How we work

A decision at every stage.

An engagement should make uncertainty smaller before the commitment gets larger. We agree the next output, how it will be judged and who can authorise the next step.

The delivery path

Every gate can lead to a narrower scope, a revised approach or a documented stop.

Delivery stages

What each stage leaves behind.

These are indicative planning ranges for a bounded workflow, not historical results or delivery guarantees. Access approvals, data preparation, integrations and client reviews affect the schedule. Dates and stage boundaries are agreed in the proposal.

Assess

Planning range · 2–4 weeks

Determine whether an agent addresses a real process constraint.

Deliverables: Current-state workflow, baseline measures, data inventory, risk register and a recommendation with alternatives.
Client contribution: A process owner, representative redacted examples, current volumes and a description of the existing process.
Decision gate: Continue only if the task is measurable, the data can be used appropriately and an owner accepts the proposed scope.

Design

Planning range · 1–2 weeks

Turn the selected use case into a testable implementation plan.

Deliverables: Architecture and data-flow diagrams, tool contracts, review rules, evaluation rubric and draft acceptance plan.
Client contribution: Technical and security review, access decisions, reviewers for the evaluation set and agreement on failure severity.
Decision gate: Confirm permitted data flows, success thresholds, budget limits and rollback design before build work starts.

Build

Planning range · 4–8 weeks

Implement a bounded pilot and test it against the agreed conditions.

Deliverables: Versioned source and configuration, integrations, evaluation results, failure analysis and a release candidate.
Client contribution: A test environment, approved data, integration support and recurring reviews of outputs and exceptions.
Decision gate: Release only if acceptance criteria are met and limitations are accepted in writing. Otherwise revise, narrow or stop.

Operate

Initial transition · 2–4 weeks

Introduce the system in a controlled scope and establish ownership.

Deliverables: Rollout and rollback records, dashboards, alert routes, runbooks, training walkthrough and handover inventory.
Client contribution: Named business and technical owners, an escalation route and agreement on any ongoing support coverage.
Decision gate: Complete handover when the owner can run, monitor and recover the agreed system. Ongoing support is separately scoped.

Delivery standards

Define “good enough” before building.

Start from an observed baseline

Record how the task is performed today: completion quality, handling time, error modes and human effort. Select measures that reflect the business task, not just text similarity.

Build a reviewable evaluation set

Use authorised, representative examples with expected outcomes or a scoring rubric. Include missing inputs, ambiguous requests, tool failures, permission boundaries and adversarial instructions. Version the dataset and keep acceptance cases separate from tuning examples.

Assign acceptance authority

The business owner signs off on task usefulness; the technical owner checks operability; the security owner reviews access and data boundaries. Unresolved severe failures block release even when an average score looks acceptable.

The acceptance record

Set thresholds with the client; no universal success percentage is assumed.

Release and observation

Test recovery as well as the happy path.

Begin with a limited cohort or a supervised workflow. Keep the previous working configuration available and nominate the person who can halt execution.

Trace task IDs, model and prompt versions, tool outcomes and approvals without retaining unnecessary sensitive content.
Observe completion, failed runs, timeouts, queue load, human escalation and spending.
Test duplicate requests, bounded retries, provider failure and rollback before expanding access.
Review sampled outputs and incidents after launch; turn confirmed failure modes into regression cases.

The proposal sets observation periods and acceptance thresholds. Deployment alone does not complete acceptance.

Production release gate

Each item needs an owner and evidence. An unchecked item is a release decision, not a footnote.

When an agent is not appropriate

A useful answer can be “do not build it”.

We recommend a deterministic workflow when rules are stable and exceptions can be handled directly. We recommend keeping work with people when outcomes cannot be checked, access cannot be bounded or the potential harm cannot be contained.

Stop conditions belong in the plan

If the pilot cannot meet the agreed quality, cost or safety limits, pause expansion. Document the evidence and remaining risks; agree whether to narrow the task, change the approach or close the project.

Closure includes the completed artefacts, known limitations, access revocation and agreed data return or deletion. Fees and ownership follow the written agreement; no failed pilot is presented as a production success.

Choose the simplest adequate approach

A staged recommendation can end without an agent implementation.

Start with one workflow.

Describe the task, the constraint and what a useful result would look like.

Discuss a project