# What should a small consultancy automate first with AI?

> Start with one internal hand-off after a finished client call: turn agreed actions into reviewable project tasks, with the source and a person still in control.

For a small consultancy, the first useful AI workflow is often a hand-off everyone already recognizes: a client call ends, someone works out what was agreed, and the team turns the commitments into follow-up work.

Start by automating the internal sorting, not the promise to the client. Ask an assistant to find candidate actions in the finalized meeting record, connect each one to its source, and propose tasks on the right project. A person decides what was actually agreed and approves the tasks. That is a narrow job with a visible result and a clear point for human review.

## Why start after the call?

The workflow has a known input: a finished call’s notes or transcript. It has a useful output: a short list of proposed internal actions, filed on the project where the team works. And the person who owns the client relationship can check the result before it becomes a commitment.

That makes it easier to spot mistakes than asking an agent to run a broad process such as “manage this client.” The agent does not need to guess what counts as progress, decide which clients need attention, or send a message in your name. It is helping with one repeated hand-off.

## A practical first pass

**Hypothetical example, not an automatic Starlings feature:** a consultant connects Claude to a workspace and asks it to review the finalized notes from a meeting filed on a project. The prompt asks for three things:

1. Separate explicit commitments from ideas or questions that were not agreed.
2. For each candidate action, quote or point to the supporting passage and name the likely owner only when the record makes it clear.
3. Draft internal project tasks for the consultant to review; leave uncertain items as questions instead of turning them into tasks.

The consultant checks the source, edits the titles and owners, and asks the assistant to create only the accepted tasks. The team can then see the tasks in the project and the related conversation, rather than relying on someone to copy a summary out of a private chat.

In Starlings, a member’s connected Claude can read only records available to that member. It can read a finalized meeting artifact when the member hosted that meeting or when the meeting is filed on a subject they can access. Creating tasks requires a connector with write access, which the member must explicitly allow. The member can begin with read-only access and pair again with **Also write** only if the trial is useful.

## Keep the first workflow small

Pick a recurring project call where the team already takes notes. Agree on what counts as an action before inviting an agent into the process. For example: “We agreed to do this” is a candidate task; “we could consider this” is not. Keep client emails, pricing, scope changes and other external commitments out of the first pass.

For the first few calls, compare the proposed list with what the consultant would have written unaided. Track whether it saves time, whether it misses or invents commitments, and how much editing each result needs. A workflow that produces more review than it saves is not ready to expand.

There is an important product boundary here: Meet does not give an agent live audio or a live transcript. It does not store the audio or video of its calls. The workflow begins after a meeting has ended and its text artifact is available to an authorized person. The example above also does not send an email; it proposes internal work for a person to review.

## What to automate next

Only after the first hand-off is reliable should you try another. A useful next step might be preparing an internal client brief from the project’s visible history before a call. Keep the same discipline: name the sources, mark uncertainty, and make the output easy to check where the team will use it.

For the access model and Claude setup, see [how to connect Claude to your workspace](/help/howto/connect-claude). For why the request and result belong with the project, read [why an agent should work in the same project](/field-notes/why-agents-belong-in-the-same-project).

---

Canonical page: https://starlingshq.com/field-notes/what-small-consultancies-should-automate-first