The AI stack for a five-person agency, and what to leave out

A five-person agency does not need an AI strategy. It needs five specific jobs covered and a monthly ceiling it refuses to exceed, because the failure mode here is not missing a tool — it is paying for eleven of them.

Most "AI stack" advice is written for a company with a platform team. A five-person agency has no platform team, no procurement process, and no tolerance for a tool that needs a fortnight of setup before it does anything. What follows is the smallest set of purchases that covers the work a small agency genuinely does every week.

The five jobs

Almost every task an agency tries to hand to AI falls into one of five buckets. Buy one tool per bucket and stop.

1. Writing and editing. Proposals, case studies, ad copy, email sequences. One general-purpose assistant is enough. The differentiator between the major assistants matters far less than whether your team maintains a shared library of prompts and tone rules.

2. Research and competitive work. Pulling apart a competitor's positioning, summarising a long report, turning ten sources into a briefing. This is where a long-context model earns its keep, and where the search-integrated assistants save the most time.

3. Production. Images, ad variants, simple video cuts, transcripts. Buy for volume, not for quality, because production quality is judged by the client's eye, not by a benchmark.

4. Operations and admin. Meeting notes, CRM hygiene, invoice chasing, scheduling. Unglamorous, and usually the fastest payback of the five, because the work is repetitive enough to measure.

5. Development and internal tooling. Only relevant if you build client sites or internal dashboards. If you do not, this bucket costs you nothing and you should not buy into it.

What each bucket costs

Rather than quote list prices that will be stale by the time you read them, work in bands. As a rule of thumb for a small team, plan on:

• One seat-based assistant at roughly the price of a mid-tier SaaS subscription per person, per month.
• One long-context or research tool, shared across the team rather than per seat.
• One production tool on a usage meter, with a hard spend cap switched on.
• One meeting-capture tool billed per seat, only for the people who are actually in client calls.
• Total, for five people: treat it as a fixed line in your overheads, not a per-project cost.

Every serious vendor publishes a per-seat plan and a usage-based plan. Per-seat is easier to predict; usage is cheaper if your work is spiky. Pick per-seat unless you have measured your usage for a full quarter.

What to leave out

Agents that run your business. Anything advertised as autonomously executing multi-step work in your systems is, for a five-person agency, a liability with a login. You do not have the oversight capacity to run one safely, and the failure is invisible until a client notices.

A second tool in a bucket you have already covered. Two writing assistants do not make you twice as fast. They make your brand voice inconsistent and your bill larger.

Anything that requires a migration. If a tool needs you to move your data, your docs or your process into it before it does anything useful, the switching cost dwarfs the value at your size.

Enterprise tiers. SSO, audit logs and role-based access control exist to satisfy a compliance department you do not have. Buy them the day a client contract demands them, not before.

The number that actually matters

Track one figure monthly: total AI spend divided by hours saved that you can name. If you cannot name the hours, you are not saving them. A tool that costs the price of a nice lunch per person per week is cheap if it returns an afternoon; it is expensive if the team quietly stopped opening it.

Audit quarterly. Cancelling a tool nobody opened last month is the highest-return action most small agencies can take with AI, and it takes four minutes.

What we don't know

We cannot tell you which vendor wins in any of these buckets for your specific mix of work, because the answer changes with your client base and the vendors change their pricing constantly. We also cannot measure your team's adoption, which is the variable that actually decides whether any of this pays for itself.

There is no reliable public data on how much time small agencies genuinely recover from these tools. Every figure you see quoted, including optimistic ones, comes from either a vendor or a study of large enterprises.

What to do next

1. List the tools your team is already paying for and mark which of the five buckets each one covers. Cancel anything that is the second tool in a bucket.
2. Set a fixed monthly ceiling for AI spend and put it in your overheads, so a new tool has to displace an old one rather than quietly adding to a subscription pile.
3. Pick one bucket — usually operations — and measure the hours saved for four weeks before you touch the other four.