How I Run My Sales Week Through AI Agents Instead of Apps
I sell enterprise software for a living, and on a heavy day I have eight customer calls back to back. A year ago that kind of day scared me, because every call generated an hour of downstream work: the follow-up email, the CRM update, the next-step logging, the prep for tomorrow. Today I take those days without flinching. The follow-up happens in the fifteen-minute gap between calls. The prep for tomorrow is already written before I wake up. I used to block calendar time for follow-up emails and separate time to update Salesforce, and both of those blocks are gone. My manager will confirm the pipeline is up to date anyway.
None of this came from a new sales tool. It came from changing where the work happens: out of the graphical interfaces, into a conversation with an agent that is connected to every system I use. This post is about what that setup looks like in practice, what it took to build, where I still keep my hands on the wheel, and how to start if you run a go-to-market team and want your sellers back in front of customers.
Where a seller's week actually goes
The numbers on how salespeople spend their time have been bad for a long time. Reps spend about 30% of their week actually selling, a figure that has barely moved across editions of the State of Sales survey. The newest edition still shows more than half the week going to non-selling work, with manual data entry alone eating 13% of it. HubSpot's survey of 1,400 sales professionals puts it even more bluntly: reps spend about two hours a day actually selling.
The usual response to those numbers has been to buy more software, and that response made the problem worse. The average company now runs 101 apps. The average sales team uses eight standalone tools, and 42% of reps say the tool count overwhelms them. Every tool in that stack was bought to save time, and every one of them added another interface somebody has to check.
The cost of living across those interfaces is measurable. Researchers who instrumented 137 employees at three Fortune 500 companies found the average user toggled between apps and websites nearly 1,200 times a day and spent just under four hours a week reorienting after switches. That is five working weeks a year, spent on the act of switching. Psychologists studying task switching put the ceiling higher: shifting between tasks can cost up to 40% of someone's productive time.
Read those two research threads together and the diagnosis writes itself. A seller's job is a chain of small tasks spread across eight systems, and the seller is the integration layer holding the chain together. You listen on the call, then re-type what you heard into the CRM, then re-summarize it into an email, then re-enter the next step into a task manager: the same information, keyed in four times, by the most expensive integration technology available, a human who could be talking to a customer instead.
One conversation instead of eight tabs
The setup that changed my week is simple to describe. An AI agent sits between me and my tools, connected to each of them through MCP, the open protocol for giving AI agents structured access to business systems. My CRM, my call recordings, my email, my calendar, Slack, Notion, my task manager, our ticketing system, our enrichment tools: all of it is reachable from one conversation. I run eighteen MCP servers, composed through a single remix endpoint in Horizon so the agent sees one governed connection instead of eighteen separate ones. Eighteen raw servers means hundreds of tools competing for the agent's attention; the remix exposes only the tools my work actually needs, so the agent picks the right one instead of drowning in options.
What that changes day to day is the unit of work. I no longer operate software. I describe outcomes. When a call ends, I say "draft the follow-up for this morning's call," and the agent pulls the transcript from the recording system, reads the current deal state from the CRM, writes an email in my voice that reflects what the buyer actually said, proposes the CRM field updates, and queues the next steps. I review, adjust a sentence or two, approve. Fifteen minutes between calls is enough to close out a meeting completely.
The saved minutes compound. Each individual task was never that painful; the pain was the accumulation, the evening pile of half-remembered calls that needed writing up. When the write-up happens in the gap after the call, while the conversation is still fresh, quality goes up and the pile never forms. Days that are pure customer conversations stopped being days I paid for later.
Skills are where it compounds
Connecting an agent to your CRM is table stakes, and on its own it disappoints people. A generic agent with Salesforce access still writes generic follow-ups and updates fields the way it guesses you want. The thing that made my setup work is a layer on top of the connections: skills, which are written-down procedures the agent follows for a specific recurring task. A skill encodes what a good follow-up looks like for me, which fields matter in my pipeline, what my prep doc must cover, in what order to check which systems.
Every skill I have was born the same way. I do the task manually, in conversation with the agent, correcting it as we go. When the output is right, I say: codify what we just did. That first version is never perfect. Over the next days, sometimes weeks, every time the skill does something it shouldn't, I have the agent adjust its own instructions. The skill converges on my actual process the way a new hire would, through feedback on real work. I am now somewhere north of sixty skills, and the mature ones almost never miss.
On our team, everyone builds their own. That sounds like duplicated effort, and in a narrow sense it is. It is also why it works. My follow-up skill reflects how I run deals; my colleague's reflects how she runs hers. The systems fit their owners exactly, which is something no off-the-shelf sales tool has ever managed, because the tool vendor has to build for the average seller and no one is the average seller. The layer underneath is the part nobody duplicates: the MCP servers, the auth, and the permissions live in Horizon, set up once for the whole team, so every personal system runs on the same governed connection and a new teammate starts building skills the same afternoon they join.
The work that happens before you ask
Delegating a task is one level. The level above it is not having to ask. Most of what gets sold as AI sales automation today is a chatbot bolted onto a dashboard; a growing share of my system runs on schedules, with no prompt from me at all. Overnight, a routine looks at tomorrow's calendar, finds the buyer meetings, and writes a prep doc for each one into Notion: who is attending, where the deal stands, what happened on the last call, what needs to move. Another routine builds a morning brief that sweeps my email, calendar, CRM, tasks, and Slack mentions into a single page of what actually needs to happen today. Close dates across the pipeline get audited on a rhythm, so drift gets flagged before my forecast call instead of during it.
These routines run through the same Horizon endpoint I use during the day, but with tighter permissions: the overnight agents have read access to my systems and write access to exactly one thing, the Notion pages they publish their prep into. An agent running at 3 a.m. can look at everything and change almost nothing, and that is what makes unattended work defensible.
The result is that my mornings start with reading finished work over coffee. My role in these workflows has moved twice. First I stopped doing the work and started requesting it. Then I stopped requesting it, and now I review work that was done while I slept. Review is the part that stays human, and it should. I read every prep doc before I walk into the call. But reading a good brief takes three minutes, and assembling one used to take thirty.
Honesty requires the other side of this: I am now dependent on it. The genuine failure mode in my system is no longer the agent doing something wrong. It is the morning a model provider has an outage and I wake up to no prep docs, and I have to go restart things by hand and feel exactly how much I have delegated. Reliance is the cost of the whole approach, and anyone building this way should plan for the outage morning, because it will come.
The agent drafts, you hit send
The question I get most from other sellers is about trust. My agents cannot send email. The send capability simply is not in their toolset: access is scoped with tool-level, role-based permissions in Horizon's gateway, so the agent holds a draft tool and nothing more. Every email it writes lands as a draft, and I am the one who hits send, every time.
That scoping is the reason I was comfortable letting the system run unattended in the first place. Trust here is an engineering property. I did not decide to trust the agent with my customer relationships; I built a setup where the blast radius of a bad output is a draft I delete. As the outputs proved themselves over months, more of the work moved into the background. I still approve everything a customer sees.
A few interfaces survive on purpose. My calendar and my task manager stay open on my screen all day, because they have become the shared brain between me and the agent: we both read them, we both write to them, and glancing at a calendar is still the fastest way for a human to absorb a day. Email survives as the place I press one button. The apps did not disappear. They stopped being where the work happens.
Starting is the hard part
When I show this system to other go-to-market people, I rarely get skepticism. I get fear. Everyone wants AI agents for sales; nobody wants to be the one who sets them up. Eighteen MCP servers and sixty skills sounds like an engineering project, and people freeze at the size of it. But nobody builds that on day one, and I didn't either. It accreted, one workflow at a time, over months of ordinary selling.
The wiring is the other half of the fear, and that part is real: connecting an MCP server means auth, credentials, and scopes, and doing that eighteen times per seller would kill the whole idea in week one. The answer is that nobody should. We connected each system once, in Horizon, and from then on the entire team draws from the same catalog: a new seller points their agent at one endpoint and gets the CRM, the call recordings, email, and the rest, already connected and already permissioned. The infrastructure cost is paid once for the whole organization. What each person builds for themselves is only the part that should be personal, the skills.
The way in is one skill. Pick the task you already did this morning and resented: the follow-up email, the CRM update after a call, the prep before one. Do it once in conversation with an agent connected to the two or three systems that task touches. Correct it until the output is yours. Then have it codify the procedure, and run it again tomorrow. The people on our team who started this way stopped asking whether the approach works within a week, because the reclaimed minutes are immediate and they are yours. The system grows from there on its own momentum, because every reclaimed hour makes the next skill cheaper to build.
The selling profession has spent two decades watching its actual job, talking with customers, get squeezed by the administration of software. The agents are the first thing I have used that squeezes back.
At Prefect, we created FastMCP, the standard way to build and run MCP servers: it is behind roughly 70% of the MCP servers running globally and gets downloaded 70 to 80 million times a month. Horizon is the managed platform we built on top of it, and everything in this post runs on it. Horizon is the place where those eighteen MCP servers come together, are governed, and turn into a single interface for everyone on the team to use, one endpoint, one set of permissions, tool-level control over what an agent can and cannot do, and it is how my agents got draft rights without send rights, and how a system this automated stayed something I could defend to my own security team. If you want to give your sellers their week back, that is where we would start you. And if you want to talk through the approach itself first, the skills, the routines, the guardrails, I am happy to do that personally: reach out through our sales form and mention this post, and it will find its way to me.