No Black Box. Here Is Exactly What Happens.
One of the most common questions we get from prospective clients is: "What does working with you actually look like?" It is a fair question. AI consulting can feel like a black box -- you hand over money and vaguely hope something useful comes back. That is not how we operate.
This is our actual process, week by week, for setting up a new client's first AI workflow. We are sharing it because transparency builds trust, and because we think the process itself is one of the things that separates a good AI engagement from a bad one. For the higher-level philosophy behind our approach, see the AI consulting playbook. We have refined this over dozens of engagements, learning from every mistake along the way.
Fair warning: this is not a highlight reel. We include the parts where things typically go sideways, because that is where the real value of experience shows up.
The difference between a successful AI engagement and a failed one is almost never the technology. It is the process around it.
Week 1: Discovery
The first week is entirely about understanding your business. We do not touch any technology. We do not write any code. We listen.
What we look at
- Current tools and systems: What software does your team use daily? CRM, project management, communication, accounting, industry-specific tools. We map out the entire ecosystem because AI needs to fit into it, not replace it.
- Data flow: How does information move through your business? Where does it originate, who touches it, where does it end up? This is where we find the bottlenecks that AI can actually fix.
- Pain points: What tasks consume the most time relative to their value? What falls through the cracks most often? What do your best people wish they did not have to do?
- Team dynamics: Who is excited about AI? Who is skeptical? Who will be the day-to-day user? Understanding the human side is just as important as the technical side. An AI workflow that your team hates is a failed project regardless of how elegant the technology is.
What questions we ask
We conduct structured interviews with key team members -- typically the business owner, operations lead, and two to three people who do the actual work the AI will affect. Our standard discovery questionnaire covers about 40 questions, but the ones that matter most are:
- Walk me through the last time this task took longer than it should have.
- What would you do with an extra 10 hours per week?
- What is the most error-prone part of your current process?
- If one thing could just "happen automatically," what would it be?
- What have you already tried that did not work?
How we assess current tools
We audit your existing tech stack for API availability, data export capabilities, and integration options. This matters because the easiest, most cost-effective AI implementations connect to what you already have. If your CRM does not have an API, that changes the approach entirely. We use tools like MCP to map integration possibilities -- what can talk to what, and how.
Deliverable at end of Week 1
A discovery document that includes: a map of your current systems, a ranked list of AI opportunities with estimated impact and complexity, and our recommendation for which workflow to tackle first. This document is yours regardless of whether you continue working with us.
Week 2: Strategy
Week two is where we translate discovery into a concrete plan. This is the most collaborative week -- we need your input and decisions.
Prioritizing opportunities
We score each opportunity on three dimensions: business impact (how much time or money it saves), technical feasibility (how straightforward the implementation is), and team readiness (how likely your team is to adopt it). The sweet spot is high impact, high feasibility, high readiness. That is always where we start.
Common first workflows we implement: automated document processing, meeting summary and action item extraction, customer inquiry response, internal knowledge base search, and report generation. These are proven patterns with predictable results.
Defining success metrics
Before we build anything, we agree on exactly how we will measure success. These are specific numbers, not vague aspirations:
- Time saved per task (measured in minutes, not percentages).
- Error rate reduction (compared to a baseline we establish in Week 1).
- User adoption rate (what percentage of the team uses it regularly by Day 30).
- Cost per transaction (how much each AI interaction costs at projected volume).
We document these in a one-page success criteria sheet that both parties sign off on. This prevents the scope creep and goal-post shifting that kills most AI projects.
Choosing the right approach
Based on the specific workflow, we decide on the technical approach:
- Prompt engineering: For straightforward tasks where the AI just needs good instructions. This is the fastest and cheapest approach.
- RAG (Retrieval-Augmented Generation): For tasks that require your company-specific knowledge -- product catalogs, SOPs, historical data. We use Supabase for the vector database because it is reliable, affordable, and integrates well with everything else.
- AI agent with tools: For multi-step workflows where the AI needs to access external systems, make decisions, and take actions. This is the most powerful approach but also the most complex.
Deliverable at end of Week 2
An implementation plan: specific workflow to be automated, technical approach, integration points, success metrics, timeline for Weeks 3-4, and a cost projection for ongoing operation.
Weeks 3-4: Build and Integrate
This is where the actual building happens. We work fast because momentum matters -- the longer a project takes, the more likely it is to stall.
Setting up the first AI workflow
We build using Claude as the primary model for most business tasks. The development happens in Claude Code, which lets us iterate rapidly -- building, testing, and refining in tight cycles. A typical first workflow goes through three to five iterations before it performs at a level we are comfortable putting in front of your team.
The first iteration is intentionally rough. We get something working quickly and then test it against real data from your business. This is where most issues surface -- edge cases, unexpected data formats, workflows that are more nuanced than they appeared in discovery. We would rather find these in Week 3 than Week 8.
Connecting to existing tools
Integration is where projects commonly get stuck. We use MCP (Model Context Protocol)as our standard integration layer. MCP gives us a universal way to connect the AI to your CRM, email, project management tools, databases, and file storage without building fragile custom integrations for each one.
For data storage and retrieval, we deploy on Supabase -- it gives us a production-ready database with vector search capabilities (for RAG), edge functions (for webhooks and triggers), and authentication. The infrastructure is yours. We do not host your data on our servers.
Common integrations we set up in the first 30 days: CRM read/write, email send/receive, document storage, calendar access, and communication tools like Slack. Each one is an MCP server that the AI agent can access as needed.
Testing
We test with real data, not synthetic examples. Your actual documents, your actual customer inquiries, your actual reports. We run the AI workflow alongside your current manual process for at least 3-5 days, comparing outputs. This parallel testing period is non-negotiable -- it is how we catch errors before they reach anyone who matters.
What typically goes wrong
Transparency time. Here are the most common issues we encounter during build and how we handle them:
- Data quality surprises. Your CRM has duplicate records. Your documents use inconsistent formatting. Your data is messier than anyone realized. We budget time for data cleanup in every engagement because this happens in about 80% of projects.
- Scope creep from stakeholders. Someone sees the prototype and says "can it also do X?" Yes, probably, but not in this sprint. We maintain strict scope discipline in the first 30 days. X goes on the backlog for the next phase.
- Integration limitations. The CRM API does not support a feature we need. The email system has rate limits. A vendor's webhook implementation is unreliable. We design around these constraints rather than fighting them.
- AI accuracy on edge cases. The AI handles 90% of inputs perfectly and stumbles on the weird 10%. We build confidence thresholds and escalation paths so that uncertain cases get routed to a human instead of producing bad outputs silently.
End of Week 4: Train and Handoff
The final phase is about making sure your team can actually use what we built and that it does not fall apart the day we step back.
Team training
We conduct hands-on training sessions -- not lectures, actual working sessions where your team uses the AI workflow on real tasks while we watch and coach. Typically we run two sessions: one for the primary users and one for the manager or owner who needs to understand the reporting and oversight layer.
Training covers: how to use the workflow, how to review AI outputs for accuracy, when to override or escalate, and how to provide feedback that improves performance over time.
Documentation
Every workflow we build comes with documentation that a non-technical person can follow: step-by-step usage guide, troubleshooting for common issues, escalation procedures, and a reference card for the prompts and configurations we set up. This lives in a shared document your team owns -- essentially the beginning of what we describe in building an AI knowledge base for your team.
What "done" looks like
A project is done when all five of these conditions are met:
- The AI workflow is running on your infrastructure, not ours.
- At least two people on your team can operate it independently.
- The success metrics from Week 2 are being tracked and are trending positive.
- Documentation is complete and accessible.
- There is a clear plan for the next 30 days of optimization.
What Happens After Day 30
Day 30 is not the end. It is the end of the beginning.
Ongoing optimization (Days 31-60)
We stay engaged for the first month after handoff, monitoring performance, tuning prompts based on real-world usage patterns, and addressing issues that only surface at production volume. Most workflows improve 20-30% in accuracy and speed during this optimization phase as we refine the system based on actual usage data. For a realistic picture of what these gains translate to financially, see the ROI of AI consulting.
Expanding to more workflows
Once the first workflow is stable and delivering measurable value, we start the cycle again with the next item on the priority list from Week 2. The second workflow is always faster -- typically two weeks instead of four -- because the infrastructure, integrations, and team familiarity are already in place.
Building internal capability
Our long-term goal for every client is that they need us less over time, not more. We progressively hand off more responsibility to your team, train power users to create simple workflows on their own, and transition from implementation partner to occasional advisor. A good AI consultant works toward their own obsolescence in any given engagement.
That is the process. No magic. No black box. Just structured, disciplined work with a technology that, when implemented thoughtfully, produces real, measurable results. The difference between a successful AI engagement and a failed one is almost never the technology. It is the process around it.
A good AI consultant works toward their own obsolescence in any given engagement. The goal is not permanent dependency -- it is building your team's capability.