The process
Problem to deployed AI
in 7 days.
Four steps. One week. Here's exactly what happens at each stage — and why the diagnosis we do upfront is what makes the build worth it.
The core problem
“Most AI projects are well-executed answers to the wrong question.”
Why this happens
The problem isn't AI capability. The models are powerful. The problem is framing. When an AI project starts with a solution someone saw at a conference rather than a diagnosed operational problem, everything that follows is optimised for the wrong outcome.
The strategy layer — the thinking that happens before any code — is what most SMBs don't have. They have tools. They have vendors. They don't have someone who can sit down, understand their operations, and say "here's what will actually work." That's what we do first, every engagement.
01
Step 01
Day 0
Your Problem
We don't need a spec. We don't need a deck. We need one sentence — however rough, however incomplete. 'Our team wastes 3 hours a day on X' is enough. A voice note is enough. A half-formed frustration is enough.
The goal of day zero is to capture the problem before the internal edit starts. Most bottlenecks get polished before they're submitted, which strips out the context that actually matters — the frustration, the specific workflow, the moment the cost became obvious.
what we're actually doing here
- Describe the bottleneck before rationalizing a solution
- Identify who in the business is affected and how often
- Note what's been tried before (if anything)
- Resist jumping to solutions — the stated solution is rarely the right one
you end up automating a workaround instead of fixing the actual problem.
02
Step 02
Within 24 hours
AI Diagnosis
The diagnosis comes first. Everything else is execution of the diagnosis. We deliver it within 24 hours because speed matters here — not to be fast, but because a diagnosis written close to the problem is more honest than one written after a week of stakeholder management.
The diagnosis has four components: the real problem (which is often different from the stated problem), the AI approach most likely to work given your constraints, explicit scope exclusions with reasons, and the one metric that will tell us if it worked.
what we're actually doing here
- Reframe the bottleneck from operational impact, not technical framing
- Identify the right AI approach for this specific problem
- List everything explicitly out of scope — and why
- Name the single success metric
you build a technically correct AI that solves the wrong problem.
03
Step 03
Days 1–5
Build
We build one working agent or automation, end to end. Not a mockup. Not a slide deck. A deployed thing your team can actually use. Something real that touches your actual data and workflows.
AI-assisted development means this is genuinely achievable in 3–5 days for most operational problems. The constraint isn't time — it's scope discipline. We build exactly what the diagnosis specifies, nothing more. Everything else is a conversation for after it's working.
what we're actually doing here
- Select the single workflow that delivers the most operational impact
- Build against real data and real systems from day one
- Deploy to a live environment daily — no staging theatre
- Document every cut decision made during build
you've built something technically impressive that nobody in operations actually uses.
04
Step 04
Day 7
Decision
Three paths, pre-priced, no pressure. You've seen the agent work in your environment. Now you decide what to do with it.
Path one: go to production. We scope and price the full deployment based on what you saw. Path two: take the code. Full handoff, documented, deployable. You own it. Path three: retain us. Keep us as your ongoing AI team — we maintain, improve, and add agents as your needs evolve.
what we're actually doing here
- Review the agent against the original success metric
- Evaluate what changed in your operational assumptions
- Choose a path with full cost transparency
- No lock-in, no pressure — every path is pre-priced
you've spent $50K on an agency build of something that a $1,200 sprint would have validated first.
Sample diagnosis (illustrative)
What an AI diagnosis looks like.
REAL PROBLEM
New client documents aren't the bottleneck. Classification is. Paralegals spend ~3 hours per new matter manually reading and sorting email attachments before legal work can start. 80% of documents fall into 6 repeatable categories.
HYPOTHESIS BEING TESTED
A Claude classifier reading attachment text will route documents into the 6 standard categories with higher accuracy than manual review, at a fraction of the time.
IN SCOPE
- Email attachment intake (PDF and Word)
- Claude-based document classification into 6 categories
- Structured summary delivered to paralegal inbox
- Manual review and override for low-confidence items
EXPLICITLY OUT OF SCOPE (and why)
- Client portal — change management too heavy for a hypothesis test; add in phase 2
- OCR for handwritten documents — edge case, marginal return in v1
- Case management system integration — phase 2 if classification accuracy holds
SUCCESS METRIC
New-matter document sorting drops from 3 hours to under 30 minutes per matter.