FIELD · Handbook · 6 min read
FIELD: the Kuzushi Labs engagement methodology
A connected method for evaluating, scoping, delivering and owning customer work.
Arunank Sharan · Kuzushi Labs ·
In this article
FIELD is Kuzushi Labs’ methodology for evaluating and scoping client work, carrying an agreed customer outcome through delivery, adoption and ongoing ownership. It gives the commercial lead, delivery team and customer one shared record of what is worth doing, what must be true, and what evidence supports the next commitment.
The unit of work is a customer workflow with an accountable owner. A model, connector or application is a means of improving that workflow. FIELD applies to AI products, enterprise integrations, analytics, internal platforms and operational automation; the depth of assessment follows the consequence of failure.
Five stages, five decisions
| Stage | Core question | Required artifact | Decision |
|---|---|---|---|
| F — Frame the outcome | Whose work improves, and why is it worth changing? | Outcome contract, baseline, alternatives, sponsor and constraints | Fund discovery, redirect or decline |
| I — Investigate the workflow | What actually happens, including exceptions and waiting? | Observed workflow, source inventory, access probes, assumption register | Select a feasible first journey |
| E — Engineer the solution | What is the smallest complete intervention that meets the contract? | Scope, resource model, authority boundaries, delivery and acceptance plan | Commit to a bounded build |
| L — Launch with evidence | Does the proposed release work for this population under these conditions? | Versioned evidence, rollout limits, monitoring and recovery plan | Pilot, expand, revise, hold or stop |
| D — Develop ownership and reuse | Can the customer operate it and realize its value? | Operating owner, benefit review, handover evidence, reuse and exit plan | Transfer, improve, extend or retire |
The stages are revisited when evidence changes. A failed access probe returns the work to investigation. A pilot that increases review effort returns the value model to framing. A new tenant, language or write permission changes the scope and can invalidate earlier evidence.
The connected engagement record
Each important promise has a chain:
Outcome → requirement → mechanism → test → evidence → decision.
For example: “Reduce invoice handling effort” depends on “Never apply an unapproved correction,” which depends on a scoped approval record and an idempotent business action. Tests submit expired approvals, repeat a request after a timeout and compare the authoritative ledger. Evidence identifies the exact policy, connector and release tested. The finance owner then decides whether the bounded workflow may enter a pilot.
A benefit forecast cannot compensate for failed authority. A passing accuracy benchmark cannot compensate for missing recovery. An artifact from a previous configuration cannot silently authorize a changed release. This separation is the central operating rule of FIELD.
Assessment coverage
Use this map at intake and revisit it at every material scope change. Each dimension needs an owner, evidence status and next decision; a blank cell is an unresolved item.
| Dimension | What the engagement record must establish |
|---|---|
| Strategic fit | Business priority, competing initiatives, urgency and consequence of doing nothing |
| Customer value | Eligible population, baseline, outcome denominator and benefit owner |
| Client readiness | Sponsor authority, user access, domain experts, delivery capacity and procurement path |
| Workflow | Trigger, steps, waiting, handoffs, exceptions and final business state |
| Stakeholders | Who uses, pays, supplies data, approves, operates and absorbs failure |
| Scope | Included journey, exclusions, dependencies, acceptance and change mechanism |
| Data | Rights, provenance, quality, freshness, lineage, retention and deletion |
| Integration | API behavior, quotas, network access, identity, schemas and recovery |
| Quality | Task correctness, coverage, abstention, subgroup outcomes and review burden |
| Authority | Read/write boundaries, approvals, spending limits and irreversible effects |
| Performance | Workload distribution, throughput, concurrency, latency and capacity limits |
| Reliability | Availability, degraded service, recovery time and acceptable data loss |
| Security and privacy | Threats, least privilege, tenant separation, sensitive data and auditability |
| Economics | Adoption, human effort, recurring costs, investment and downside sensitivity |
| Delivery | Critical path, staffing, milestones, dependencies and decision deadlines |
| Commercial terms | Price basis, payment triggers, usage exposure, support and change control |
| Adoption | In-workflow access, incentives, accessibility, training and feedback |
| Operations | Monitoring, incident response, support hours, escalation and staffing |
| Governance | Decision rights, evidence freshness, exceptions and periodic review |
| Ownership | Repositories, accounts, credentials, documentation and recovery competence |
| Reuse and exit | Second consumer, portability, vendor substitution, export and retirement |
Tailor the weight, retain the logic
A reversible internal report may need a one-page outcome contract, a small evaluation set and an operator. A write-capable agent touching payments needs explicit effect semantics, separation of duties, reconciliation and tested recovery. A multinational deployment adds language, residency, identity and support boundaries.
Use three assessment states: supported, unknown, and failed. Add “not applicable” only with a rationale and owner. Avoid collapsing them into a single percentage. Priority scores can help choose discovery work among otherwise eligible opportunities; they do not grant data access or release authority.
A practical first session takes 90 minutes: 20 minutes on the workflow and baseline, 20 on constraints and ownership, 20 on alternatives and value, 20 on unknowns and the first experiment, and 10 on the decision record. These are facilitation allocations, not a required ceremony.
How to use this publication
Start with qualification, then follow one journey through scope and evidence. Use the assessment kit in a live engagement. The workbench lets a team vary the economic assumptions and inspect its release prerequisites. The application library applies the same method to 36 distinct situations.
Reading the numbers. The applications are worked analyses with explicit planning assumptions, calculated implications and proposed acceptance thresholds. They demonstrate how decisions are made; their figures are not reported customer measurements or undisclosed client results. A target becomes a result only when the corresponding measurement is recorded. Specific assumptions make an analysis reproducible without exposing confidential client records.
FIELD is an attributed synthesis of established practices and Kuzushi Labs’ delivery approach. The particular combination, templates and decision discipline are ours; the underlying ideas have substantial prior art. The source register identifies those influences and the limits of the claims made here.