One developer with a coding assistant
is a faster developer.
A roster of agents across the whole lifecycle is a faster organisation.
The AI Delivery Lifecycle (AIDLC) puts a named agent on duty at all eight phases — from the moment an intent is captured to the telemetry that proves it worked. Gains stop being local to the keyboard and start compounding across the portfolio. This page sets out the mechanism, the return, and how to adopt it without betting the organisation on it.
- Lifecycle elapsed time
- −68%
- Cumulative output by quarter 8
- 3.0×
- Modelled payback
- Month 10
- Agents across eight phases
- 12
Figures on this page come from the delivery model set out in full below, with every assumption stated. They are a model you can re-run with your own numbers — not a customer average.
Why the coding assistant plateaus
A coding assistant makes writing code faster. Writing code is a quarter of the lifecycle. Everything else — deciding what to build, designing it, reviewing it, testing it, proving it is safe, releasing it, learning from it — runs at exactly the speed it always did.
- Shape & decideDiscovery, specification, architecture, planning
- BuildWriting the implementation
- AssureReview, test, security, compliance evidence
- Ship & learnRelease, documentation, telemetry, feedback
| Approach | Shape & decide | Build | Assure | Ship & learn | Total |
|---|---|---|---|---|---|
| Traditional delivery | 3.0 days | 2.5 days | 2.5 days | 2.0 days | 10.0 days — Baseline |
| Developer + coding assistant | 3.0 days | 1.2 days | 2.5 days | 2.0 days | 8.7 days — 13% faster |
| AIDLC — full agent roster | 0.9 days | 1.0 days | 0.8 days | 0.5 days | 3.2 days — 68% faster |
A named agent on duty at every phase
Twelve role-agents across eight phases. Each has a defined remit, produces a specific artefact, and hands off to the next. Human gates sit where judgement genuinely matters, and nowhere else.
Intent
Intent ClassifierClassified intent, value hypothesis, duplicate and overlap detection, routing to the right squad
Human gate Human accepts or rejects the intent
Context
Codebase Analyser AI ConsolidationImpact map, reuse candidates, the constraints pack drawn from the Context Library — specs, ADRs, threat models, personas, NFRs
Risk-gated Auto-passes unless the impact map is flagged
Plan
Composer ArchitectBrief, workstreams, work packages with acceptance criteria, data model and API contracts, test strategy, ADR
Human gate Human approves the plan at every trust level below fully autonomous
Code
Code GeneratorImplementation against the approved plan, inline documentation, a pull request per work package — from the Studio's own runner, GitHub Copilot or Claude
Risk-gated Auto-passes; every change still arrives as a reviewable PR, and its merge is held until the Definition of Done passes
Review
Code Reviewer Security ScannerFindings on the pull request itself, pinned to the commit they judged; coding actions sent back to the coding agent; a recorded verdict per item
Human gate Human verdict on the Definition of Done — the agents' approval and a named person's, both required
Verify
Test Architect Compliance AuditorUnit, integration and E2E suites; regulatory and governance validation; DPIA and control evidence
Risk-gated Auto-passes on a green suite; failures always wait for a human
Release
Release PublisherDeployment artefacts, release notes, changelog entries, traceability back to the originating intent
Human gate Two-key sign-off — automation plus a named human, at every trust level
Telemetry
Telemetry Analyst AI ConsolidationInstrumentation, effectiveness reporting, cost per shipped increment — and the next round of intents, evidenced
Risk-gated Auto-passes; findings feed the Effectiveness Report
Why delivery compounds instead of stepping up once
An assistant gives you a one-off step change and then a flat line. A lifecycle roster gives you a slope — because context accumulates, rework falls, and each squad that adopts inherits everything the last one learned.
- Developer + coding assistant
- AIDLC agent roster
| Quarter | Developer + coding assistant | AI Delivery Lifecycle roster |
|---|---|---|
| Q1 | 14 | 10 |
| Q2 | 14 | 16 |
| Q3 | 14 | 24 |
| Q4 | 14 | 34 |
| Q5 | 14 | 44 |
| Q6 | 14 | 56 |
| Q7 | 14 | 68 |
| Q8 | 14 | 80 |
| Cumulative over eight quarters | 112 | 332 |
Parallelism
Twelve agents work the same intent at once — the Test Architect writes suites while the Code Generator is still implementing, and the Security Scanner runs against every commit.
Where an assistant alone stops: A single developer with an assistant is still one thread. The assistant waits for the human, and the human waits for the assistant.
Context that accumulates
Every ADR, threat model, persona and API contract lands in the Context Library and is injected into every subsequent run. Output quality rises month over month without anyone retraining anything.
Where an assistant alone stops: Assistant context dies with the editor session. Nothing learned on Tuesday is available to the person sitting next to you on Wednesday.
Defects caught upstream
Review, security and compliance agents run before merge, not after release. Rework — the single largest tax on delivery capacity — falls, and the freed capacity goes back into new value.
Where an assistant alone stops: An assistant that only writes code can only generate more code to review. Volume rises; the review queue rises with it.
Governance as throughput
Audit evidence, cost attribution and control validation are by-products of the run, produced automatically. Approval stops being a queue and becomes a gate that clears in minutes.
Where an assistant alone stops: Assistants produce no evidence. Every governance artefact is still assembled by hand, weeks after the fact.
The return, with every assumption on the table
Adoption costs money before it saves money. Here is a deliberately conservative model for a 60-person product engineering organisation — including the months where the programme is under water.
Model assumptions
- Product engineering headcount
- 60 people
- Fully loaded cost per person
- £95,000 / year
- Annual engineering cost base
- £5.70m
- Capacity on new value today
- 32% — the rest is rework, coordination, ceremony and waiting
- Programme cost
- £85k in month 1 for enablement, then £48–54k / month for platform, inference and support
- Capacity released
- Ramping from 0% in month 1 to 34% by month 12
Benefit is counted only as released capacity redeployed to new value — no revenue upside, no headcount reduction, no avoided-cost claims. Swap in your own cost base and ramp and the shape of the curve holds; only the payback month moves.
- Net investment outstanding
- Net return delivered
| Month | Cumulative net position (£000s) |
|---|---|
| M1 | -85 |
| M2 | -140 |
| M3 | -174 |
| M4 | -193 |
| M5 | -194 |
| M6 | -175 |
| M7 | -140 |
| M8 | -85 |
| M9 | -14 |
| M10 | 72 |
| M11 | 170 |
| M12 | 277 |
How to start, and how to scale
Nobody arrives at an AI-first lifecycle in one move, and organisations that try tend to fail loudly enough to set themselves back a year. Five rungs, each with an exit criterion you have to actually meet before you climb.
- Rung 0
Prove
Weeks 1–4One squad, one product area, fully supervised. Every agent action is proposed and a human approves it before it runs. You are buying evidence, not throughput.
Turn on in AI Product Studio
- One studio
- Three agents — Intent Classifier, Composer, Code Generator
- Supervised trust level
- Cost ceiling per run
Exit criterion
The squad ships one real increment through all eight phases and the audit trail satisfies your risk function.
- Rung 1
Standardise
Months 2–3Stop improvising. Write down what "ready" and "done" mean, load the constraints your agents must respect, and make the same run reproducible by anyone in the squad.
Turn on in AI Product Studio
- Context Library — ADRs, threat models, NFRs, personas
- Definition of Ready / Definition of Done per phase
- Recorded gate decisions on every work package
- Work Packages for slice-level tracking
Exit criterion
Two different people get comparable output from the same intent. Rework on agent-produced work falls below your human baseline.
- Rung 2
Widen
Months 4–6Three to five squads, semi-autonomous. Agents plan and execute inside cost and risk thresholds you set; humans approve each work package and decide what ships. This is where the compounding starts to show.
Turn on in AI Product Studio
- Semi-autonomous trust level
- Full twelve-agent roster
- Security Scanner and Compliance Auditor on every run
- Per-squad cost attribution
Exit criterion
Lead time for change has halved against your pre-programme baseline, with no rise in change failure rate.
- Rung 3
Industrialise
Months 7–12The lifecycle becomes the operating model rather than a pilot. Policy is set centrally, evidence is produced automatically, and the Effectiveness Report — not anecdote — decides where to invest next.
Turn on in AI Product Studio
- Organisation-wide AI policy and guardrails
- DPIA and model-risk register
- Effectiveness Report and cost-per-increment reporting
- Immutable audit across every studio
Exit criterion
Governance signs off on the evidence pack without a manual assembly exercise. Cost per shipped increment is trending down quarter on quarter.
- Rung 4
Operate at portfolio scale
Year 2Multiple studios across multiple value streams, each at the trust level it has earned. Autonomy is granted per studio on evidence, never granted by default, and can be withdrawn the same way.
Turn on in AI Product Studio
- Studio-per-value-stream
- Fully autonomous trust level where earned
- Portfolio governance and strategy-to-value traceability
- Cross-studio agent marketplace and blueprints
Exit criterion
New value streams reach Rung 2 in weeks rather than quarters, because the blueprints, policy and context already exist.
Your first ninety days
Rungs 0 and 1, spelled out. One squad, one product area, and a baseline you took before you changed anything.
Establish the ground truth
- Baseline the four DORA measures before you change anything — without a baseline you cannot prove anything later.
- Pick one squad and one product area with real, visible delivery pressure. Not a sandbox.
- Stand up the studio, connect the repository, set the cost ceiling, and run at the supervised trust level only.
- Get your risk and compliance functions into the Definition of Done and the release approval in week one, not week twelve.
Make it repeatable
- Load the Context Library with the constraints that actually bind you — architecture decisions, threat models, non-functional requirements, personas.
- Write the Definition of Ready and Definition of Done for each of the eight phases.
- Bring the full agent roster into service, starting with Code Reviewer, Test Architect and Security Scanner.
- Review the first Effectiveness Report as a leadership team and agree what "good" looks like numerically.
Earn the next trust level
- Move the pilot squad to semi-autonomous once the evidence supports it — and only then.
- Onboard the second and third squads using the first squad's blueprint rather than starting over.
- Publish cost per shipped increment alongside throughput, so the conversation is about value and not about token spend.
- Set the organisation-wide AI policy from what you have learned, not from what a vendor deck asserted.
What actually changes at enterprise scale
At one squad this is a productivity story. At portfolio scale it is an operating-model change — the binding constraint on delivery moves off engineering hours entirely.
| Dimension | Before — project-and-ticket delivery | After — AIDLC at scale |
|---|---|---|
| Unit of work | A ticket, sized in story points, queued against a team's capacity | An intent, broken into work packages by the Composer and executed by a roster |
| Constraint on delivery | Engineering hours — the binding constraint on everything | Decision throughput and governance capacity, both of which scale far more cheaply |
| Role of the engineer | Implementer, with review and testing squeezed in at the end | Reviewer, architect and arbiter of quality — judgement work, not typing work |
| Where quality is enforced | Late, by people, under release pressure | Upstream, by agents, on every commit, with the evidence produced automatically |
| Audit and evidence | Assembled by hand, weeks after the fact, from whatever survived | An immutable, append-only trail generated as a by-product of the run itself |
| Cost visibility | Headcount by cost centre; the cost of an individual feature is unknowable | Cost per intent, per agent and per shipped increment, attributed in real time |
| Scaling a new value stream | Recruit, onboard, and wait two or three quarters for productivity | Instantiate a studio from a blueprint and inherit the policy, roster and context on day one |
| Risk posture | AI use is shadow, ungoverned and invisible to the risk function | Every model call is policy-bounded, cost-capped, logged and attributable to a named human owner |
Measure these eight, from day one
The four DORA measures prove you have not traded stability for speed. The four AI measures prove the compounding is real rather than anecdotal.
Lead time for change
The headline claim. Measure it end to end — intent captured to running in production — not first commit to merge.
Reported in: Effectiveness ReportDeployment frequency
Compounding shows up here first. If throughput rises but deployments do not, the constraint has moved to release.
Reported in: ReleaseChange failure rate
The honesty check. Speed that costs you stability is not a gain, and this is the number that proves it either way.
Reported in: TelemetryTime to restore
Agents should shorten diagnosis, not lengthen it. If this rises, your context is thin.
Reported in: TelemetryAgent acceptance rate
The share of agent output accepted without amendment. Rising acceptance is context maturing; falling acceptance is drift.
Reported in: VerifyRework ratio
Work redone after review or release. This is where released capacity actually comes from, so measure it directly.
Reported in: Effectiveness ReportCost per shipped increment
Programme cost divided by increments delivered. The only number that makes the ROI argument to a CFO.
Reported in: Effectiveness ReportFindings caught pre-merge
Security and compliance issues stopped before merge versus after release. The shift-left claim, quantified.
Reported in: Risk & DPIAStart at Rung 0 this week
One squad, one product area, fully supervised, with a cost ceiling and a complete audit trail. That is a four-week commitment, not a transformation programme — and it produces the evidence every subsequent decision depends on.