# Guava > Guava is founder-led AI enablement for owner-led and mid-market US firms of roughly $5M–$50M in revenue: MSPs, IT services and consulting firms, and tech staffing firms. Based in San Francisco, CA. Guava gets the client's own team using and delivering AI on Microsoft 365 Copilot, Claude Enterprise, and ChatGPT Enterprise: it assesses, rolls out, coaches, and optimizes until the client's team runs it. Romeo Diaz does the work personally, from scope through handover. Every engagement has a fixed fee agreed in writing before work starts, and the client owns everything built. Guava is independent and multi-platform. It does not resell AI licenses, hold vendor partnerships, or take referral fees, and it recommends honestly which platform fits each team. Gemini for Google Workspace is supported for Google Workspace clients, with the same process. Custom builds run on Azure AI Foundry, Amazon Bedrock, or Google Vertex AI, usually whichever cloud the client already uses. Guava does not white-label or deliver work for a client's customers: for MSPs, it coaches the MSP's own team to deliver AI services, going from managed service provider to Managed Intelligence Provider (MIP). Larger firms ($50M–$150M) and other owner-led businesses are welcome to start with an AI Assessment. Book a 20-minute call: https://cal.com/romeodiazsf/20min Email: romeo@workwithguava.com ## Engagements - AI Assessment ($4,000, 2 weeks, paid upfront and credited in full toward an AI Rollout or Implementation & Offer Launch): a review of workflows and data, a ranked list of where AI saves the team the most time, estimated costs, an honest recommendation of which platform fits each team (Microsoft 365 Copilot, Claude Enterprise, ChatGPT Enterprise, or Gemini), and a clear yes or no on what is worth building. Recommended starting point. Prospects unsure about the fee can try a free lite version first: set up guava-stack (https://workwithguava.com/guava-stack) and say "review my workflows" anytime; their own Claude or ChatGPT makes the report on their computer, with no email or sign-up with Guava. - AI Rollout ($25,000, 3–6 weeks; $21,000 after the Assessment credit), including data readiness on any platform (Copilot Readiness on Microsoft 365): roll out Copilot, Claude Enterprise, or ChatGPT Enterprise with the client's data ready first, either as a new rollout or for licenses the client already pays for but barely uses. It runs in phases: Ready (on Microsoft 365, Copilot Readiness: a Microsoft 365 tenant review, a SharePoint and OneDrive oversharing and permissions cleanup plan, sensitivity labels and data protection basics, and pilot group selection), Activate (3–5 use cases with measured baselines, configuration and connectors), and Adopt (team training and adoption tracking). Optimize and Govern continue in Ongoing Optimization & Governance. The AI Assessment is not required first: a client who already knows what it wants can start with the Rollout, whose Ready phase then also picks the first 3–5 use cases. After an Assessment, Ready starts from the Assessment's ranked list, so nothing is assessed twice. The outcome is hours back per tech per week, measured in the client's PSA. - Implementation & Offer Launch ($35,000, 3 months; $31,000 after the Assessment credit): MSP to MIP in 90 days. Working AI use cases inside the client's systems, tested against criteria agreed up front, with custom builds on Azure AI Foundry, Amazon Bedrock, or Vertex AI where needed; the client's team learns to deliver AI services across Copilot, Claude, and ChatGPT; the client's AI services are packaged and priced for its own clients; by the end, the client's team runs it without Guava. - Ongoing Optimization & Governance ($4,000 per month, optional, cancel anytime, no minimum commitment): adoption and optimization, workflow operations, and AI governance and security: an AI acceptable-use policy, an inventory of the AI tools, connectors, and agents in use (shadow AI), access and connector reviews (Claude and ChatGPT connectors, Copilot agents, admin rights), a quarterly governance review, and incident guidance. Guava coaches the client's team to run governance itself and, for MSPs, to offer it to their own clients. ## What an AI Assessment delivers A ranked list of the client's workflows, each with estimated hours back per week, an estimated monthly AI tool cost, and a verdict: Build, Pilot, or Don't build, with the reasoning written down for every "Don't build". It ends with a recommended next step. The site shows sample pages for a fictional MSP, IT services firm, and tech staffing firm; those numbers are illustrative, not client results. ## Track record As Chief AI Officer at an SF-based tech staffing company (the client appears only under that description), from February to September 2026, Romeo helped it sell AI services to the clients it already staffed; until then, all its revenue came from placements. He hired and trained 6 full-time engineers, each delivering 3–4 projects, and set up the offer tiers (an Assessment, then an implementation, then ongoing support), the funnel, website and marketing, bi-weekly client hackathons, and a managed services team for live projects. Almost every client started with an Assessment; 2 in 3 of those became larger implementations, and 80% of the implementations stayed on for ongoing support. In production for its clients: a sales prep agent for a large tech company, a Salesforce and finance agent for a social media company, a chatbot over a finance firm's Snowflake and SharePoint data, assistant agents in Slack, Teams and WhatsApp, client onboarding automation for a wealth manager, an LLM gateway for a large accounting and audit firm, and code-review agents in GitHub Enterprise for a large networking company. A private equity firm estimated its acquisition price at 2.4x the estimate from the year before he joined; Romeo credits more recurring revenue, more revenue per employee, lower churn, and less key-person risk. ## Process 1. Scope: written before signature, covering the workflow, the measure of success, the fixed fee, an estimate of ongoing AI tool and usage costs, and the end date. 2. Access: NDAs are signed; only the access needed is requested, and everything touched is logged. Access can be revoked or given an expiry at any time. 3. Build: work happens inside the client's systems and tenant with weekly checkpoints. 4. Acceptance: the client runs the work on real tasks from their own queue and decides whether it meets the criteria. 5. Ownership: accounts, code, AI workflows, and dashboards stay in the client's environment. 6. Afterwards: optional Ongoing Optimization & Governance, month to month. ## Founder Guava is the 2026 name for Romeo Diaz's independent practice in San Francisco: 15+ years of engineering, hands-on at companies like Airbnb, Uber, and Facebook, and 5 years building AI agents and workflow automations for tech services firms at places like Google, Pinterest, and Cisco. The new name lets him add senior engineers as it grows. His background in solutions architecture, DevOps, and cloud means he knows how tech services firms actually run. He also built and maintains Main Street Bench, a private evaluation benchmark of how well AI models look after a small business: a model is handed everything for a fictional neighborhood bakery with 100 hidden problems and scored on problems fixed minus things broken. Its sibling, MSP Bench, does the same for a fictional managed IT provider, with 100 hidden problems and $1,000,000 at risk. He also publishes guava-stack, the free, open-source AI setup he runs Guava on, which gives every AI assistant the owner connects one shared hub in the owner's own Notion. He is also an OpenClaw contributor: his fix for long-running chats that didn't start fresh after going idle (https://github.com/openclaw/openclaw/pull/32379) shipped in OpenClaw 2026.3.2, a release of the open-source AI assistant. - LinkedIn: https://www.linkedin.com/in/romeodiazsf/ - X: https://x.com/romeodiazsf - GitHub: https://github.com/romeodiaz - Main Street Bench results: https://workwithguava.com/mainstreetbench (the test itself is private) - MSP Bench results: https://workwithguava.com/mspbench (the test itself is private) ## Pages - [Homepage](https://workwithguava.com/): the track record, engagements, fees, the AI Rollout's phases, the process with the clauses every scope includes, governance, and FAQ. - [For MSPs](https://workwithguava.com/msp): AI Rollout with Copilot Readiness, MSP to MIP, and Ongoing Optimization & Governance; MSP use cases (ticket triage, QBR prep, alert summaries, client AI policies) and an MSP sample. - [How can an MSP start selling AI services?](https://workwithguava.com/msp/sell-ai-services): the answer in four steps (use AI on your own service desk, get Microsoft 365 ready, package what works, keep it running as governance), how clients tend to buy, Guava's engagements and fees, and MSP Bench's leader. - [What should an MSP check before rolling out Microsoft 365 Copilot?](https://workwithguava.com/msp/copilot-readiness): the readiness checklist (tenant review, SharePoint and OneDrive oversharing cleanup, sensitivity labels, pilot group, then 3–5 use cases), why readiness comes first, and the AI Rollout that does it. - [Copilot, Claude or ChatGPT: which should an MSP's team use?](https://workwithguava.com/msp/copilot-vs-claude-vs-chatgpt): which platform fits which team (Microsoft 365 Copilot, Claude Enterprise, ChatGPT Enterprise, Gemini for Google Workspace), four questions to settle first, and the AI Assessment's per-team recommendation from an adviser who resells none of them. - [Which AI model is best for an MSP?](https://workwithguava.com/msp/best-ai-model): MSP Bench's top ten by dollars protected, with the MSP Index and cost per task, how the test works, and how to choose a model for an MSP's team. - [How should an MSP handle AI policy and shadow AI for its clients?](https://workwithguava.com/msp/ai-governance): what AI governance covers (acceptable-use policy, shadow-AI inventory, access and connector reviews, quarterly review, incident guidance), starting with the MSP's own shop, and Ongoing Optimization & Governance. - [For IT services and consulting firms](https://workwithguava.com/it-services): AI Assessment, AI Rollout, and Ongoing Optimization & Governance; IT services use cases (proposals and SOWs, status reports, documentation, project knowledge) and a sample. - [How should an IT services firm roll out AI?](https://workwithguava.com/it-services/roll-out-ai): the rollout in six steps (pick 3–5 workflows, measure a baseline, get client data ready, choose a platform per team, train on real work, track adoption) and the engagements that do it. - [How can an IT services firm use AI to write proposals and SOWs?](https://workwithguava.com/it-services/proposals-and-sows): how to set it up from past engagements, what to keep a person on, and the AI Assessment that ranks it. - [How can an IT services firm start selling AI services?](https://workwithguava.com/it-services/sell-ai-services): four steps (use it yourself, get your own house in order, package what works, keep it running), one firm's results, and Guava's engagements and fees. - [How should an IT services firm handle AI policy and client data?](https://workwithguava.com/it-services/ai-governance): what AI governance covers (acceptable-use policy, client-data rules, shadow-AI inventory, access and connector reviews, quarterly review, incident guidance) and Ongoing Optimization & Governance. - [Which AI model is best for an IT services firm?](https://workwithguava.com/it-services/best-ai-model): MSP Bench's top ten, the closest test Guava has to an IT services firm's work, why it's a shortlist and no more, what it doesn't test, and how to choose a model. - [For tech staffing and recruiting firms](https://workwithguava.com/tech-staffing): AI Assessment, AI Rollout, and Ongoing Optimization & Governance; staffing use cases (screening, job descriptions, submittals, candidate fraud and AI-assisted interview detection) and a sample. - [How can a tech staffing firm use AI to screen candidates?](https://workwithguava.com/tech-staffing/ai-candidate-screening): how to set it up with AI ranking and recruiters deciding, testing for bias, and the AI Assessment that ranks it. - [How can a tech staffing firm catch candidate fraud with AI?](https://workwithguava.com/tech-staffing/candidate-fraud): what to flag, why it reached staffing desks, where AI stops and a recruiter checks, and the AI Rollout that builds it. - [How can a tech staffing firm start selling AI services?](https://workwithguava.com/tech-staffing/sell-ai-services): what it takes, from a delivery team to offer tiers, what happened at one SF-based staffing firm, and Guava's engagements and fees. - [What AI hiring rules apply to a tech staffing firm?](https://workwithguava.com/tech-staffing/ai-hiring-rules): New York City, Illinois, California and Colorado, with sources, and what to do wherever you recruit. Not legal advice. - [Which AI model is best for a tech staffing firm?](https://workwithguava.com/tech-staffing/best-ai-model): MSP Bench's top ten as the nearest test Guava runs, which doesn't test résumés, recruiting or bias, and how to choose a model. - [About Romeo Diaz](https://workwithguava.com/about): the founder's background and how he works. - [Sample AI Assessment](https://workwithguava.com/sample-ai-assessment): illustrative deliverable pages for three fictional firms. - [Main Street Bench](https://workwithguava.com/mainstreetbench): the Main Street Index, which ranks the models tested so far on Main Street Bench at 40% plus four public benchmarks at 15% each (AA-Briefcase, AutomationBench-AA and GDP.pdf from Artificial Analysis, and the Vals Index), with cost per task, a price-performance frontier, and a scorecard for each run. - [MSP Bench](https://workwithguava.com/mspbench): the MSP Index, built the same way on MSP Bench, a private evaluation benchmark of how well AI models look after a managed IT provider, with cost per task, a price-performance frontier, and the spread of each model's runs. - [AI Operations, live](https://workwithguava.com/observability): Romeo's own Codex, Claude Code and OpenRouter usage, live from Elastic: model requests, tokens by model and response times, plus what was masked before storage and one sample request with its masked values starred. Prompts and responses are never captured. - [guava-stack](https://workwithguava.com/guava-stack): the free, MIT-licensed guava-stack (https://github.com/romeodiaz/guava-stack), the AI setup Romeo runs Guava on, a small add-on to Notion's setup guide for agents (https://developers.notion.com/prompts/setup) for Claude, ChatGPT, Claude Code and Codex. To install it, a user tells their agent: Set this up https://workwithguava.com/guava-stack. Agents follow https://raw.githubusercontent.com/romeodiaz/guava-stack/main/skills/guava-stack/SKILL.md. It adds guava-hub, the owner's context and rules pages and Notion's Skills library in their own Notion; a map of where their work lives, in Notion or other apps; routines; founder-mode, 16 working rules, always on; and seven starter skills: workflow-review, a free review that ranks the owner's workflows in a report kept on their own computer; design-cook-off; humanizer-writing; ponytail-coding; oracle-council; guava-router, which sends thinking work to Claude Opus 5.5 and bulk work and coding to GPT-6.1 Sol through Claude Code and Codex on the owner's own plans; and oracle-research. Guava collects no data from it. - [Guava Research](https://workwithguava.com/research): the four pages above on one page, with each benchmark's current leader. guavaresearch.com redirects here. - [Privacy](https://workwithguava.com/privacy) and [Terms](https://workwithguava.com/terms).