Gemini Enterprise Go-live in 30 Days: Real Playbook (Not the Sales Pitch)
30 days is the deadline we promise for the first agent in production. Here's week by week what happens — including what goes wrong and how we fix it.
Fabiano Brito
CEO & Founder
right scope
no buffer
non-negotiable
The 30 days start at kickoff — not at contract signing, not at license activation, not at the first meeting. Confusing those dates is the first error that wrecks the forecast. Agree internally: D0 is the day stakeholders, data, and the Google Cloud environment are ready.
Pre-requisites before D0
sa-east1. 3. Named executive sponsor (business director, not IT). 4. DPO available for 4 hours of DPIA discussion in week 2. 5. Use case chosen with 1 measurable KPI (not "productivity"). We've seen projects fail for lack of quota — which takes 5-15 days to release with Google.
Week 1: Deep discovery
Days 1-2: Immersion workshop
4 hours with the business area. We map:
- Current process step by step (with timings).
- Systems involved (ERP, CRM, spreadsheet — all of them).
- Decision rules (written and tacit).
- Current and desired volume.
- Quantitative success criteria.
Days 3-4: Technical inventory
We survey available APIs, auth, latency, rate limits. Watch out: 60% of projects discover here that "the ERP has an API" means "a 2009 SOAP without documentation". That finding may change scope — better now than on day 25.
Day 5: Adjusted Go/No-Go
Refined scope presented to the sponsor. If it changed too much, we replan before promising 30 days.
Week 2: Architecture and governance
Days 6-7: Technical design
We define:
- Model (Gemini 2.5 Pro vs Flash according to cost/latency).
- Required tools (with fallback for each).
- RAG strategy (if applicable).
- Prompt + guardrails layer.
- Human hand-off (when the agent stops deciding).
Days 8-9: DPIA and security
Session with DPO. We fill the DPIA together. Configuration: training opt-out, sa-east1 residency, DLP, ACL, audit log. Detailed in DPIA for Gemini Enterprise projects.
Day 10: Navigable prototype
Working version with synthetic data. Sponsor approves or redirects — cheap to adjust now.
Week 3: Agent build
Days 11-13: Implementation
Real build:
- Connectors configured in Vertex AI Agent Builder.
- Prompts iterated against 20-30 gold-set cases.
- Tools with retry, timeout and error handling.
- Structured logs from the first line of code.
Days 14-15: Automated evaluation
Gold set of 50-100 cases with expected answers. We run it against every prompt change. Metrics: relevance@1, faithfulness, completeness. Without this, you're iterating in the dark.
Week 4: Pilot and go-live
Days 16-18: Closed pilot
5-10 real users on real cases. Daily feedback. Adjustment backlog. Golden rule: only fix what shows up in 2+ users — isolated cases go to "sprint 2".
Days 19-22: Hardening
Load, simulated failures, edge cases. Operational documentation, runbooks, alerts. Training the team that will operate (not just use).
Days 23-25: Audience training
90-minute session with full audience. Hands-on. Recorded material for those who can't attend.
Days 26-28: Soft launch
Full audience, but with hot support channel. Daily monitoring of target metrics.
Days 29-30: Delivery ceremony + evolution plan
Results presentation to sponsor. Plan for the next 60 days (optimization, scaling, next agent).
What goes wrong (and how we avoid it)
1. Weak sponsorship
If the sponsor is a manager, not a director, decisions stall in W2/W3. Mitigation: we require named C-level or director sponsorship with budget autonomy.
2. Dirty data
Unstandardized documents, duplicates in the base, inconsistent ACL. Mitigation: quality assessment in week 1, with remediation plan or scope adjustment.
3. Unstable tools
ERP API drops 4x/day. Mitigation: circuit breaker in the agent + clear manual path.
4. Scope change in week 3
Sponsor discovers a new idea. Mitigation: formal scope freeze on day 10, with a "backlog v2" channel.
5. Team resistance
Operations sees AI as a threat. Mitigation: involvement from week 1, focus on "more interesting work", not "fewer people".
When 30 days isn't enough
- More than 3 legacy systems without API.
- Automated decision with legal effect (needs legal validation + committee).
- Sensitive data without prior DLP configuration.
- Heavy sector compliance (BACEN, ANS, ANEEL) without prior governance work.
In those cases we promise 60-90 days — and we deliver. The point is not pretending 30 will do.
The worst 30-day project is the one that promises without the 5 pre-requisites. Better to say "this will need 60" than to pretend and blow up in week 3.
Does your case fit in 30 days? A free diagnostic answers that.
30 minutes: scope mapping, pre-requisites, KPI and feasibility. We leave with a week-by-week plan and value estimate. If it doesn't fit in 30, we honestly promise 60-90.