Systems Management .co.uk

AI can write the code. Engineering decides whether it ships.

Systems Management helps organisations put AI into production the way safety‑critical industries put anything into production: with clear requirements, a sound architecture, evidence from testing, and a release process that someone is accountable for.

It doesn’t replace agile, it accelerates it: less rework, tighter financial control, lower token spend, and faster, controlled delivery into the business.

RC‑2026.11.3claims‑triage‑assistant
  1. G1Requirements baselinedPASS
  2. G2Architecture reviewedPASS
  3. G3Model eval ≥ thresholdPASSaccuracy 0.94 vs 0.90 target · 1,200 labelled cases
  4. G4Integration & fallback testsPASS
  5. G5Safety & bias reviewPASS
  6. G6Change approval (CAB)PASS
Rollback plan: verified→ Released

Example release record. Every gate leaves evidence.

SHEET 01Rationale

Why engineering matters more now, not less

Generative AI has made building things cheap and fast. It has not made them correct, safe, or maintainable. The disciplines that kept aircraft, railways and banks running are exactly what AI systems are missing.

Probabilistic parts

A model is a component that can be wrong

Traditional software fails the same way every time. Models fail statistically. That needs defined acceptance thresholds, evaluation sets and fallbacks designed in from the start.

Speed of change

Output now outpaces review

When code, content and configuration are generated in minutes, the bottleneck moves to verification. Test management and release discipline become the real constraint on delivery.

System behaviour

Risk lives in the integration

Most AI incidents come from how a model is wired into data, people and other systems. Systems engineering looks at the whole, including interfaces, failure modes and the humans in the loop.

Accountability

Regulators expect evidence

The EU AI Act, ISO/IEC 42001 and sector regulators ask the same question: show me how you know it works. Good governance answers that with traceable records, not slide decks.

Financial control

Discipline is the cheapest cost control there is

AI makes it cheap to produce the wrong thing quickly. Every regenerate, retry and “try another prompt” is billed, and without agreed acceptance criteria teams keep iterating until something looks right. Engineering discipline puts AI spend on the same footing as any other cost line.

  • Tokens spent on evidence, not guesswork. Requirements and evaluation sets decide which prompts and models go forward, so the smallest model that passes is the one that ships.
  • Cost designed in. Caching, retrieval scope and context size are architecture decisions, made once and reviewed, not left to whoever writes the prompt.
  • Budgets at the gate. Cost per transaction is a release criterion alongside quality and safety, and usage is tracked against budget after go-live.
  • Less rework. Defects caught at requirements and design never become the regeneration, retesting and redeployment that consume tokens and people later.
40–50%

of project effort typically goes on avoidable rework

~100×

the cost of fixing a problem after delivery, compared with catching it in requirements and design

Source: Boehm & Basili, “Software Defect Reduction Top 10 List”, IEEE Computer, 2001
SHEET 02Practice areas

Six disciplines, one joined‑up system

Each area stands on its own, but the value comes from connecting them so that a requirement can be traced through the architecture, into a test, and out through a release.

SE

Systems engineering

Turning a business need into a system that can be built, tested and operated.

  • Stakeholder & system requirements
  • Interface and failure‑mode analysis
  • Requirements traceability
EA

Enterprise architecture

Deciding where AI belongs in the estate and where it doesn't.

  • Target state & transition roadmaps
  • Capability and application mapping
  • Architecture decision records
TM

Test management

Planning and running verification that produces evidence, not just green ticks.

  • Test strategy & entry/exit criteria
  • Model evaluation and regression suites
  • Defect and risk‑based prioritisation
GV

Governance & assurance

Controls that are proportionate, auditable and actually followed.

  • AI risk registers and impact assessments
  • Design authority & review boards
  • Alignment to ISO/IEC 42001 and the EU AI Act
RM

Release & change management

Getting changes into production safely, repeatedly and on purpose.

  • Release gates and readiness reviews
  • Change advisory process
  • Rollback, canary and model versioning
AI

AI integration engineering

The practical work of fitting models into real systems.

  • Guardrails, fallbacks & human review
  • Data and prompt configuration control
  • Production monitoring and drift review

Also ESG management: the same disciplines applied to sustainability reporting under the GHG Protocol, ESRS and GRI, for fashion, hospitality and supply chains. esg.systems-management.co.uk

SHEET 03Lifecycle

The V‑model, updated for AI

Every level of definition on the left has a matching level of verification on the right. AI adds new things to define and new things to prove, and a loop that brings production evidence back into the next baseline.

V-model lifecycle for AI systems Definition stages descend on the left: business need, requirements, architecture, detailed design. Build sits at the base. Verification stages ascend on the right: component verification, integration, system verification, operational validation. Each pair is linked by traceability, and production evidence loops back to the business need. production evidence → next baseline acceptance criteria system test cases interface tests evals Business needoutcomes · risk appetite Requirementsincl. model behaviour Architecturewhere AI sits, and why Detailed designdata · prompts · guards Build · train · integrateversioned models & data Component verificationunit tests · model evals Integrationinterfaces · fallbacks System verificationend‑to‑end · red team Operational validationmonitoring · drift review DEFINITION ↓ ↑ VERIFICATION

Production evidence loops back into the next baseline

  1. acceptance criteriaBusiness needoutcomes · risk appetiteOperational validationmonitoring · drift review
  2. system test casesRequirementsincl. model behaviourSystem verificationend‑to‑end · red team
  3. interface testsArchitecturewhere AI sits, and whyIntegrationinterfaces · fallbacks
  4. evalsDetailed designdata · prompts · guardsComponent verificationunit tests · model evals
  5. Build · train · integrateversioned models & data
FIG 3.1 · after ISO/IEC/IEEE 15288 technical processesdashed = traceability link
SHEET 04Release gates

Gates that ask the right question

A gate is only useful if it has a clear question, a named decision‑maker and evidence that can be checked. This gate model is our starting point, tailored to each organisation.

GateQuestion it answersEvidence requiredAI‑specific check
G0ConceptIs this worth doing, and is AI the right tool?Problem statement, options appraisal, outline business caseUse‑case risk classification
G1Requirements baselineDo we agree what “working” means?Signed‑off requirements, acceptance criteria, traceability matrixMeasurable performance thresholds and unacceptable behaviours
G2Design baselineWill this architecture meet the requirements safely?Architecture description, ADRs, interface specs, threat modelFallback path and human‑in‑the‑loop design
G3Test readinessAre we ready to prove it?Test plan, environments, test data, entry criteria metEvaluation set frozen and independent of training data
G4Release readinessIs it safe to change production?Test summary report, open defects accepted, rollback rehearsed, CAB approvalModel card, eval results vs threshold, bias & safety sign‑off
HOLD if any threshold missed
G5Operational acceptanceIs it doing what we said, in the real world?Service metrics, incident log, user feedback, benefits trackingDrift monitoring and scheduled re‑evaluation
SHEET 05References

Built on recognised frameworks

We use established standards so the work is familiar to auditors, suppliers and new team members, and adapt them so they stay lightweight.

ISO/IEC/IEEE 15288System life cycle processes
INCOSE SE HandbookSystems engineering practice
TOGAF · ArchiMateEnterprise architecture method & notation
ISO/IEC/IEEE 29119Software testing
ITIL 4Change enablement & release management
ISO/IEC 42001AI management systems
NIST AI RMFAI risk management
EU AI ActRisk classes & conformity obligations
SHEET 06Engagements

Ways to work together

2–3 weeks

Assurance health check

An independent review of an AI initiative or programme: requirements, architecture, testing and release controls.

You get: a RAG‑rated findings report and a prioritised remediation plan.
4–8 weeks

Lifecycle & governance set‑up

Design a proportionate delivery lifecycle, gate model and design authority that teams will actually use.

You get: gate definitions, templates, RACI and a pilot on a live project.
Ongoing

Embedded leadership

Part‑time technical leadership in chief architect, test lead or release manager roles while you build the in‑house capability.

You get: accountable engineering leadership in your governance forums.
SHEET 07Case studies

The same discipline, proven at scale

These programmes pre‑date generative AI, but the problems are the ones AI programmes face now: suppliers to hold to account, requirements to trace, services that cannot fail, and change that has to land safely.

Global logistics · Data‑centre migrationSE · TM · RM

Holding a supplier to account with traceable evidence

Challenge
An outsourced supplier was migrating a cross‑border data centre during a transition from internal IT to managed services. Its initial proposal was not fit for purpose, and slippage was turning into a dispute.
Approach
Mapped each user community's needs into requirements, then traced them through the supplier's migration approach, test strategy and individual test cases. Led the technical review and acceptance of a replacement solution, and moved suitable services to cloud platforms.
Result
The trace showed exactly where the supplier's plan failed, and the supplier was held liable for the direct costs of slippage. It also exposed early that the in‑cab driver mobility system could not be migrated, so a replacement was engineered before it became an outage.
National healthcare · Platform designSE · EA

A service for 250,000 concurrent users that could not go down

Challenge
Design a national platform with resilience to match its clinical importance.
Approach
Led the infrastructure architecture for two high‑resilience data‑centre network designs, built to fail over from one to the other.
Result
Failover within two hours for 250,000 concurrent users. At go‑live it was the largest Citrix farm in the world.
Central government · Estate consolidationSE · GV

Repeatable engineering under the strictest security regime

Challenge
Consolidate UNIX and Windows platforms from eight sites into three data centres on virtualised platforms, to the highest government security standards.
Approach
Applied systems engineering methods to create repeatable designs and governance processes, with security requirements traced across every platform.
Result
Faster, consistent delivery with full DR and business continuity, and a client that matured into running a formal managed service.
Global travel catering · Cloud migrationSE · EA · GV

A £4m outsource replaced by a programme costing under £1m

Challenge
One of the world’s largest specialist caterers, running food and drink franchises across airports and travel locations (if you have bought a coffee or a burger at an airport, there is a good chance it came through them), had planned a £4m outsourced programme with an external supplier to move its systems to the cloud.
Approach
Reviewed the programme using systems engineering disciplines. Careful requirements engineering defined what the business actually needed, and every requirement was traced through design, build and acceptance. Services were consolidated into the UK.
Result
Delivered on time for under £1m, around 20% of the planned budget, with all the required cost savings achieved and the £4m outsource removed entirely.
Central government · Test servicesTM · RM

Turning a test service from a cost into an asset

Challenge
An infrastructure test service was due to be decommissioned, and new environments were slow to provision.
Approach
Realigned the service with the client's development and integration strategy, made it able to run DR and BCP recovery tests, and used lean and Scrum methods for environment delivery.
Result
The client invested £2m in the service instead of closing it, and catalogue compute was delivered within 12 hours of request.
Defence · Secure accessEA · GV

Protectively marked data, reachable anywhere

Challenge
Give staff worldwide access to protectively marked data, and secure the links to partners across the international supply chain.
Approach
Designed a remote access service to government security guidance, and reviewed the security architecture and business purpose of every collaborative link to partner organisations.
Result
More than 4,000 users connected securely over the internet, saving half a million pounds a year.
ESG ratings · Scoring operationsSE · GV

A scoring cycle cut from four months to nine weeks, delivered on time to the day

Challenge
One of the largest ESG scoring companies needed its most complex scoring automations to run reliably and deliver on time, with around 350,000 individual scoring events to manage in each cycle. Earlier attempts, both internal and through several tier‑one consultancies, had failed to build and operate processes that could do it.
Approach
Applied systems engineering methods to the process design, building in control and feedback loops so that every stage was measured, checked and corrected while the cycle ran. The design was refined over four annual cycles.
Result
The scoring cycle fell from three to four months to nine weeks. Every annual delivery landed on the scheduled day, with significant cost savings across the business.
SHEET 08About

Thirty years engineering systems that had to work

Systems Management is a systems engineering and enterprise architecture practice. Our experience spans three decades of designing and delivering large infrastructure and IT programmes across the UK, Europe and Scandinavia, for government, defence, healthcare and logistics organisations.

The common thread is the discipline this site is about. We map user needs to requirements, trace those requirements through design, test strategy and test cases, and govern change so it lands safely. The case studies above show what that looks like in practice.

AI hasn't changed those fundamentals. It has raised the stakes. Systems Management brings that engineering discipline to teams adopting AI, so they can move quickly without losing track of what they've built or why they trust it.

  • Traceable by defaultEvery requirement has a test, every test has a result, every release has a record.
  • Proportionate controlGovernance scaled to risk, so low‑risk changes flow and high‑risk ones get scrutiny.
  • Evidence over assertionDecisions rest on data that someone else can check.

Methods TOGAF · ITIL · PRINCE2 · MSP · Lean and Six Sigma · Scrum · UML and model‑based design · requirements management and traceability · DR and BCP planning · security architecture

Putting AI into production? Let's engineer it properly.

Every week without clear requirements is paid for twice: once in rework, once in tokens. Tell us what you’re building.

Email us today to arrange a discovery session. We’ll come back with an initial view of where your programme is losing time and money, or soon will, and how to stop it.