← Edge Notes
AI-Native10 min read
Essay · ai-native

The AI automation stack for a 10-person company

Every AI stack recommendation you read was written for a company with a CTO, a data team, and a DevOps function. If you have ten people and one person who is somewhat technical, those recommendations are noise. Here is what actually works at your scale.

The wrong frame: enterprise AI scaled down.

The default frame for AI stack advice is enterprise AI, compressed. Take what Shopify or Stripe or Google is doing with AI, remove the parts that require 50 engineers, and describe what is left as a stack for small companies. This produces recommendations that are technically correct and practically useless: use a vector database for your knowledge retrieval, build an evaluation framework for your prompts, instrument your LLM calls with observability tooling.

A 10-person company does not have an MLOps engineer. It does not have a data infrastructure. It may not have any engineers at all. The AI stack for this company needs to be composed of tools that can be configured by someone who is not an engineer, that run reliably without maintenance, that fail loudly when they fail, and that cost less than $1,000 per month for the whole team.

This is a different design problem than the enterprise AI problem. The enterprise problem is optimisation at scale: reducing latency by 12 milliseconds, improving hallucination rate from 3 percent to 1.8 percent, cutting inference cost by 40 percent. The small company problem is different: getting any automation working at all, making sure it keeps working without a dedicated person watching it, and ensuring that when it breaks, someone notices before a customer does.

The stack described below is designed for this problem. It is opinionated, it has trade-offs, and it will not be right for every 10-person company. But it is grounded in what actually runs reliably without engineering maintenance, what integrates with the tools small companies already use, and what costs within a budget that makes sense at sub-$50K ARR.

Layer one: the AI assistant that everyone uses.

The first layer of the stack is a team AI assistant — a single subscription that every person on the team has access to and that is actively used for individual work tasks. This sounds obvious, but most small companies are in one of two failure modes: either only one or two people use AI tools (high tool, low leverage), or everyone has their own individual subscriptions that are not connected and not shared (high cost, low institutional benefit).

The right tool here is a team plan of an AI assistant — Claude, ChatGPT, or Gemini, depending on your primary use cases. Claude performs better on long-document analysis and careful reasoning. ChatGPT has the broadest ecosystem of third-party integrations. Gemini integrates most naturally if your team is already in Google Workspace. Pick one, pay for the team plan, and get everyone using it actively before you build anything on top of it.

The discipline here is adoption, not selection. A 10-person team where everyone uses the AI assistant for at least one real task per day will extract more value from a standard team subscription than a team where the founder has an enterprise account and nobody else has touched it. Measure adoption, not just access. The tool is only in your stack if people are using it.

Layer two: the automation platform that connects your tools.

The second layer is a no-code automation platform with built-in AI capabilities. This is where you build the workflows that run without human intervention: the automation that processes new form submissions, the automation that routes incoming emails, the automation that generates a draft response to a support ticket and puts it in the agent's queue, the automation that pulls weekly data from your analytics platform and formats it into a briefing document.

Make or Zapier with AI steps are the right tools at this layer for most 10-person companies. They connect to the tools you already use — Gmail, Slack, HubSpot, Notion, Airtable, Google Sheets — without requiring any code. They have built-in AI steps that can summarise, classify, extract, and draft without requiring you to manage an API connection. They have monitoring and error logging that tells you when an automation fails. And they cost between $20 and $100 per month for the volume a small company actually runs.

The mistake at this layer is over-engineering. A 10-person company does not need a custom webhook infrastructure. It does not need to manage its own API keys for every AI model. It does not need to build retry logic and error handling from scratch. The no-code platform handles all of this. Your job is to configure the automation correctly, test it with real data, and check the error logs once a week.

Start with three automations and do not add more until all three are running reliably. The most common failure mode is building ten automations, having four of them break in the first month, and then spending more time debugging automations than the automations were saving. Three working automations beat ten aspirational ones.

Layer three: the document intelligence tool.

Every 10-person company has a knowledge problem. Policies, procedures, product documentation, sales scripts, onboarding materials, and historical decisions are scattered across Notion pages, Google Docs, Slack threads, and the memory of whoever has been at the company longest. Every new hire takes three months to become functional because there is no organised way to access accumulated knowledge.

A document Q&A tool solves this. Tools like Notion AI, Guru with AI search, or a purpose-built tool like Dust allow your team to ask questions in natural language and receive answers drawn from your internal documents. "What is our refund policy for enterprise customers?" gets an accurate answer in 10 seconds instead of a Slack message to the one person who knows.

The limiting factor at this layer is document quality, not tool quality. A document Q&A tool is only as good as the documents feeding it. Before you configure the tool, spend a week auditing and consolidating your internal knowledge. Remove outdated documents. Consolidate duplicates. Write down the things that currently only live in people's heads. The tool multiplies the value of your documentation — it does not substitute for it.

The 10-person company does not need to build AI. It needs to configure AI well enough that the configuration keeps working without a full-time person watching it.

Layer four: AI writing assistance integrated into existing workflows.

The fourth layer is AI writing assistance built into the tools your team already uses for writing — not a separate tool they have to switch to. The best implementations of this are tools like Notion AI, Gemini in Google Workspace, or Microsoft Copilot, which surface AI assistance inline in the documents your team is already working in.

The value here is friction reduction. A standalone AI writing tool that requires copying text in and out is used by the motivated early adopters on your team and ignored by everyone else. An AI writing tool that appears when you highlight text in a document you are already editing gets used because the switching cost is near zero.

The specific use cases that produce the most value at this layer are first-draft generation for templated content (sales emails, proposal sections, job descriptions), tone improvement for external communications, and translation of internal shorthand into clear customer-facing language. These are all high-frequency writing tasks that do not require specialised knowledge — which makes them good candidates for AI assistance without high risk of the AI producing incorrect content.

What to avoid at this scale.

Custom AI development is the first thing to avoid. Building your own LLM integration, managing your own vector database, or writing custom prompt chains is engineering work that requires engineering maintenance. A 10-person company that builds a custom AI tool is committing to maintaining it — which means debugging it when it breaks, updating it when the underlying model changes, and having an engineering conversation every time a business stakeholder wants to change how it works. Off-the-shelf tools with configuration are almost always better at this scale.

Over-automation of judgment-heavy tasks is the second thing to avoid. AI automation works well when the task has a clear correct answer and the inputs are consistent. It fails when either of these conditions does not hold. Automating the routing of complex support escalations, automating the qualification of enterprise sales leads, or automating any decision with significant downside risk before you have strong confidence in the automation's accuracy produces errors that are harder to detect than the manual process and more expensive to fix.

Tool sprawl is the third thing to avoid. Each tool your team uses is a tool someone has to learn, a credential someone has to manage, and a subscription someone has to pay for and evaluate. A 10-person company with 12 AI tools is a company with a tool management problem masquerading as an AI strategy. Pick the layers above, fill each layer with one tool, and resist the temptation to add unless there is a clear use case the current stack cannot serve.

Frequently asked questions.

What is a realistic monthly AI tool budget for a 10-person company?

A well-configured AI stack for a 10-person company typically costs between $400 and $900 per month, covering a team AI assistant licence, an automation platform, a document Q&A tool, and an AI writing assistant. This is less than the cost of one additional hire and should produce the equivalent of one to two days per week of reclaimed time across the team, if the automations are well-chosen. The mistake is paying for capability you do not use — a $200 per month tool that only one person in the company uses is not a team investment, it is an individual subscription with overhead.

Should a small company build custom AI tools or use off-the-shelf products?

Off-the-shelf for almost everything, with rare exceptions. Custom AI tool development requires engineering time to build, ongoing engineering time to maintain, and a debugging process when things break. For a 10-person company, that engineering capacity is usually not available, and the custom tool often degrades without maintenance faster than the off-the-shelf alternative. Build custom only when you have an unusual workflow that no existing tool addresses, when the workflow is high-volume enough to justify the maintenance cost, and when you have someone on the team who can maintain it without it becoming their full-time job.

What AI automation breaks most often in small company deployments?

Automations that depend on consistent input format. The automation that processes inbound emails fails when a client sends a PDF instead of plain text. The automation that extracts data from invoices breaks when a vendor changes their invoice layout. The automation that routes support tickets by topic fails when customers describe problems in unexpected language. Before deploying any AI automation, map the ways the input could vary from the expected format and confirm that the automation handles those variations gracefully or fails loudly rather than silently producing wrong outputs.

How do I know which workflows to automate first?

Start with the workflows that are high-frequency, low-judgment, and currently done manually. The daily activity that takes 30 to 60 minutes and follows the same steps every time is a better first automation than the weekly report that requires someone to interpret the data. List every repeated task in your team's week, estimate the time cost per week, and rank by time saved multiplied by automation confidence. The automations at the top of that list should be your first three. Resist the temptation to start with the automation that would be impressive — start with the one that is clearly worth doing.

Working on this?

If this is the problem on your desk, a 30-minute intro is the fastest way to find out whether I can help.

Request an intro →