AI for SMBs
AI for non-technical SMBs that actually works.
Most companies are stuck at the bottom of the AI maturity ladder, trying to jump straight to agents and automation before they can write a reliable prompt. The work that actually produces meanigful ROI, teaching your team to use the AI they already have, then building the custom tools on the data they already have, is the work most companies skip. That is the work we do.
The problem
According to BCG and McKinsey, roughly 70%-85% of AI projects fail in the first year. The pattern behind the number is consistent across most of the companies we've looked at. The money and the attention jump straight to agents and automation, where the vendor pressure lives and the board slides are easiest to produce. The work that makes any of it land gets skipped: teaching your team to use the AI they already have, and wiring it into the data they already have.
Most mid-market operators we talk to are still stuck at Layer 1 of the maturity ladder below, running their business through a chat window that has no memory of anything. They want to be at Layer 4, but they haven't done the work of Layer 2 or 3 yet.
The 4-layer AI maturity ladder
Where your company actually is, versus where LinkedIn thinks it should be.
- 1
Layer 1: Raw prompt
“Summarize this email.” “Draft a response.” Everyone starts here. Useful, but you start from scratch every single time.
- 2
Layer 2: Persistent context
Your role, your company's terminology, your quality standards, examples of good output written into a saved set of instructions the AI reads every session. Onboard the AI once instead of re-onboarding it every morning. Output goes from useful-sometimes to actually-reliable.
- 3
Layer 3: Connected data
The AI reads your actual files, spreadsheets, CRM records, emails instead of you copy-pasting everything into a chat window.
- 4
Layer 4: Agents and automation
Automated workflows that run without you guiding each step.
LinkedIn is obsessed with Layer 4 right now. Agentic this, autonomous that. But almost every company we talk to is still stuck at Layer 1. Layer 4 built on top of a bad Layer 1 produces confident garbage at scale.
Training and Build: two products, two doors
Where your team is on the ladder decides where you start.
Currently at Layers 1–2
Training
Most teams live at Layer 1 and don't know Layer 2 exists. The training takes your operators from starting-from-scratch prompts to persistent-context workflows on their real day-to-day work. Tailored, private, single-company.
Explore Training→Currently at Layers 3–4
Build
When your team hits the ceiling of what chat-interface work can do, the next move is connected data and automation: custom software that reads your actual files, coordinates across multiple systems, and runs in production without supervision.
Explore Build→Training teaches your team to work with Layer 3 concepts inside the AI tools they already have: connecting to data sources a chat interface already supports, loading saved files, using the built-in retrieval a vendor ships. Build ships the custom Layer 3 and 4 systems your team could not assemble on their own: multi-source data pipelines, production-grade retrieval, orchestrated workflows, cross-system consistency. The concept lives in both; the engineering lives only in Build.
If an off-the-shelf agent covers your use case, use it
If your team just needs an agent that reads email, drafts replies, and manages calendars, you don't need custom engineering. The major LLM vendors now ship agent-mode features that handle this class of work.
If any of these cover your use case, train your people, then go use the tools. The Build service is for problems where off-the-shelf agent features can't reach: multi-system orchestration, custom business logic, domain-specific workflows, production reliability.
OpenAI
ChatGPT agent
ChatGPT Plus and higher tiers.
Best for: solo operators without enterprise IT.
Native Gmail, Google Calendar, and Outlook integration. Most turnkey of the three. If your team runs on consumer Google or personal Microsoft accounts, this is the fastest path.
Anthropic
Claude Cowork
Claude Pro and higher tiers.
Best for: local file work and browser-based tasks.
Strongest for document analysis, coding, and research workflows. Inbox and calendar integration requires MCP connectors or Zapier bridging, which is less turnkey than the other two if email is the primary use case.
Microsoft
Microsoft 365 Copilot Cowork
M365 E-tier SKU, enterprise-only, GA May 2026.
Best for: companies already on Microsoft 365 Enterprise.
Native integration with Outlook, Teams, and the M365 stack. If IT has already approved M365 E-tier, this is the least-friction option. Nothing new to license, nothing new to install.
Why this works.
People and processes are 70% of the work
BCG’s own AI research puts algorithms at 10%, tech and data at 20%, and people and processes at 70%. Training handles the people layer. Build handles the process layer. Neither alone is enough, which is why most AI initiatives stall. Tooling is the smallest piece of the problem.
Layer 2 is where most of the ROI lives
The jump from raw prompting to persistent context is where output quality goes from sometimes-useful to actually-reliable. It doesn’t need custom engineering. It needs teaching. That’s what the cohort delivers in four weeks.
Private, single-company training beats public cohorts
Generic curriculum fails because your workflows are specific. Twelve of your operators in a room working on the tasks they’re actually stuck on today produces artifacts the company can use the next morning. Open cohorts optimize for per-seat revenue. We optimize for your team running without us.
Governance is a core deliverable, not a stub
The #1 reason AI implementations die at month 9 is that the AI champion takes a new job and nobody else can maintain the system. Build ships runbooks, monitoring dashboards, escalation protocols, and maintenance training so the system survives personnel changes. Vendors don’t sell this because it isn’t their value capture. We sell it because it’s the only reason anything lasts.
Training and Build sequence by capability, not by delivery
The logical order inside a company is Training first, then Build. You can’t hand custom software to a team that doesn’t know how to drive it, and you can’t train your way through a problem that needs real engineering. But inside an engagement the two can run in parallel. The Cohort + Build offering ships a scoped Build during the four weeks of training, so operator capability and the custom system land together.
Tell us where your team is on the ladder.
We'll tell you honestly whether Training or Build is the right next move, and if neither is, we'll tell you that too. The first conversation takes 30 minutes.
Looking for a full custom analytics platform instead? See our Platform work →