AI, Automation & Smart Building Systems

How It Works & Technology Overview

This category spans two related but distinct practices. AI services: deploying AI assistants (Microsoft Copilot and similar), custom LLM integrations against vendor APIs, and the data-governance work that precedes them. Workflow automation and IoT: building workflow automations (Power Automate, Zapier, Make, n8n) and deploying occupancy/environmental IoT sensor networks that feed dashboards or analytics platforms. (Building automation proper — BAS/BMS, lighting control, HVAC integration — is its own category: Building Automation & Controls.)

AI deployments are governance-first: an AI assistant indexes whatever the user can already access, so tenant permission hygiene (sharing links, inherited permissions, stale groups) determines whether a rollout is safe. Automation projects are process-first: the automation encodes business rules, so undocumented exception paths — not the happy path — determine build size.

Common Deployment Scenarios

• Copilot/AI-assistant readiness + rollout: license assignment, data-security remediation (sharing audit, sensitivity labels, DLP baseline), pilot group, prompt training, adoption support.
• Departmental workflow automation: approval chains, document generation, system-to-system sync for a well-defined process, with error handling and a support owner.
• Custom LLM integration: an internal chatbot or document-processing pipeline against OpenAI/Anthropic/Google APIs, usually including retrieval over company content — data preparation dominates.
• Building sensor networks: occupancy, air quality, leak detection feeding dashboards or the BMS; effectively a wireless + cabling project with an analytics layer.

Products, Vendors & Characteristics

• AI assistants: Microsoft 365 Copilot (per-user subscription, requires eligible M365 licensing; value depends on tenant data hygiene); Google Gemini for Workspace; ChatGPT Enterprise/Claude for Work (bring-your-own-governance).
• Automation platforms: Power Automate (native to M365, licensing tiers gate premium connectors), Zapier/Make (fast, SaaS-to-SaaS), n8n (self-hosted option). Connector availability, not build skill, often decides feasibility.
• LLM APIs: OpenAI, Anthropic Claude, Google Gemini — usage-metered; costs scale with volume and must be modeled in recurring pricing.
• IoT platforms: sensor ecosystems are gateway-plus-cloud products; connector availability into the client's dashboard/BMS decides feasibility more than sensor specs do. When sensors integrate into a BAS, that integration scope belongs to Building Automation & Controls.

What Makes It Hard (Gotchas & Failure Modes)

• Data readiness is the project. Clients assume their file stores are clean; discovery scans reliably find broken permissions and over-shared content. Deploying an AI assistant over that estate is a data-exposure incident. Remediation is a prerequisite phase and adds 40–60% to naive timelines.
• Exception paths sink automations. The manager describes the perfect-day process; frontline staff handle a dozen unwritten exceptions. Each exception is branching logic, error trapping, and testing.
• Change management failure. Technically flawless automations get bypassed by staff who don't trust them; AI rollouts without champions and training become shelfware. Adoption work is billable scope.
• Subjective success criteria. "Improve the employee experience" projects can't close. Force measurable acceptance criteria before pricing.
• API drift and model churn. LLM APIs, connectors, and SaaS endpoints change; unowned automations rot. A maintenance agreement (or explicit disclaimer) belongs in every proposal.
• Pilot scope creep: automations approved for one department silently expand to three; every added department brings its own exception paths and stakeholders — re-scope, don't absorb.

Effort & Estimation Patterns

• AI-assistant deployment, ~50-seat org: 3–4 days deployment (licensing, policy baseline, pilot) plus 1–2 days training/enablement. Data-security remediation is separate and discovered by scan — commonly 1–3 additional days at this size, far more with large legacy file shares.
• Workflow automation, well-defined process: 1–3 days; multiply by the exception-path count found in discovery, not the happy path. Add 10–20% for documentation and handover.
• Custom LLM integration/pilot: 1–3 weeks depending on data preparation and retrieval quality targets; treat data cleanup as its own phase.
• Custom integration engagements: never quote from a verbal description — "just update the code" ranges 4–40 hours; require a scope document or price a paid discovery block.
• IoT sensor deployments: estimate as a wireless/cabling project first (mounting, power, gateway placement, RF verification), then add platform configuration; per-sensor field time 20–45 minutes plus network/gateway setup.
• Recurring: managed AI/automation support (monitoring runs, fixing broken connectors, model/prompt tuning) prices as a monthly retainer; API usage costs pass through with margin.