10 days on-site → 6–8 AI projects + 12-month roadmap targeting 30% profitability uplift

The problem:

Core parts of the business still relied on:

• •relatively straightforward bidding/pricing logic
• •human-heavy customer support and risk processes
• •manual partner (hotel) communications and supply quality improvements

Meanwhile, competitive pressure demanded faster iteration and better unit economics. The client knew AI could help — but lacked clarity on:

• •what to prioritize next
• •what would produce the biggest profit impact
• •how to structure an AI stack and governance to ship reliably

Their current approach:

• •teams tried tools and prototypes, but projects stalled before production
• •no consistent prioritization framework across departments
• •no shared governance/cadence to drive adoption
• •unclear skill gaps and ownership, so initiatives drifted

The consequences:

• •critical initiatives delayed despite strategic urgency
• •duplicated effort (or fear of duplication), slowing decisions
• •fragmented pilots that didn't compound into a platform advantage

The goal wasn't an "AI strategy deck." The goal was an execution-ready Business Case + roadmap, grounded in real operational data, with sequenced projects that build on each other.

1. 10-day on-site audit: interviews + shadowing across core functions

Over 10 days, we:

• •ran interviews with leads and operators across Supply, Risk, Customer Support, Dev, Finance, and Operations
• •shadowed day-to-day workflows to capture handoffs, exceptions, and bottlenecks
• •mapped end-to-end processes into a single workflow log (the "source of truth")

Output: 40 workflows documented with owners, systems, cycle time drivers, failure modes, and measurable pain.

2. Process mining with real data: Zendesk ticket analysis

To avoid assumptions, we pulled and analyzed Zendesk tickets, focusing on:

• •time to first response and time to resolution
• •triage patterns (where tickets get stuck, bounced, or escalated)
• •common categories and repetitive "high-volume" request types
• •where human effort was being spent that could be automated or assisted

This gave us a factual baseline for cost-to-serve and bottleneck severity — not opinions.

3. Business Case model: quantify bottlenecks + estimate solution uplift

We converted the workflow log into a prioritization and business case model using:

• •Reach (volume/frequency, % of operations touched)
• •Impact (margin, conversion, cost-to-serve, speed)
• •Uplift estimate per solution (what improves, by how much, and why)
• •Effort & dependencies (data readiness, integration complexity, change management)
• •Risk & governance needs (human-in-the-loop, evaluation, monitoring)

This enabled objective prioritization and leadership alignment: what to do first, second, and third.

4. Roadmap design: 6–8 projects that stack, not scatter

Instead of a long wishlist, we consolidated everything into 6–8 AI projects designed to compound:

• •shared data foundations and reusable components
• •a consistent agent layer (instead of one-off bots)
• •clear sequencing so later projects become cheaper and faster

This is what made the roadmap executable and prevented duplication.

5. Build vs buy vs hybrid: vetting vendors, partners, and in-house options

For each project theme, we reviewed and validated:

• •what should be built in-house (competitive edge, deep integration)
• •what could be handled by third-party tools (speed, commodity features)
• •where a hybrid approach made sense (platform + vendors)

We also met and evaluated external partners to pressure-test feasibility and tradeoffs.

Roadmap themes we prioritized