Put AI to work on the things only your firm does.

Not by asking what AI can automate. Six questions, in order: where is the firm losing money, time or clients; what constraint causes the loss; why the existing team and systems cannot remove it; whether AI can remove it and how much of it; where human judgment must remain; and what proprietary system should be built around it that the firm then owns. The answer to the sixth is the initiative, and it is usually smaller than the idea the firm walked in with. You are not buying AI. You are buying the removal of a constraint, and you own the system that removes it.

One workflow, observed rather than described. We sit with the people who do the work and record the last ten real cases: what triggers it, who touches it first, what judgment is applied and on what information, what happens next and in which system, what breaks the standard path, who takes over and how they find out, and what number records that it closed. The gaps between the official process and the real one are the findings, and they are why AI implementations fail when they start from the SOP. It is the first deliverable of Build Scoping and the morning of the Growth Systems Intensive.

Human in the loop says a person is somewhere in the process. The Exception Architecture says where. For each workflow it states, in writing, what the AI handles on its own, what it asks for more information about, what it escalates and to whom, what it refuses to attempt, what it logs, and what remains a named person's decision. The rule for placing the person is blast radius: if an error at a step is reversible and internal, the AI handles it; if it is irreversible, client-facing or regulated, a person approves it and the approval is logged beside what the AI proposed.

A chatbot is the visible tenth of an AI system. Under it sit nine layers: identity and access, guardrails, agent orchestration, context and memory, retrieval, the model layer, tools and integration, the data foundation, and evaluation and governance. The risk does not scale down with headcount: a twelve-person firm that lets an assistant read its documents carries the same confidentiality obligation as a firm of twelve hundred. What scales down is the implementation, and the build path above is sized to a firm, not an enterprise.

The Accelerator proves retrieval and the model layer on one use case. Build Scoping fixes identity and access, the data foundation (the Data-Ready Gate) and the model choice. The Proof of Concept builds orchestration, memory, guardrails and the evaluation set (the Golden Set). The Production Build adds the integrations and hardens everything. Governance Install and the Fractional CAIO own the policy, the audit trail and the ongoing evaluation. Nothing in the path is a chat window on its own.

Three triggers. First, when your obligations outgrow vendor terms of service. Client-confidential legal documents or PHI shouldn't flow through consumer AI tools. Second, when the value is in your proprietary knowledge. A copilot trained on your firm's precedents, protocols, or playbooks compounds in a way a generic subscription never will. Third, when workflow fit matters. Off-the-shelf tools automate generic tasks; custom AI automates yours. If none of the three apply yet, we'll tell you to keep the subscriptions.

Governance is designed in, not bolted on. For healthcare: PHI is processed under a signed BAA, transmitted via encrypted channels, and never routed through non-compliant LLMs. For law firms: systems respect confidentiality and UPL guardrails, with advertising-adjacent outputs reviewed against Florida Bar Rules 4-7.11 through 4-7.22 and ABA Model Rules where applicable. For wealth managers: SEC marketing-rule and books-and-records constraints shape what the AI may generate and what gets logged. Audit logging and traceability are scoped to each system's data flows, risk level, and operating environment, and we define that scope in writing before the build starts.

Starting floors are published on the pricing page, and the exact quote is scoped to your firm, because no two firms' data, risk posture, and workflow are alike. Professional firms start at a firm-sized budget: the $500 AI Use Policy Kit, the $1,997 AI Business Accelerator (a 90-day cohort, done with you), the $2,950 Executive AI Intensive, or the 90-Day CAIO Bootstrap from $4,500, with every paid step credited toward the next. Mid-market operators with a real data estate and a compliance review start at $4,000 Build Scoping (credited in full), then a Proof of Concept from $20,000. Most work begins one of two ways. One is a defined-scope project: a copilot, a document-analysis pipeline, or a governance framework. The other is a monthly fractional Chief AI Officer retainer that sequences the roadmap over time. Either path starts with the free AI Executive Readiness Assessment. It shows where your firm stands across six pillars before anyone writes a proposal. See pricing.

A private AI assistant trained on your firm's own knowledge base: precedents, SOPs, intake histories, and research memos. Your team queries it in plain language. A law-firm copilot drafts first-pass documents from your templates and flags conflicts. A medical-practice copilot answers protocol questions from your own clinical standards. A wealth-management copilot summarizes client files within your compliance perimeter. Your data stays in your environment, and the copilot's knowledge compounds as the firm's does.

We're vendor-neutral by design. That's half the point of the governance work. The mix depends on your security posture and workload. Builds draw on frontier models (Anthropic Claude, OpenAI, Google Gemini) via their enterprise APIs. They use open-weight models run in your own environment when data cannot leave it. And they use retrieval infrastructure (vector stores, document pipelines) sized to the firm. Model choice is an architecture decision we justify in writing, not a default.

Two retainer tiers. The CAIO Signature retainer starts at $4,000 per month and covers AI strategy and quarterly roadmap ownership, vendor governance, architecture decisions, and board-level reporting in business terms. The CAIO Estate retainer starts at $10,400 per month and adds active build management, team training, and custom-project execution within the retainer. On the build side, a Custom AI build starts with a proof-of-concept pilot from $20,000, and the full production build is scoped from what the pilot proves, depending on data complexity and compliance requirements. Every path starts with the free AI Executive Readiness Assessment so the first conversation is grounded in evidence.

Three-layer architecture for regulated firms. First, data-path isolation: PHI is processed under a signed Business Associate Agreement, transmitted through encrypted channels only, and never routed through a non-compliant LLM. This applies to all healthcare clients and aligns with HIPAA 45 CFR Part 164 requirements. Second, model selection: open-weight models (such as Llama 3 or Mistral variants) can be deployed entirely within the firm's own environment when data cannot leave it. Third, access control: role-based permissions, and audit logging scoped to the interactions your risk posture requires, so there is a record of what was queried and what was returned.

A proof-of-concept pilot ships in about four to six weeks. Full copilot and document-analysis builds then ship in roughly 6 to 12 weeks from kickoff. That timeline typically breaks down as: 2 weeks for discovery and architecture (understanding the firm's data, workflow, and security posture), 3 to 5 weeks for build and iteration (model selection, retrieval pipeline, interface, integration testing), and 1 to 2 weeks for compliance review and a monitored soft launch. Larger projects involving multiple pipeline stages, BAA review, or regulatory sign-off take longer. We scope each project individually and put the timeline in writing before work begins.

Yes, with the right architecture. Custom document-analysis AI performs first-pass privilege review, extracts key terms, flags inconsistencies across contract versions, and drafts summaries from the firm's own templates. All of this operates as an attorney productivity tool, not a legal advice system. The AI does not communicate with clients, does not give advice, and every output is reviewed by an attorney before use. Systems are scoped against Florida Bar Rules 4-5.3 (responsibilities regarding non-attorney assistants) and relevant ABA Model Rules so the attorney-supervision requirement is embedded in the workflow, not left to chance.

Retrieval-augmented generation (RAG) is the architecture that lets an AI model answer questions using the firm's own documents rather than general training data. When a staff member queries the copilot, the system retrieves the most relevant chunks from a vector store of the firm's precedents, SOPs, intake histories, and memos, then feeds them to the language model as grounded context. The result is a copilot that cites the firm's own protocols with attribution rather than hallucinating from general knowledge. For a 10-partner law firm this means the copilot answers from that firm's matter history. For a 6-physician concierge practice it means the copilot references that practice's own clinical protocols.

Against a Golden Set. Before the build, the people who do the work today assemble real cases with the correct answer already known, and the partners agree on a pass rate. The Proof of Concept runs against every case, a human reviews the results, and it passes or fails on that number. Every miss goes back into the set for the next run. The decision to go to production is a number, not a feeling.

Five checks of the data the system will read, done before anything is built: quality, freshness, lineage, access, and a named owner. For a firm that means the practice management system, the document store, the inbox, and the intake forms. Most firms fail at least one check, and finding that out in a two-week scoping engagement is far cheaper than finding it out after the build.

No. But most firms that skip it spend the first consultation describing symptoms rather than gaps. The free AI Executive Readiness Assessment scores six pillars in ten minutes: strategy, governance, data readiness, security posture, team capability, and adoption readiness. It produces a 0 to 100 score, a readiness band, and your single lowest-scoring pillar. Most firms discover their biggest gap is not the one they expected. Starting from that data makes the first scoping conversation specific rather than exploratory.