Whole-org · Kestrel · falsifier-gated · workspace-backed
The function-by-function
transformation map.
Every Kestrel function triaged — which tasks AI takes end-to-end, which it assists, which stay human. Residual headcount named and justified. Every claim backed by a live workspace that actually does the job. The ones without a workspace are clearly labelled.
01 · The org we're working from
Real. Named. Sized where sourced.
This is the Kestrel org model from triangulated public sources — theorg.com, LinkedIn, and vacancy corpus. Named individuals are for organisational structure only. Sizes marked sourced come directly from theorg.com. Sizes marked estimated are inferred from vacancy density and org ratios.
Technology team breakdown (theorg.com, sourced):
02 · The triage
Function by function. Honest verdict.
Each function opens to show the task triage and residual team. Evidence labels tell you how confident each verdict is.
Technology
| Team | HC | AI-takes | AI-assists | Human-only | Evidence |
|---|---|---|---|---|---|
| Software Dev + Eng | 44 | ~20% Boilerplate, test gen, docs |
~45% Code review assist, refactoring, debugging |
~35% Architecture, production incidents, novel features, review accountability |
JD analysis |
| QA | 13 | ~40% Test generation, regression scripts, test plan drafts |
~30% Exploratory test guidance, CAPA effectiveness tracking |
~30% UAT coordination, release sign-off, customer-facing quality accountability |
Quality Lead → |
| Eng & Technical Solutions | 19 | ~10% Integration docs, config templates, API specs |
~45% Solution design support, RFP analysis, technical assessment |
~45% Client relationships, novel integrations, accountability for delivery |
JD analysis |
| Engineering Management | 12 | ~5% Status reports, OKR tracking dashboards |
~30% Hiring briefs, team health monitoring, capacity forecasting |
~65% People decisions, exec relationships, accountability, culture |
JD analysis |
| Data & Reporting | ~19 est. | ~55% Routine SQL, dashboard refresh, data-quality monitoring, debt alerts |
~25% Analysis drafts, anomaly interpretation, estimated-read exposure reports |
~20% Metric governance sign-off, pricing decisions, Ofgem-facing accuracy |
Senior Data Analyst → |
Residual team — Technology
| Architects & principals (Sw Dev/Eng) | Accountable for system design, production stability, novel features. Human-core of the 44 engineers. |
| Tech leads & senior engineers | Own the review + merge accountability AI can't carry. One per squad. |
| QA leads + release owners | Sign off. UAT. Client-facing quality. AI generates; humans validate and accept. |
| Solutions architects (E&TS) | Client relationships and integration design are irreducibly human at this deal size. |
| Engineering managers | People management, exec credibility, culture — not offloadable. |
| Analytics owner + 2–3 analysts | Metric governance, sign-off on AI-generated outputs, Ofgem-facing accuracy. |
Residual estimate: ~68–75 FTE (saving 32–39 FTE, 30–36% reduction over 3-year steady state). £ saved: 32–39 × £65k avg. Nottingham tech salary = £2.1–2.5M/yr.
Salary basis: Glassdoor/Reed median Software Developer Nottingham 2025 £55–75k; blended ~£65k including data/QA/solutions roles. Labelled as assumption.
Product
5 live workspaces cover Billing, Payments, Credit Risk, Asset Management and Home & Business Moves. These represent the majority of Kestrel's Meridian module PM remit. Head of Product Platform and Product Marketing Manager are inferred from JD analysis (no workspace).
| Role / module | AI-takes | AI-assists | Human-only | Evidence |
|---|---|---|---|---|
| PM — Billing | ~35% Exception monitoring, back-billing alerts, unbilled-book tracking, board-pack data pulls |
~40% Spec drafts, backlog analysis, Ofgem-compliance gap analysis |
~25% Regulatory sign-off, customer escalations, roadmap prioritisation |
PM Billing → |
| PM — Payments | ~35% DD funnel analysis, retry-recovery modelling, arrears monitoring |
~40% Dunning strategy analysis, collection board pack, FCA-risk review |
~25% FCA compliance decisions, commercial terms, PSR-group accountability |
PM Payments → |
| PM — Credit Risk | ~35% Debt aging, arrangement adherence, write-off modelling, recovery funnel |
~40% Treatment strategy analysis, vulnerability-segment review, coverage gap briefings |
~25% PSR accountability, Ofgem regulatory decisions, collections-partner oversight |
PM Credit Risk → |
| PM — Asset Management | ~40% Fleet health monitoring, smart-rollout tracking, read-staleness alerts, settlement-class gaps |
~35% Regional gap analysis, SMETS2 programme reporting, supplier SLA review |
~25% Supplier negotiations, BEIS/Ofgem compliance sign-off, exchange-programme accountability |
PM Asset Mgmt → |
| PM — Home & Business Moves | ~35% Move-flow analytics, tenure cohorts, churn-by-tenure modelling |
~40% Tariff-landing analysis, partner-journey SLA review, retention reporting |
~25% Journey design ownership, partner accountability, customer-complaint sign-off |
PM Moves → |
| Head of Product Platform | ~10% | ~40% Technical strategy analysis, OKR dashboards, architecture review support |
~50% Platform strategy ownership, CPO/CTO alignment, cross-team accountability |
JD analysis |
| Product Marketing Manager | ~45% Content drafts, competitor analysis, pricing-page copy, release notes |
~35% Positioning strategy, go-to-market brief drafts |
~20% Brand ownership, exec-facing narrative, partner announcements |
JD analysis |
Residual team — Product
| Senior PMs per module | One accountable PM per Meridian module. Regulatory decisions and roadmap ownership cannot be automated — they require a named person to accept accountability. |
| Head of Product Platform | Platform strategy at the CTO boundary. Inherently human — cross-function, exec-facing. |
| Product Marketing Manager | Halved capacity needed once AI takes the content-generation load; one senior role owns the brand narrative. |
£ saved: 8–10 × £60k avg PM salary Nottingham = £480–600k/yr. Salary basis: Glassdoor/Reed median PM Nottingham 2025 £55–70k.
Customer Success
| Activity type | Verdict | Why |
|---|---|---|
| Account health monitoring, renewal-risk scoring, onboarding tracking | AI-takes | Deterministic signal → alert. No client judgement required for the detection. |
| QBR prep, usage reports, success-plan drafts, meeting notes | AI-assists | AI drafts; CSM owns the client narrative and accuracy sign-off. |
| Executive relationships, escalation handling, renewal negotiation, complex problem solving | human-only | Kestrel sells to energy retailers: these are enterprise relationships worth £M. Trust is personal. AI cannot carry commercial accountability at this level. |
Operations & Commercial
| Sub-function | Verdict mix | Human-only residual reason |
|---|---|---|
| Finance (reporting, forecasting, AP/AR reconciliation) | 20% takes / 45% assists / 35% human | Audit sign-off. Statutory accounts require a named signatory. Forecasting judgement stays human. |
| Legal (contract review, compliance monitoring) | 10% takes / 45% assists / 45% human | Regulated professional accountability. AI is a research/draft tool; the legal sign-off must be human. |
| People / HR (policy, onboarding, performance support) | 20% takes / 40% assists / 40% human | Employment decisions, disciplinary, culture — irreducibly human. High-sensitivity, relationship-critical. |
| Commercial (partnerships, supplier management) | 15% takes / 40% assists / 45% human | Commercial deals at Kestrel's scale (Helios Group, energy retailers) are senior relationship work. AI assists with analysis and draft terms; the deal is human. |
Marketing
| Activity | Verdict | Why |
|---|---|---|
| Content production (blogs, case studies, docs, SEO copy, social) | AI-takes | High-volume, spec-driven content generation. AI with human brief + edit loop handles the bulk. |
| Campaign analysis, competitor tracking, performance reporting | AI-takes | Data pull and pattern detection. Human interprets the so-what. |
| Messaging strategy, brand positioning, partner announcements | AI-assists | AI drafts; senior marketer owns the brand voice and exec-level accuracy. |
| C-suite comms, board narratives, investor announcements | human-only | Helios Group relationship visibility, regulatory sensitivity. These documents carry personal and corporate accountability. |
03 · Aggregate impact
What the numbers say — and where they come from.
These figures are built bottom-up from the per-function triage above. Ranges reflect the uncertainty in headcount assumptions. The proved segment (Technology + Product, 133 staff) is the high-confidence anchor. The remaining ~167 is extrapolated.
04 · Implementation roadmap
Three phases. Proof before scale.
The proved workspaces already exist. The roadmap is: institutionalise the proofs, expand to the full function set, and build the operating model for continuous improvement.
Prove & adopt
Deploy the 7 live workspaces into production. Establish governance. Measure baseline productivity.
Months 1–6Expand the coverage
Build workspaces for the remaining proved-segment roles. Integrate with live Meridian data. Extend to Customer Success.
Months 7–18Org transformation
Restructure headcount through natural attrition + role redesign. Build the AI-augmented operating model. £3M+ run-rate.
Months 19–36- Deploy 7 live workspaces from this proof into the tools used by the Technology Data team, QA, and the 5 Product modules covered.
- Establish the human-control layer: every AI output has a named human accountable for sign-off. Define which tasks can go straight to action vs. which need review.
- Baseline measurement: track task-level productivity before/after. Target: demonstrate 1.5× throughput on the automatable tasks within 90 days of deployment.
- Falsifier gate: if any workspace produces an output that can't be independently re-derived from live data, it's pulled and the vendor notified. No AI output goes to a customer or regulator without re-derivation.
- Outcome: £0 headcount savings in Phase 1 (no redundancies). Productivity gain harvested as capacity for higher-value work.
- Build workspaces for the remaining Kestrel Product modules: PM-API, PM-Identity & Access, PM-Energy Products Quoting — the three roles currently without workspaces.
- Extend to QA: integrate workspace with the real Meridian test environment. AI generates regression suites; QA lead reviews and accepts.
- Customer Success tooling: AI account-health monitoring, QBR prep assistant, renewal-risk dashboard. Human-in-loop for all client-facing communications.
- Software development tooling: standardise AI code generation, review assist, and test generation across all squads. Establish the AI-review accountability model (AI writes; senior engineer accepts).
- Finance/Legal: contract-review assist, recurring reporting AI. Human sign-off on all statutory documents.
- Outcome: 15–20% productivity uplift across the covered org. Backfill freeze on roles that become demonstrably AI-augmented.
- Headcount redesign: restructure through natural attrition. For every 2–3 departures in AI-augmented roles, hire 1 senior accountable lead rather than a direct replacement.
- Target residual: Technology ~70–75 (from 107), Product ~16–18 (from 26), other functions at similar 30–35% reduction. Total org target: ~195–210 FTE from 300 base.
- AI-ops team: 3–5 FTE specialised in prompt governance, workspace quality, and AI-output falsification. This is the team that replaces the headcount; budget from savings.
- Run-rate savings: £3–4M/yr from proved segment alone; £5–6M/yr if the full org reduction is achieved. Use a portion to fund AI-ops and tooling (est. £500k–800k/yr ongoing).
- Board metric: AI-augmented output per FTE (not headcount alone). A smaller team with measurably higher throughput is the success condition — not just a lower payroll.
05 · Methodology + falsifier notes
How this was built. What it isn't.
theorg.com/org/kestrel (accessed June 2026) for named C-suite + 6 Technology team sizes and total team count (43). LinkedIn profile + vacancy corpus for role-level structure inference. Named individuals are for org modelling only and reflect publicly available data.
JD decomposition: every Kestrel vacancy and the 3-year posting corpus, decomposed to task level. Each task classified by: (a) whether the output is checkable/deterministic → AI-takes; (b) whether human judgement is required on the output but AI produces the draft/analysis → AI-assists; (c) whether accountability, relationships, regulatory sign-off, novel architecture, or emotional intelligence is the primary value → human-only.
Proved — a live workspace exists, the AI actually performs the task in your browser over representative data, the SQL is visible and re-runnable, and the output is re-derived not stored. This is the only tier that supports a firm claim. Inferred — JD analysis + comparable proved roles; directional, no workspace. Extrapolated — statistical extrapolation from proved segment to functions with no sourced headcount or workspace; use as a range only.
Three buckets: human-only tasks → HC unchanged. AI-assists tasks → 60% of original HC needed to do the same work at the same quality (conservative; based on the assumption that AI handles the generation load and a human spends ~40% less time on review-and-correct than generation). AI-takes tasks → 10% oversight HC (1 FTE per ~10 automated FTE-tasks freed). These multipliers are stated assumptions; they are not sourced from an Kestrel internal study. The board should run an A/B on one team before applying them org-wide.
This is not a redundancy plan. The residual figures are a statement of what output-equivalent headcount looks like after AI augmentation at steady state — not a list of roles to eliminate. Realising the savings requires: (a) the AI actually doing the work reliably, (b) a managed transition through natural attrition and role redesign, and (c) investment in AI-ops and governance. None of that is free or instant.
The proof is clickable
Open a workspace. Watch the AI do the job.
Every green "proved" badge above links to a live workspace. The AI is actually running — not a pre-recorded reel. Click into any of the 7 roles and re-run the SQL yourself.