Capabilities

What we are competent to build.

SteelWorks Intelligence Systems designs, builds and operates structural AI systems for mid-market and enterprise operations — systems with an org chart, defined decision rights, approval gates and a security function. Not a chatbot. Not a pile of automations. An operating structure that happens to be staffed by specialized agents.

Every capability on this page runs in production inside our own company before it is offered to anyone else. What follows is the method, the practice areas, and the evidence behind each one.

Start with an assessment Talk to us about an engagement

The operating model

Five stages. None of them skipped.

Architecture precedes build, always. An operation running without a named owner and a defined decision boundary is an incident that has not been assigned a date yet.

01 — Assess

Where the operation actually breaks, which decisions are being made without a defined owner, and what is already running that nobody is measuring.

02 — Architect

The org chart first: units, decision rights, approval gates, escalation paths, and the boundaries an agent may never cross. The chart is a deliverable in its own right and it is reviewed before a line of the system is built.

03 — Build

Operations instanced against the chart, on infrastructure you designate, with reporting your team reads rather than a dashboard only we understand.

04 — Assure

Governance and security measured against a published framework before any part of the system is trusted with production work. Findings are owned and dated, including the ones that are not flattering.

05 — Operate

Monitoring, failure triage, model re-scoring as new models ship, control re-measurement, and a standing review that catches an operation whose owner has quietly drifted.

Practice areas

Five things we do, and the mechanism behind each.

Structural systems architecture

The chart is the product. Which unit owns each operation, what that unit decides unaided, where a human approval gate sits, what escalates and to whom, and what is permanently out of bounds regardless of instruction. Systems designed this way can be audited, handed over and extended, because the authority model is explicit rather than emergent.

deliverable: org chart, decision rights, gate map
reviewed: before build begins
running here: 29 units, 70 specialist agents

Operational agent teams

Cross-functional teams of specialized agents managed through hierarchy rather than prompt chaining. Each team has a command layer that can stop work, a delivery layer that does it, and an assurance layer that checks it independently of the agent that produced it. The structure is what makes the output reviewable by someone who was not in the loop.

structure: command, delivery, assurance
failure mode: stop, escalate, never silently proceed
running here: continuous, unattended

Model selection and routing

Every model is graded continuously against our own task mix, and routing is set by measured evidence rather than vendor claim or preference. A model that has not cleared a minimum evidence floor on a task does not get that task, no matter how well it scores on five examples. When a stronger model ships, the ladder changes because the numbers changed.

basis: graded turns on live work
guard: minimum-evidence floor per task
re-scored: as new models ship

AI governance and assurance

Measured against the NIST AI Risk Management Framework. Prompt-injection screening on inbound content, secret handling and privilege boundaries, single-writer guards that stop two processes quietly overwriting each other, an external-action gate that checks whether an action is permissible before it checks whether it is authorized, and a kill switch that works. The same instrument we run on ourselves, on a schedule, with the results dated.

framework: NIST AI RMF
output: scorecard, owned and dated findings
cadence: standing, not one-off

Vertical systems

The work that turns a general system into one that belongs to an industry: the corpus it reasons over, the evaluation set that decides whether it is good enough, and the regulatory envelope it must operate inside. Built for energy, oil and gas, distribution, security, financial services, healthcare, retail and marketing.

See the eight markets

components: corpus, evaluation set, envelope
decides: fit for the industry, not fit in general

Evidence

What we can prove, rather than assert.

Most of this field demonstrates capability with a prepared scenario. We would rather show you the telemetry from a system that has been running our own company unattended, measured by the same instruments we would install in yours.

Capability	How it is measured here
Unattended operation	12,314 scheduled runs recorded since 16 October 2026, 99.6% completed clean
Model selection	407 graded turns across 14 models; the count grows as turns are graded
Governance	25 NIST AI RMF controls, each measured and dated
Adversarial testing	55 prompt-injection fixtures held as a standing suite
Acceptance grading	276 implementation probes on the register, 99 passing on the production host after the 13 September migration
Published surface	18 pages in the sitemap; every figure on them is read from live fleet state and audited after each deploy

We publish the unflattering figures as well. The governance scorecard currently reads 56.7% at the Developing tier — the gaps are enumerated, owned and dated, and we would rather a prospective client read that than a claim of maturity we cannot evidence. A vendor who reports only their good numbers is telling you which numbers they do not collect.

The layer model every engagement is architected against. Yours would be drawn the same way, against your operation.

Deployment

Where it runs is your decision, not our default.

Systems can be deployed on infrastructure you already own, inside your own cloud tenancy, or on dedicated hosting with defined isolation and service levels where the workload and the data warrant it. Data residency, tenancy isolation and model hosting are engagement decisions taken with your security team. We do not move a client onto an arrangement that suits us.

What we will not do

Overstate the system, or report a result we did not measure.

We build and operate systems that orchestrate strong models under governance. Where a claim would require a capability we have not built, we say so instead of implying it. Every public claim passes a gate before it ships: no fabricated ratings, no unverifiable numbers, no citation we did not see produced. Where an audit leg could not be assessed, the report says which one and why.

Engagements

Every engagement begins with a scoped assessment.

We do not scope a system we have not examined, and we do not publish a rate card. Governance requirements, deployment constraints and the state of what is already running differ enough between operations that a list figure would be a fiction — and a number quoted before the assessment is a number someone will have to walk back.

Tell us what the operation is and where it breaks. You get a written reply from a principal, not a sequence from a marketing system.