Five offers. One method.
Each offer has a written scope, a named buyer, a boundary and acceptance criteria agreed before work begins. Start with the one that fits the problem; most enterprises start with 01 or 03.
What decision is in front of you?
01Workflow architecture and baseline
Decide whether to build
2–3 weeks02Assured workflow engineering
Build one workflow
6–10 weeks03Agent release assurance
Test an agent you already have
4–6 weeks04Reliability operations
Keep a live workflow dependable
Monthly05Dedicated engineering delivery
Hand over an engineering workstream
3–24 monthsHow they connect
01 decides whether to build and sets the baseline. 02 builds it. 03 tests an agent you already have, yours or ours.
04 keeps a live workflow dependable. 05 supplies a managed team when the work is a programme, not a single workflow.
What is the same in all five
- A written scope, a named buyer and a boundary agreed before work begins.
- A named technical lead and a named delivery lead.
- Acceptance criteria written before work begins.
- Consequential actions stay under human approval. Authority is enforced in code outside the model.
- Each proposal states its direct cost, its margin and its value measure, so you can test them.
The five offers side by side
Who buys each one, what it produces, how long it runs and the boundary it keeps.
| Offer | Buyer | Output | Duration | Boundary |
|---|---|---|---|---|
| Workflow architecture and baseline | CTO, Head of AI, operating leader | Process map, alternative comparison, data and access review, failure-cost estimate, evaluation plan, architecture options with a recommendation | Two to three weeks | A recommendation not to build is a valid outcome |
| Assured workflow engineering | Product or platform owner, with the CTO or Head of AI as sponsor | One integrated workflow with controls, recovery paths, runbook, acceptance report and handover | Six to ten weeks | No unrestricted autonomous action on financial, legal or physical systems |
| Agent release assurance | AI platform lead, quality lead or security lead | Outcome, permission and regression tests, trajectory review, evidence pack, remediation backlog | Four to six weeks | An assessment is not a legal certification |
| Reliability operations | Platform owner or operations lead | Regression testing on every change, model-change evaluation, incident analysis, cost monitoring, controlled releases, monthly report | Monthly, under contract | Support hours and service objectives are written into the contract |
| Dedicated engineering delivery | CTO, VP Engineering, Head of Platform, consulting delivery lead | A managed team that owns a defined workstream, with weekly status and a tested exit plan | Three to twenty-four months | Not individual engineers under your management; not sold by the hour |
By company size
The same five offers and the same method for growing companies, mid-market companies and enterprises. What changes is where you start.
Growing companies
Start with 01 to find out whether AI fits a process. Build it with 02, then keep it running with 04. Each step is a separate decision.
Mid-market
Take one workflow through 01 and 02, and connect the next once the first meets what was agreed. 04 keeps them dependable; 05 when the work becomes a programme.
Enterprise
Most enterprises start with 01 or 03. Programmes move to 05, a managed team that owns a workstream under your controls.
Acceptance is agreed before the build.
- Define
- Bound
- Engineer
- Verify
- Operate
Define
We confirm the workflow owner, the current process, the intended improvement and the acceptance criteria. We compare AI with simpler options before choosing it.
Output: Result, owner, acceptance criteria, alternatives
Bound
We agree data access, allowed tools, actions that need approval, the risk owner and operating constraints. The boundary is written down before the build.
Output: Data access, tools, approvals, risk owner
Engineer
We build the integrations, state management, context handling and recovery behaviour the scope requires.
Output: Integrations, state, context, recovery
Verify
We test normal and difficult cases, inspect the resulting state of the systems involved, and document limitations.
Output: Normal, hard and adversarial cases
Operate
We agree deployment, support, monitoring, change control and documentation, and deliver an executable handover plan.
Output: Deployment, monitoring, handover
Human control stays explicit.
Each engagement identifies which decisions stay with people and which low-risk actions may be delegated. Changes to autonomy, tools or model versions trigger proportionate review and testing.
Start with one workflow.
Tell us the process, the cost of failure and the systems involved. We will respond with a proposed first step.