The AI Stack for a 10-Person Company: What to Buy, Skip, and DIY

← Back to Blog

A ten-person company does not need an AI strategy. It needs five tools that earn their monthly fee and a rule for saying no to the rest. That is the whole argument of this article, and it is harder to follow than it sounds, because the pressure to buy is constant: new agents, new 'co-pilots,' new platforms that promise to run your marketing while you sleep. The version that works at this size is boring. It is a small stack, under a hundred and fifty dollars a month, each piece justified by a specific job it does better than a person doing it by hand. And the whole thing runs on tools you can cancel in a month, which keeps the downside small.

Before you buy anything, decide what a tool has to prove. Our rule of thumb: a tool earns its place when it saves more hours per month than it costs in dollars, at a modest wage rate — ten dollars an hour saved justifies a ten-dollar subscription. That standard kills most purchases on paper, which is the point. The stack below is what survives the test for most small teams. Read it with your own numbers in mind, and treat the prices as a shape: what matters is the order — writing and drafting first, speculative agents never. If you want the fuller version of the thinking behind that order — what AI can and cannot do, and how to run a pilot that tells the truth — our practical playbook for AI in business is the companion piece.

The Buy-vs-DIY Test

Before you compare tools, answer three questions about the job you want done. First, volume: how often does this task happen? A task that happens twice a month is not worth automating — the setup costs more than the time it saves. A task that happens forty times a week is a candidate, because the time it consumes is real and recurring. Second, rules: can you write down exactly how the task should be done, including what counts as done? If the steps are clear enough to explain to a stranger, a tool can follow them. If the task changes with context and judgment, you are building a bespoke system, and that is a different project with a different budget.

Third, risk: what happens when the tool gets it wrong? If the answer is 'nothing important,' you have a candidate for full automation. If the answer is 'a customer gets a wrong answer' or 'a contract gets misfiled,' you have a candidate for a human-in-the-loop version, where the tool drafts and a person approves. Risk is the question most buyers skip, and it is the one that separates tools that stay from tools that get unplugged after the first mistake. The test works in both directions: it tells you what to buy, and it tells you what to leave alone. The order matters too — volume first, because a low-volume task will never justify the setup, no matter how clever the tool.

Run your three answers through the list below and you will find that most of what the market is selling fails the test at this size. Autonomous agents fail the risk question. 'AI strategy' retainers fail the volume question. Fancy forecasting fails the rules question. The stack in the next section passes all three, which is why it is small — and why it is the only stack we recommend to ten-person companies.

• Volume: does the task happen often enough that the time it saves is real?
• Rules: can you write down the steps, including what 'done' looks like?
• Risk: what breaks if the tool errs, and who checks the output?

The Core Stack

Here is the stack we actually recommend, in the order we recommend buying it. The total lands under a hundred and fifty dollars a month, and every piece in it replaces a recurring weekly chore rather than a someday ambition. You will notice what is missing: no 'AI phone agent,' no autonomous marketing copilot, no predictive dashboard. Those are real products — they are just not the right first purchases for a team of ten, and most of them show up in the skip list below. The order also matches the buy-vs-DIY test: each purchase passes volume, rules, and risk before it earns a budget line.

The chart below shows the weekly hours a ten-person team typically recovers per function once the stack is in place — the shape, not the precision, matters. Document data entry leads because it is the least glamorous and the most mechanical: invoices, applications, and forms arrive in every format and someone types them into a system. Writing and drafting and inbox triage follow close behind. Reporting and lead routing trail, not because they are unimportant, but because they depend on data that has to be clean first. Buy in that order and each tool funds the next.

• Writing and drafting — the safest first buy, nothing ships without an edit
• Meeting notes — kills the summary and follow-up email chore
• Inbox triage — reclaims a morning a week for whoever runs the operation
• Document extraction — the quiet workhorse, verified against rules
• Routine support — absorbs the repetitive questions at any hour

Illustrative weekly hours recovered per function for a 10-person team — the pattern, not the precision, is the point.

Writing and drafting

One subscription covers most of what a small team needs: first drafts of proposals, job descriptions, emails, and posts, written in a shared voice. The workflow is simple — a person briefs, the model drafts, a person edits. The editing is not optional; it is the part that carries your voice and catches the confident mistakes. Teams that skip the edit get generic output; teams that edit get three times the output of the same person typing. This is the safest first buy because the risk question answers itself: nothing ships until a person approves it.

Meeting notes and inbox triage

Two tools do the jobs nobody wants to admit takes hours. Meeting notes: the recorder joins the call, produces a summary, action items, and decisions, and the summary lands in the thread before the call ends. No one takes notes, and the follow-up email writes itself. Inbox triage: the model reads the overnight pile, drafts replies to the routine fifty percent, flags the urgent, and leaves the rest for a human pass at the start of the day. Together these two reclaim a morning a week for whoever runs the operation.

Document extraction and routine support

Document extraction is the quiet workhorse: a model reads the invoice, the application, the insurance card, and writes the fields into your system, flagging anything that does not parse. Accuracy is verified against rules, so a bad read gets reviewed instead of filed. Routine support is the last piece — a chatbot trained on your FAQ that answers the same twenty questions at any hour and hands the complicated ones to a person. It does not replace support; it absorbs the volume that makes support miserable, and it is the piece most teams buy last and wish they had bought first.

What to Skip

The skip list matters as much as the buy list, because the market is optimized to sell you things you do not need. Autonomous agents — the ones that promise to 'run your outreach' or 'manage your social' while you sleep — fail the risk test at this size. They are impressive in a demo and expensive in practice: they act without context, they cost more than they save, and when they go wrong, they go wrong at volume. We have watched teams spend a quarter cleaning up what an agent did in a weekend. The 30-minute audit in the automation audit guide is a better use of that money, because it finds the busywork a simple tool can kill. Run the audit before you buy the agent, and the agent usually becomes unnecessary.

Next on the skip list: 'AI strategy' retainers and AI consultants who sell you a roadmap and a deck. At ten people, you do not need a roadmap; you need one working tool and the discipline to measure it. The strategy is three sentences — the buy-vs-DIY test above — and you just read it. Spend the retainer money on the tools instead. Finally, skip overkill forecasting. Predictive dashboards that promise to tell you next quarter's revenue are seductive and mostly wrong at this scale, because your data history is thin and the future is shaped by a handful of deals, not a pattern. If a dashboard cannot beat your own spreadsheet forecast, it is decoration.

The common thread in everything on this list is the same: it fails one of the three questions from the first section. Agents fail risk. Retainers fail volume — they are a one-time cost attached to a recurring problem. Forecasting fails rules — the output is a guess wearing a chart. When a vendor demo makes you feel behind, run the three questions and let them do the arguing. You are not behind. You are just being sold.

• Autonomous agents that promise to run a whole function unattended
• 'AI strategy' retainers and roadmap decks at ten-person scale
• Predictive forecasting dashboards built on thin data history
• Any tool you cannot cancel inside a month

Security and Privacy Defaults

The question every team asks is the right one: is it safe to paste company data into these tools? The answer depends entirely on which tool. Consumer chatbots are not business tools — anything you paste into them can be used to train models and can surface in someone else's answer. Business tiers with data-protection commitments are different: the vendor contracts not to train on your data, encrypts it in transit and at rest, and lets you delete it. The rule we use with clients is simple: treat the free tier like a public forum and the paid business tier like a locked file room, and never paste anything you would not put on your own website. That one sentence has prevented more incidents than any security policy we have written.

Second default: data stays in the business boundary. Customer names, financial details, and anything protected by regulation do not leave your approved tools, full stop. If a model needs to see a document to extract from it, it sees the document inside the business tier, not in a free chat window. The practical version of this rule: before you connect a tool to anything, write down what it can touch and what it cannot, and share that list with the team. Ambiguity is where mistakes live. Post the list next to the coffee machine; a policy nobody has seen is a policy that does not exist.

Third default: one owner and one reviewer per tool. The person who owns a tool owns its output, its renewals, and its mistakes. A second person reviews anything customer-facing before it ships. Add a monthly pass: does the tool still earn its fee, does it still have an owner, is it still in the approved list? Tools without owners are the ones that leak data and drift out of date; tools with owners get fixed when they break.

The test for any tool is not whether it is impressive. It is whether you would let it touch your customer list unsupervised.

What can touch what

Draw the line in one sentence: approved tools touch business data, everything else touches nothing. The approved list is short — your writing tool, your notes tool, your extraction tool, your support bot. Anything not on the list gets no data, no logins, no integrations. When someone asks 'can we try this new tool with our real data?', the answer is a conversation, not a default yes. The default is no until the tool earns a place on the list.

The defaults that cover most incidents

Four rules catch most problems before they happen. One: no customer data in free tiers. Two: no personal accounts for business tools — every login is a company login, so access dies with the employee. Three: anything that goes out gets a human review, at least in the first quarter. Four: every tool has an export path, so your data is never trapped when the vendor raises prices or the product gets acquired. Write these four down and you have covered the incidents we see most often.

The 30-Day Rollout

Buying the stack is the easy part. Getting it used is the project, and the failure mode is always the same: a tool bought, connected, and abandoned after two weeks because nobody was responsible for making it stick. The rollout below is the antidote — one tool at a time, one owner, one measurement, thirty days. It is the same shape as the four-step starting playbook in our AI in Business guide, scaled to a single purchase. If you already have a process that works, use it. If you do not, use this one. The discipline — one tool, one owner, one number — is the whole trick; the tool is interchangeable.

Week one: pick one tool from the core stack — the one that solves the loudest problem — and connect it to nothing yet. Set the owner, set the reviewer, and write the one number that will prove it works: hours saved per week, or minutes per task, or messages handled. Week two: run it in parallel with the old way. The tool drafts, a person approves, and nobody's job depends on it yet. This is the week you find the edge cases, and the edges are why the human review exists. Keep the parallel run honest: if the tool needs constant correction, that is data, not failure.

Week three: measure against the number you wrote in week one, and decide. If the tool clears the bar — it saves more than it costs — expand it to the full team and connect the integrations you deferred. If it does not clear the bar, say so in writing and kill it; a tool that fails in week three costs you a subscription, not a quarter. Week four: review the whole stack. Which tools earned their fee, which are borderline, which should go? This is also the moment to check the sprawl risk — every tool you keep should appear in the inventory we describe in tool sprawl is eating your margin, with an owner and a cost.

Thirty days after you start, you should have one tool in daily use with a number proving it, one tool in the pipeline, and a written decision on anything you tested and dropped. That is the entire goal — not a stack of ten tools, but a small set that earns its keep. The next purchase gets the same thirty days, and the stack compounds. By the end of the quarter, the busywork that used to eat Fridays has a home, and you have the receipts to prove it.

• Week 1: pick one tool, name the owner and reviewer, write the one measurement
• Week 2: run it in parallel, with a human approving every output
• Week 3: measure against the number and decide expand or kill, in writing
• Week 4: review the whole stack and update the tool inventory