Most platform sites make you guess. This one lays out what Kernos actually delivers — and takes your requirement — so you know where you stand before any contract, call, or invoice.
If one of these is yours, Kernos was built for it — and there is a card for it below.
That instinct is correct, and it is exactly what we built: every write proposed, checked, approved, then executed.
Approval DAGs with segregation of duties, and an append-only chain that can replay any decision, even six months later.
Your own agent, connected to our MCP server, does the delivery work on your side of the table. More on that below.
Construction compliance and e-commerce revenue recovery run today as packaged scenarios: verified before the payment run, not after.
The forward-deployed engineer model — vendor engineers embedded at your site, priced by the engineer-year — is how enterprise AI delivery works today. It produces real outcomes, and it prices out most of the market.
A vendor staffs engineers on your premises for months. Discovery, integration, and delivery all flow through their calendar and their rate card. You rent judgment you cannot keep — priced by the engineer-year, queued by their bench.
Kernos packages its domain knowledge as an MCP server: 24 semantic tools plus a manual that teaches your agent the platform. Your own AI — Claude Code, Codex, whatever your team already runs — connects to it and does the delivery work on your side of the table. No embed, no rate card.
Where we still show up in person: complex industry domains deserve hands-on help, and that's what the fixed-price pilot is for. The difference is that you buy a scoped project, not a headcount with a bench.
Not a feature list — a delivery map. Each card says what lands, what signals it, and where its edge is.
Business objects, relations, and lifecycle rules become a typed model your systems and agents share. Signals: "the same customer means three things in three systems." Edge: not a data warehouse replacement.
Agents never mutate a system of record directly. Every write is proposed, checked against policy, approved by a qualified human, and only then executed. Signals: "we can't let AI touch the ERP unsupervised." Edge: not for fully autonomous writes.
OData, IDoc, BAPI, events, CDC — with idempotent retries, dead-letter queues, and three-way reconciliation. A write counts only when SAP's own event confirms it. Signals: payment runs, journal entries, vendor master data. Edge: other ERPs go through connectors, case by case.
Multi-node DAGs with segregation of duties, delegation, and batch decisions. The initiator cannot approve their own proposal — the engine refuses it. Signals: payment gates, procurement queues. Edge: not a BPM suite for human-only processes.
Append-only, hash-linked records carrying correlation and causation IDs — walk backward from a disputed outcome to every decision that fed it. Signals: SOX and EU AI Act evidence demands. Edge: it produces evidence, not a certification.
Construction compliance (COI verification, payment gating) and e-commerce revenue recovery run today as packaged scenarios, not templates. Signals: subcontractor certificates, settlement disputes. Edge: other verticals start from the ontology, not a pack.
Before you submit, score yourself against these — they are the same tests we apply. Fall short on one? Submit anyway. Telling you what's missing is also a useful answer.
A process that runs every day or every week earns governance; a one-off project doesn't. If the answer to "how often" is "just this once," no platform will pay for itself — and we'll tell you that.
"Block payments when the certificate is expired" can be encoded and enforced. "Decide which vendor feels right" cannot. If the decision logic can't be written down, the first step is writing it down — and we can help with that, but it changes the project.
Requirements that shift every sprint can't stabilize into an ontology, an approval DAG, or an audit trail. A stable goal with changing inputs is work; a changing goal is discovery. We deliver the first and staff the second differently.
Customer-facing conversation with no system writes? Dify-class builders do it well and cheap — we say so on the comparison page.
If no human may ever approve anything, our core mechanism is your obstacle, not your asset. Kernos is built for the opposite bet.
Read-only BI over a warehouse is a solved problem elsewhere. Kernos earns its keep where agents must act, not just look.
Describe it in plain words — one sentence is enough, and rough is fine. We do the translating. An engineer reads it and replies with a fit verdict — strong, partial, or outside scope — usually the same day. No sales sequence, no mailing list, and the verdict is advisory, not a delivery commitment.
An engineer reads your description — not a sales team. If the honest answer is "not us," you get that answer, with a pointer to what would fit better. We would rather be useful than persistent.
It lands in our assessment queue and an engineer reads it — not a bot. We reply with a fit verdict and, when it's a strong fit, the shortest path to a pilot. Nothing enters a mailing list.
Recurring, rule-bound, and stable — the three marks above. A payment run that happens weekly, follows contract rules you can quote, and has a defined approver is a textbook case. If your process fails one mark, say so honestly in the description; a partial fit answered straight beats a mismatch discovered in month two.
Both. The workbench runs an automated check in about a minute — three per account. Prefer a human? Submit here and an engineer replies, usually the same day. The automated verdict is an initial assessment; the human one is not.
No embedded headcount. When a requirement truly needs hands-on delivery, we scope it as a fixed-price pilot — $1,500, first project, 60 days — with the approval loop and audit chain live and your team operating the result. After that, your own agents carry the routine work through the MCP server.