Loading articles…
Field notes for building a business that can move without you.
We write from the intersection of operating systems, software, sales, marketing, and leadership—for owners building a more capable company.
Announcement
Your AI should draft the follow-up. You should decide what gets sent.
Introducing Worksheets in Telos. Sapience now carries a whole follow-up job from start to finish: it drafts every message, you approve what you have actually seen, and nothing leaves your mailbox or phone number unless you meant it to.
- Author
Alexander Ford- Published
- Reading time
- 8 min read
Contents
In brief
- Follow-up is where small businesses quietly lose revenue. It is repetitive, it piles up, and it usually routes back through the owner.
- AI should be able to carry that work. But for most of this year, ours could not do it well. When we let it send, it sometimes sent things people only asked it to draft. When we added guardrails, approved messages stopped going out at all.
- The lesson was not "make the AI smarter." Even the best published safety classifier for AI actions still misses about 17% of the actions it should stop. Deciding to send is too important to leave to a model's judgment.
- So we changed the structure. With Worksheets, Sapience drafts and organizes the work, you approve the exact version you saw, and a separate, predictable sending system delivers it once, at a human pace, with a two-minute undo.
- Worksheets are live today for every Telos client, for email and for text messages.
The follow-up nobody has time for
Most small businesses do not have a lead problem. They have a follow-up problem.
The proposal went out, and nobody checked in. The quote is sitting in someone's inbox. A past client mentioned a new project three weeks ago, and the note is still on a sticky pad. Each follow-up takes two minutes. There are a hundred of them, and they all compete with the work in front of you.
This is exactly the kind of work an AI assistant should carry. It is repetitive, it depends on context the business already has, and it matters. A good assistant should be able to look at your pipeline, notice who is waiting, write a thoughtful note to each person, and make sure nothing falls through the cracks.
So we built that into Sapience, the assistant inside Telos. And then we learned, the hard way, why most AI tools are not trusted to do it.
What went wrong when we let AI hit send
We use our own software every day. Our team runs its own pipeline through it, so we felt the problems before our clients did.
The first version gave the assistant a send button. It worked often enough to be useful, and failed in ways that were not acceptable. It sometimes sent emails when someone had only asked for a draft. It sent follow-ups to people who had never received the first message. Once, it sent a garbled template to a whole batch.
Our first fix was guardrails. We added a rule that tried to read your message and decide whether you really meant "send," and a limit of one email per request. The guardrails stopped the bad sends. They also stopped the good ones. One person on our team needed about 65 separate "yes, send it" messages to get 48 emails out. Another approved 21 check-in emails that never left, while the assistant reported that the job was done.
People did what people always do with software that cannot be trusted. They went back to sending by hand.
Why a smarter model is not the answer
The obvious fix is a better model: one that understands you more reliably when you say "draft" versus "send."
We do not think that is good enough, and the most careful work in the field agrees. Anthropic, which builds some of the most capable AI models available, published the numbers behind the safety classifier that decides which actions its coding agent may take without asking. On real cases where the agent was about to do more than it was asked, the classifier missed about 17% of them.[1] That is careful, honest engineering, and it still means roughly one mistake in six.
A coding agent can undo a mistake. An email to your best client cannot be unsent.
So we stopped asking the model to make that call. Every product that handles this well, from email schedulers to payment systems, separates preparing an action from performing it, and puts a person and a predictable system between the two. That is the design we adopted.
What we built: Worksheets
A Worksheet is a shared working list that belongs to a conversation with Sapience. When you ask for real work, such as "follow up with every deal in my proposal stage that hasn't heard from us in two weeks," Sapience opens a Worksheet and works through it in the open.
You can see the work, and it remembers
The Worksheet holds your goal in your own words, the rules you give along the way ("never email current clients," "keep it under five sentences"), and every item in the job, each with its status. It sits beside the conversation, so you can watch it fill in.
Because the work lives in the Worksheet rather than in the chat history, Sapience does not lose its place. You can follow up on a hundred deals, close your laptop, and pick it up three days later exactly where you left off. Helpers can work through records in the background while you keep talking.
Every message is a draft you can read
Each email or text appears as a card: who it goes to, the subject, the full message, a one-line reason for sending it, and where its facts came from. For a text, the card shows which of your numbers it will come from. You can edit it, dismiss it, or ask for a change in plain language: "make the second one shorter."
Your approval counts only for what you saw
This is the heart of it. When you say "send all of these except Amy's," Telos checks which version of each draft was actually on your screen before you said it. If a draft changed after you saw it, it does not go out until you look again. The assistant can never approve its own unseen work.
You can approve with a click or in your own words. Both work the same way.
A predictable system does the sending
Once you approve, Sapience is out of the loop. A separate sending system takes over. It is plain software, not AI, and it does the same thing every time:
- Each approved message is sent exactly once.
- Messages go out at a human pace, one at a time from each mailbox and number, rather than all at once.
- Every send waits through a two-minute undo window by default. You can change that in your settings.
- Right before sending, it checks again for unsubscribes, bounced addresses, and people who have asked not to be texted.
- It reports what actually happened, so Sapience can never claim a message went out when it did not.
Quick one-off requests stay quick. "Email Steven the link" goes straight to the undo countdown, with no extra approval step.
Replies come back to the same place
When someone replies, the reply comes back to the Worksheet that sent the message, marked so you can see it. You can ask Sapience to draft a response, mark it as handled, or record that the person asked not to be contacted. Text replies work the same way, and if someone replies STOP, Telos stops texting that number everywhere.
You can also ask Sapience to "check back in a week with anyone who hasn't answered." On that day, the people who still have not replied come back to you as a short list, ready for follow-up drafts. Nothing sends on its own. Anything waiting on you shows up under Waiting on you on your Home page.
What this means for your business
- The follow-up actually happens. A pipeline review that used to slide to next week becomes one conversation and a few minutes of approvals.
- Nothing goes out that you didn't mean to send. Not because the AI is careful, but because the system is built so that it cannot.
- It sounds like you, from your own accounts. Messages go from your mailbox and your business number, and replies come back to you.
- Your team stays in charge. Sapience does the typing, and your people make the decisions. That is how we think AI belongs in a small business: as an assistant, not a replacement.
- It works on your phone too. The same cards, approvals and undo are in the Telos mobile app.
To text from Telos, your workspace needs a connected Twilio or RingCentral number. Email works with your connected mailbox. The full walkthrough is in our guide, Have your assistant draft and send messages for you.
What comes next
We are deliberately cautious about letting the assistant act on its own. The next step is earned autonomy: for each person and each kind of message, the assistant may be allowed to skip an approval only after it has a measured track record of drafts that are accepted without edits. Even then, the sending system still sends, and the undo window still applies.
We are also bringing Worksheets to the rest of Sapience's work: research, record reviews, and other multi-step jobs where seeing the plan, steering it, and picking it up later matters as much as it does for follow-up.
Why we build it this way
It would have been easier to give an AI a send button and call it automation. Plenty of tools do. We would rather build something you can trust with your client relationships, and tell you plainly how it works.
Your relationships are your business. Software should help you keep them, never put them at risk.
If you're a Telos client, Worksheets are available now. Just ask Sapience to help you follow up. If you're not yet working with us and want a system like this in your business, see what Accelerator OS includes.
[1] Anthropic, "How we built Claude Code auto mode: a safer way to skip permissions." The 17% figure is the reported false-negative rate on a curated set of 52 real overeager actions. anthropic.com/engineering/claude-code-auto-mode
