How we work
The operating model behind 'AI-driven, human-reviewed.'
Most consultancies are either silent about AI or claim 'AI-first' without explaining what's delegated. This page is the explicit version. What AI does, where humans gatekeep, and how we keep both honest.
The thesis
AI is good at the routine. Humans are good at judgment. We've structured our practice around both.
The default in consulting is to staff every engagement with whoever is on the bench. Our default is the opposite: a small team of senior engineers, augmented by AI on the work where AI shines.
AI handles the first 80% — the parts where the inputs are structured and the output has a known shape. Senior engineers own the final 20% — the architecture, the trade-offs, the review, and the relationship. The result is consultancy speed at consultancy quality, with a billing model that matches.
-
01
Pillar
AI does the routine work
Audits, IaC scaffolds, runbook drafts, alert summaries, post-mortem first cuts — anything where the inputs are structured and the output has a known shape. We use AI tools (Claude, Cursor, internal pipelines we've built) the way a senior engineer uses an IDE: to compress the time between a blank page and a working draft.
- → A two-week infrastructure audit produces its first findings draft in hours, not days.
- → Terraform module scaffolds for new services are generated from your existing patterns, not from a generic template library.
- → Runbook drafts are auto-generated from observability data, then validated against the live system.
- → Long incident logs distill into causation summaries before the post-mortem meeting starts.
-
02
Pillar
Senior engineers gatekeep every change
No AI-generated diff hits your production unread. The senior engineer named on the engagement reviews every PR, owns every architectural decision, and signs every deliverable. The AI moves fast; the human decides what ships.
- → Every PR carries a senior engineer's name in the approver field, not just an automation account.
- → Architectural trade-offs are documented — including the alternatives we rejected and why.
- → Plausible-but-wrong is the most expensive AI failure mode. We catch it before you do, by validating output against the real system.
- → When something breaks at 3 a.m., a human you've worked with answers — not a chatbot.
-
03
Pillar
We say what's AI-generated and what's human
When AI drafted something and a human approved it, we say so. When a human wrote it from scratch, we say that too. This matters more than it sounds: clients with AI governance requirements (a growing number) need to know which parts of a deliverable carry which kind of authorship.
- → Audit reports carry a methodology footer naming which sections were AI-drafted and human-reviewed.
- → Code review comments distinguish "AI suggested, I confirmed" from "I caught this manually".
- → We don't pretend AI didn't help — and we don't pretend AI did the whole thing.
-
04
Pillar
You pay for judgment, not for hours
Most consultancies bill for the time it takes to write things. We bill for the time it takes to decide what to write. The result: faster delivery, smaller teams, higher per-hour value. Total cost is usually 30–50% lower than a body-shop equivalent.
- → A discovery engagement that would take a typical consultancy 3-4 weeks lands in 2 with us.
- → Smaller engagement teams (2-3 senior engineers) replace 5-7 person body-shop teams without losing capacity.
- → You can read every PR. The trail of decisions is auditable and ours to defend.
Our commitments
Four guarantees we'll put in writing.
Every deliverable is human-signed.
A senior engineer puts their name on it. We are the accountable party — not the AI tool we used to draft it.
AI usage is disclosed.
We tell you what AI tools were used in your engagement and where they touched your data. If you have AI-governance constraints, we work to them.
Your data does not train models.
We don't fine-tune external models on your code, infrastructure config, or proprietary data. AI tools we use process data ephemerally only.
We will say no when AI shouldn't.
Where the work requires deep judgment, novel reasoning, or context AI can't have, we say so and bill the human time honestly.
The trade-off, named
When this model is the right call — and when it isn't.
Pick us when
- +You want consulting speed without sacrificing senior judgment
- +Your engineering team values AI tools and uses them themselves
- +You'd rather pay for one senior than four mid-levels
- +Your AI-governance posture allows commercial AI tools with disclosure
Pick someone else when
- −You need 20+ engineers on a single engagement (we cap at ~5)
- −AI tools are contractually prohibited and can't be negotiated
- −You need a partner with on-the-ground presence in a specific region we don't serve
- −You want a staffing arrangement (body-shop), not an outcome-driven engagement
The questions we get asked
Honest answers, including the awkward ones.
What AI tools do you actually use? +
Day-to-day: Claude (Anthropic), Cursor, GitHub Copilot, plus internal tooling we've built for infrastructure-specific tasks. We document the stack used in each engagement during onboarding.
How do I know AI didn't ship something subtly wrong? +
Two answers. (a) Every change goes through a senior engineer review before merge. (b) AI-generated code is annotated as such in PR descriptions, so reviewers know what to scrutinize. The discipline is the safeguard, not blind trust.
Is this faster, cheaper, or both? +
Both, usually. Engagements that would have taken 3-4 weeks at a typical consultancy land in 2. Total cost is typically 30-50% lower than a body-shop equivalent of the same scope.
What about quality? Doesn't AI introduce subtle bugs? +
AI introduces a different distribution of bugs than humans do — confidently-wrong assertions, hallucinated APIs, plausible-looking but incorrect logic. The senior-review layer is specifically designed to catch this distribution. Net quality, in our experience, is higher than a body shop where junior engineers do the same routine work without senior review.
Will you sign a contract that prohibits AI tool use? +
Yes — we have done this for clients with strict AI-governance constraints. The work takes longer and costs more, and we're explicit about the trade-off. Most clients prefer the AI-assisted model with disclosure.
Can you work in air-gapped or on-prem environments? +
Yes, with adapted tooling. We've shipped engagements where cloud AI tools were not permitted; we use locally-hosted models or pure-human delivery in those cases.
How do you handle client data privacy? +
Standard practice: data classification on day one, redaction pipelines for anything sensitive before it touches an AI tool, no fine-tuning on client data, and contractual data-handling commitments. Privacy posture is part of our standard MSA.
Do other CloudWizz engineers see my code? +
Only the engineers assigned to your engagement. Cross-engagement learning happens at the pattern level — 'we saw this auth bug shape three times' — not at the code level. We do not pool client codebases.
Want to see this operating model run on a real engagement?
A 30-minute call. We'll walk through how we'd approach your specific work — and what AI vs. human would do.
Book a 30-min call →