Start by reading a year of the team's email, read-only, to see where the time goes. Then automate the work that repeats, one job at a time, with a person approving anything that leaves the building.
18 questions owners asked us, from 12 calls. Updated 2026-10-23.
Some work has to stay on machines you own, for example where security rules keep the data on site. Everything else can usually go to a hosted model. Sort the work into those two groups before you pick a setup for each job.
Running agents on your own machines is an extra cost, and it can be a large one. Where the work allows it, a hosted model avoids that cost. Count what each job costs to run, including server time, before you pick a setup. Open-source models can cut that cost, but only keep them where the quality is good enough for the job.
Yes. The repeating steps, such as drafts and follow-up timing, can be automated, and a person approves each message before it goes out and takes over once someone replies. Early on, a person calls each new lead before any automatic sequence starts, since automatic sends make more sense at higher volume. Measure first by reading a year of the team's email, read-only, to see where the time goes.
No, not for the routing itself. When the choices come from a short fixed list, plain rules can send each lead to the right sequence. Replies stay with people, who take over the conversation once someone answers, and measuring how leads arrive comes first.
Give each AI app its own key, with a spend limit on each one, and set an expiry date on every key. Renew each key on a regular cycle, around every ninety days, and prepare the next key before the old one expires. Keep all keys in one secure password vault with reminders several weeks ahead. A person does each renewal.
Traffic spikes are usually the smaller worry. The bigger one is someone reaching the models without permission, so put a gateway in front of them and require proper login. Give each app its own key with a spend limit, so one app's heavy use stays inside its own budget.
One setup runs a virtual server, a computer in a data centre that you reach over the internet, to hold the memory system and the scheduled jobs. Whether you need one depends on what your agents have to reach and how much of your data you want to hold yourself. Sort the work first, then decide where each job runs.
Start with a folder of markdown files, one for each person, task and entity you deal with. This shows what kinds of things you need to remember. Memory is complicated, so use an existing memory tool and avoid building your own from scratch. Move to a database once the notes grow to a size that a folder of files cannot handle.
Some people keep only a summary of each call, which can help with compliance. A paid notetaker is only worth it if you have enough meetings to justify the cost. For occasional calls, the built-in AI notes in your meeting app may be enough.
Start by timing the job as it runs now, including the waiting time, so you know what is worth automating. Give the agent examples of finished letters and the steps it must follow, then have a person review each draft before it goes to the prospect. Add more of the work only once the first step is running and trusted.
Write down each step, who does it, and how long each one waits before the next starts. Both the active time and the background waiting count. Automate the step that repeats most, then measure again before you touch the next one.
Parts that depend on each other are hard to troubleshoot when one of them breaks, so keep each workflow to one job. Get that job running and measured before you connect the next part. Where an API already does the job, use it and skip custom code. Anything that leaves the building still goes past a person.
Split the work into separate workflows, each with its own goal and rules, so each one can be measured on its own. Work with high compliance requirements, such as handling money, needs a person to approve every action. Lower-risk work can run on a lighter setup. Build the next workflow only after the first one is running and trusted.
Pick a goal you can measure, such as a count of visits or clicks, so you can tell when the job works. Give the job one day: start on a single day and have something working by the end of it. The next day, start again from what you learned, without carrying the old version forward. Where an API already does the job, use it and skip custom code.
Start with read-only access to one inbox, so the tool can read mail but cannot change it. Your IT team can see how the access is set up and switch it off at any time. Begin with a single inbox before deciding whether the whole team should be connected. Nothing that leaves the building goes out without a person approving it.
Screen recording captures a person's whole workflow, but staff can read it as surveillance, and the people who run the work may push back even when the owner wants it. A recorded video call can be more efficient than screen capture. Start with the inbox to see where the time goes.
Start with the inbox. Even where a team says it works mostly in chat or meetings, work that produces a deliverable usually leaves a trace in email.
Leads can be a large part of the inbox. The inbox also holds invoicing and other back-office steps, so count all of it before you choose what to automate. Measure the whole inbox first, then pick the one job that repeats most.
Today. Done person by person: each team member's email is gone through to see the hours spent, the projects, and who owns each one.
What changes. An agent could read a year of the inbox, read-only, and draft the breakdown by project and owner, for a person to check before it goes to the client.
Today. A weekly meeting with the client to set priorities and check where each piece of work stands, with the agents doing the reporting.
What changes. Once the reports are trusted, an agent could draft the weekly status for a person to review before it goes to the client.
Everything that matters in a business goes through an inbox at some point. Read a year of it, read-only, and you can see where each person's week goes. Then automate the work that repeats, one job at a time.
