met because they were surfaced as facts, not remembered by someone. One replayable record of how the decision was made, for the review that always follows.
Where we have built it. Voyage Disruption Command, seventeen steps and two approval authorities, is running on the platform as a reference solution.
4. The regulatory or compliance package
What it looks like. A schedule or an event starts evidence-gathering across several systems at once. Every fact is cited to its source. A checklist validator confirms completeness against the regulator's or the auditor's requirements. The package is assembled as a signed document, a compliance owner approves it, and it is submitted or filed through a governed action.
Where it shows up. Pre-arrival regulatory packages in shipping and cruise. Call reports, capital and liquidity filings in banking. Rate-case filings in utilities. Submission dossiers and corrective-action records in pharmaceuticals. Accreditation and survey evidence in healthcare. Security and compliance evidence collection in technology. Audit workpapers in professional services.
What it returns. Filing preparation cut from weeks to days. No more fire drills before the deadline. A citation register that is the audit evidence, rather than a package that has to be defended after the fact.
Where we have built it. A pre-arrival regulatory package is specified for a cruise operator on the same platform components as the disruption workflow. The artifact signing and evidence export it depends on are running today.
5. Governed reply and agent-assist
What it looks like. An inbound item, a ticket, an inquiry, a submission or a renewal, is classified and answered from an approved knowledge base and the customer's own record, with every statement cited. A validator checks the draft against policy and for personal data. Inside an allow-list of low-risk categories, the reply goes out without a person. Everything else is presented to an agent or a manager as a one-click approval. Any commitment the reply makes, a refund, a discount, a date change, is executed through an action with a hard ceiling and a named permission, so the draft can never promise something the approver is not allowed to give.
Where it shows up. Customer service in hospitality, retail, telecom and banking. Group, partner and commercial quoting, the deal desk, in hospitality, cruise, insurance and software. Renewal and retention desks. Complaint handling under regulatory scrutiny. Contract review against a playbook.
What it returns. Handle time down on assisted work. Quotes out in hours rather than days. Discounts and refunds outside policy that cannot be sent, rather than are asked not to be sent. Customer answers that arrive while the customer still cares.
Where we have built it. The Governed Deal Desk is running on the platform as a reference solution: fifteen steps, three discount levels, each enforced as a separate governed action.
What all five share
Look past the industry vocabulary and the same six things are true of every shape. They are the reason the shapes are reusable, and they are also the reason a buyer can trust the first one.
- Something deterministic goes first. A match, a score, a threshold, a schedule. The model never decides whether work should start.
- Agents assemble and explain; they do not decide. Every fact is cited to the system that owns it. When a source returns nothing, the run says so rather than filling the gap.
- A person with authority sits at the step that matters. Not a person somewhere in the loop. A named role at the specific decision where money moves, a customer is told, or an irreversible action is taken.
- Authority is enforced where the action happens. The operation that pays, posts, sends or holds declares the permission it requires. An approver without it is refused, and the refusal is recorded. Policy is not a paragraph the model was asked to respect.
- One path may run without a person, and it is the most governed path of all. Auto-release, auto-close and auto-send exist only inside a band a policy row explicitly permits, and that row is cited every time.
- The record is a by-product. The decision brief, the case file, the close memo, the quote, and the replay of how each was produced, come out of running the work rather than out of reconstructing it later.
Which shape is yours?
A short test. Take the process you are thinking about and ask:
- Is the trigger a fact (a file arrived, a score crossed a line, a date passed) or an opinion? If a fact, it is one of the five.
- Does more than one system have to change when it is done? If yes, it is not a chatbot problem.
- Is there a person who is accountable for the outcome afterwards, to an auditor, a regulator, a customer or a CFO? If yes, the approval and the record are not overhead. They are the product.
- Is the hard part predicting something (demand, price, risk) rather than deciding and acting on something already known? If yes, you want a model first and a shape second, and we will say so.
Three yeses and one no is the profile of every reference solution we have built.
Where to go next
Four of the five shapes are running on NeuralVantage™ as complete, buildable reference solutions: the agents, the workflow, the controls and the tests, validated on synthetic data so you can see exactly what each one does before committing a live process to it. The fifth is specified on the same components.
Bring us one process. In thirty minutes we will tell you which shape it is, what it returns, and what it would take to run it against your systems.
Reference solutions are built to demonstrate platform capability and validated on synthetic data. Policies, thresholds, delegated authority and connector permissions are configured and approved by each operator. No outcome figures are claimed; the returns described above are the objectives each solution is measured against, not results from a customer deployment. NeuralVantage™ does not perform fraud scoring, pricing optimization or demand forecasting. Where those models exist, it governs the actions taken on their output.