M/HQ Group · Stage 2 · CrossVal

Running costs: a working note

What the owned core costs to run, expressed as a unit price per entity, with every figure derived from a stated assumption

Conceptual ranges, not quotes · validated against actual usage at WS0
Prepared by CrossVal · July 2026 · Confidential
Currency: AED, converted at 3.6725 where a source price is USD

1How to read this note

Everything here is priced per entity per year, then multiplied by today's ~1,200 entities. That makes the model transparent (change an assumption, the answer moves visibly), it scales as the Group grows, and it answers the real question: what does it cost to run one client entity on the new architecture?

Three rules. First, these are ranges from published list prices, not vendor quotes; hyperscalers do not commit to exact figures in advance. Second, no range survives contact with reality unchallenged: the WS0 baseline measures actual document counts and sizes, and this note gets re-cut against them. Third, the Stage 2 build fee is excluded throughout; this note covers what it costs to run what Stage 2 builds.

Summary

Running the owned core costs on the order of AED 25 to 140 per entity per year all-in (hosting, storage, AI usage, tenancy), which is roughly AED 30,000 to 165,000 a year at today's entity count. The single biggest driver is the fixed platform base, which falls per entity as the Group grows. The AI inference bill is smaller than most people expect; the storage bill is a rounding error. The one line that can outgrow all of these is per-seat AI assistant licences, which is why the platform decision in WS0 matters.

2The unit model: assumptions per entity

AssumptionValue usedBasis
Entities under management~1,200Group-supplied data (Stage 1)
Documents held per entity (historic)50–150Working assumption; measured at WS0
Average document size~2 MBScanned passports, licences, filings
Documents processed per entity per year (go-forward)12–36Derived from ~60 M/HQ cases/month plus BAU flow across the book
AI-drafted outputs per entity per year5–15File notes, minutes, renewals, returns contributions
Tokens per document extracted~3,0002–4 pages of OCR text plus structured output
Tokens per drafted output~5,000Input context plus generated draft
Storage multiplier (backup + non-production)×2.5Dubai–Abu Dhabi backup plus sandbox/UAT copies
Regional hosting premium+10–30%UAE/regional data centres vs US/EU list

3Unit prices, derived

3.1 Storage: a rounding error

100–300 MB per entity, ×2.5 for backup and non-production, is 0.25–0.75 GB effective. At object-storage list prices (~USD 0.023–0.025 per GB-month, regional premium included) that is under AED 1 per entity per year: AED 350–1,000 a year for the whole fleet. Storage will never be the argument for or against anything in this architecture.

3.2 AI usage: smaller than feared

12–36 documents extracted plus 5–15 drafts per entity per year is roughly 60,000–180,000 tokens. Production-grade frontier models are published at roughly USD 1–5 per million input tokens and USD 5–25 per million output tokens across the major providers (the assistant platform is selected during Stage 2, so we price the class, not a vendor). That derives to AED 0.5 to 8 per entity per year: AED 600–9,600 a year fleet-wide. Usage-based AI is cheap at this document volume; the expensive version of AI is per-seat licensing (Section 5).

3.3 Platform base: the dominant line

The fixed cost of running the core: application containers, a managed document database (managed clusters of the class this workload needs are published at roughly USD 60–600 per month, with fuller replica sets running to ~USD 1,200), backups, and the sandbox environment, with the regional premium applied. A defensible planning shape is USD 500–2,500 per month all-in, which is AED 22,000–110,000 a year, or AED 18–92 per entity per year at today's count. This line is fixed, so it is the one that improves with growth: at 2,000 entities the same platform is ~AED 11–55 per entity.

3.4 Tenancy: a modest uplift

The R&C and audit tenants share infrastructure with logical separation and firewalls, so they add roughly 20–40% to the platform and storage lines, not three times everything: AED 4–37 per entity per year.

3.5 One-time: the backlog conversion

Converting the historic corpus (50–150 documents per entity through OCR extraction) is a bounded one-off: roughly 150,000–450,000 tokens per entity, deriving to AED 1–11 per entity, once: AED 1,200–13,000 for the full book. A natural capex candidate, and small enough that the conversion decision is about data quality, not cost.

4What it adds up to

LinePer entity / yr (AED)Fleet / yr at ~1,200 (AED)CadenceTreatment
Platform base (hosting, database, backup, sandbox)18–9222,000–110,000RecurringOpex
Tenancy uplift (R&C + audit tenants)4–374,400–44,000RecurringOpex
Storage (incl. backup and non-prod)<1350–1,000RecurringOpex
AI usage (extraction + drafting)0.5–8600–9,600RecurringOpex
Recurring total~25–140~30,000–165,000AnnualOpex
Backlog conversion (historic corpus)1–111,200–13,000One-timeCapex candidate
Environment setup, data model, integrations buildInside the Stage 2 fee; not repeated hereOne-timeCapex candidate

Excluded by design: the Stage 2 build fee (in the proposal); licences for the entity-management, screening and accounting spokes (client-contracted, confirmed at the WS0 selection); and existing tool subscriptions the Group already pays.

5The line to watch: AI assistant seats

The infrastructure above is priced per entity. Assistant licences are priced per user, and published per-seat pricing for enterprise AI assistants sits around USD 30 per user per month (~AED 110) at list. At 50 seats that is ~AED 66,000 a year; at 100 seats, ~AED 132,000 a year: potentially larger than the entire infrastructure bill. Two routes exist, and the prompt-library work in Stage 2 keeps both open: per-seat licences (predictable, simple), or API usage-based access through the owned layer (costs scale with actual use, which Section 3.2 suggests is modest). The right answer depends on how many people need the assistant daily versus occasionally, which the adoption metrics from the change programme will show. This is a decision for after the prompt libraries are proven, made on measured usage rather than upfront.

6The continuing arrangement: indicative ranges

This section prices the second of the two prices in the proposal (Section 14): the month-to-month building and improvement of the owned asset after the end-of-February close. Everything here is derived from the Stage 2 rate card (blended ~AED 950/hr; engagement principal AED 1,645/hr), so these are arithmetic, not new pricing. Final terms at contracting.

The arrangement has a small fixed base (a part-time embedded lead) and a variable improvement pool the Group turns up or down. Most of it is building work rather than support: the point is to keep adding to what the Group owns.

ElementShapeIndicative range (AED)Basis
Embedded lead (base)Monthly retainer28,000–38,000 / monthA senior seat drawn from the Stage 2 build team, ~30–40 hours a month at the blended rate: runs the internal team's cadence, triages the improvement backlog, and steers any mini-projects, carrying the data-architecture handover toward Geetu
Improvement pool (on demand)Blocks of hours, monthly38,000–76,000 / month40–80 hours a month building new automations and efficiencies on the owned record; drawn only as commissioned
HypercareFirst 60 days after close55,000–95,000 totalRapid-response cover, ~60–100 hours across the period
Senior advisoryMonthly or quarterly6,500–13,000 / sessionPrincipal half-day to full day
Quarterly technology reviewQuarterly23,000–38,000 / quarter3–5 days: estate audit against the measurement baseline
Prompt-library upkeepQuarterly15,000–30,000 / quarter2–4 days: re-certification and new patterns folded in
The cost over a year

The only committed recurring cost is the base seat: AED 28,000–38,000 a month, or AED 336,000–456,000 if it runs a full year. Building is elective on top: a month with a modest improvement block runs roughly AED 50,000–75,000, and a full year of continuous building at that rate would be AED 600,000–900,000 against the one-time build fee of AED 1.27M. The arrangement is month to month: it can be thinned, paused or stopped at any month-end, so none of these annual figures is a commitment; the sensible review point is November or December 2026, once the pace of useful improvement is clear.

Treatment and staffing

Because the building work creates enduring, Group-owned assets, it is structured to support capitalisation as asset-building; the run and licence lines in Sections 3 to 5 are the opex portion. The intention is not for the Group to hire someone senior to run this: CrossVal carries that capability through the embedded seat, and any Group hire is a more junior role brought up through the super-user track.

7Basis and sources

  • Object storage: published AWS S3 list pricing (~USD 0.023–0.025/GB-month), with UAE-region premium of ~10% applied; regional providers (G42/Core42, Moro, Khazna) to be quoted at WS0.
  • Managed database: published MongoDB Atlas dedicated-cluster pricing (M-series, ~USD 60–1,200/month depending on tier and replication).
  • AI models: published per-token list prices at major providers, USD 1–5 per million input and 5–25 per million output tokens for production-grade models (the platform-class range; vendor selected in Stage 2).
  • Assistant seats: published enterprise per-seat pricing, circa USD 30/user/month.
  • Volumes: Stage 1 discovery and the Group's own figures (~1,200 entities, ~60 M/HQ cases/month); document counts and sizes are working assumptions until the WS0 baseline measures them.

End of note. Every range here re-cuts automatically when WS0 replaces an assumption with a measurement. Prepared by CrossVal for the M/HQ Group.