We put an AI CRM intelligence agent in front of our trade account list in January. It monitors around 200 accounts, flags the ones that have gone quiet, and drafts the outreach email. Six months in, I want to be honest about what it has changed and what it has not changed. The unchanged part turns out to be the more important part of the story.
What the agent actually does
The mechanics are straightforward. The agent pulls from our account data, identifies accounts that have gone quiet based on order and contact patterns, and produces a draft outreach email with the account's history surfaced. Our sales team reviews the draft, adjusts the tone where needed, and sends.
What that removes from a small sales team is the cognitive overhead of holding 200 account relationships in your head simultaneously. The manual scan of who ordered when, who went quiet after which visit, which venue changed buyers since the last conversation. That data housekeeping work is real, and it was eating meaningful time from a team of six.
According to research from Spirits Consulting Group, the administrative load on small spirits producer sales teams is one of the clearest opportunities for AI-assisted automation , precisely because the relationship density is high and the team size is small. That framing matches what we built for.
What changed
The flow of information into the sales team changed materially. Triggers, workflows, patterns, reminders , all set, all automated. The processes are the same as before, but the humans do not do them. That is the key shift.
What it produced in practice: the team spends more time on phone calls and in-person visits. Less time looking at spreadsheets and account data, more time talking to people and understanding each venue a little better. The data management layer became invisible in the right way.
The Beverage Information Group's 2026 outlook on AI in beverage logistics notes that the most durable gains from AI in small producer operations come from exactly this kind of task displacement , freeing human capacity for the relational work the data tools cannot do. We have found that to be accurate.
What did not change
The relationships. The human side of trade account management did not change because the agent cannot do it and was never designed to. The agent drafts the email. Whether it should be sent, whether the tone is right for this particular account manager, whether the moment is right , that is still a human call.
This is where the standard framing about automation gets slightly wrong. The narrative is that automation frees people to focus on what matters. That is true. What it glosses over is that the freed attention still has to be pointed somewhere, and pointing it well requires judgement that takes time to develop. The agent did not make our team better at relationships. It removed the noise that was getting in the way of that work.
Building something like sommelier.bot, which attempts to replicate expert recommendation in a conversational interface, illustrates the same principle from the other direction: even sophisticated AI interfaces built for relationship-adjacent tasks still rely on the human's underlying knowledge to work well. The tool is only as good as the human layer behind it.
The honest effort assessment
The effort to build something agentic is immense. Design and implementation is roughly 90% of your time, front-loaded. The remaining 10% is maintenance , indefinitely.
Most people assume that once an agent is deployed, it runs. That is not our experience. Configuration issues arise on a weekly basis: tweaks, improvements, flaws, edge cases. Token spend requires attention. Model selection requires periodic review , new models arrive frequently, and the model that was optimal six months ago may not be optimal now. You need to check whether you are optimised, because the default is drift.
We have built agents to help manage other agents. That is not a theoretical capability. It is a practical necessity once you have more than one workflow running.
The comparison that holds up at six months
| Dimension | Before the agent | After six months |
|---|---|---|
| Data housekeeping time | Manual, significant, weekly overhead per salesperson | Agent-handled; human reviews flagged accounts |
| Outreach first-drafts | Written from scratch, requiring account-history recall | Agent-drafted; human-edited before sending |
| Relationship quality | Depends on salesperson attention and memory | Same; attention now less fragmented by admin |
| Maintenance overhead | Low (no system) | Ongoing; model selection, config tweaks, edge cases |
| Team capacity for in-person visits | Limited by admin load | Higher; admin load redistributed to agent |
What the build taught us about agentic systems at this scale
The thing that surprised us most was not the technical complexity of the build. It was the design complexity. Getting the agent to surface the right account at the right time , not just the quietest account, but the one where an outreach would actually land , required a lot of iteration on the underlying logic.
That design work is not generalisable in a template. It requires knowing your specific accounts, your typical order cycles, your seasonal patterns. The agent we built for a six-person vermouth and amaro company is not the same agent that would work for a larger drinks distributor. The principles transfer. The specifics do not.
We also learned that the reward for this kind of build is not immediate and is not easily quantified. The effort is front-loaded. The payoff shows up gradually in team focus and in the quality of the customer conversations. We think it has been worth it. But the case for building is easier to make in retrospect than in prospect, which is worth being honest about.
Where this fits in the build-log arc
This is five months into publicly writing about what it is like to build AI systems in the margins of running a small drinks company. The CRM agent has been the most operationally significant build so far. Not the most technically interesting , that is probably the production annotation system we built for batch records. But the one that has changed the daily working pattern of the team most directly.
The full build-log, including earlier posts on the design decisions and the first months of operation, is at absolutionlabs.com/blog.
Frequently asked questions
What does a CRM intelligence agent actually do for a small drinks producer?
In practical terms, the agent monitors a list of trade accounts, flags the ones that have gone quiet based on order and communication patterns, and drafts the outreach email. The work it removes is the manual scanning of account histories, the mental overhead of remembering who went quiet when, and the first-draft writing time. The work it does not remove is the relationship itself. A produced draft email still needs a human to decide whether it reads right for this particular account before it goes.
How much maintenance does an AI agent require after it is built?
More than most people expect. There are configuration issues on a weekly basis: tweaks, improvements, edge cases, occasional glitches. You need to think about token spend and model selection regularly. New models are released frequently, and what was optimal three months ago may not be now. Building the agent is roughly 90% of the effort; maintaining it is the other 10%, indefinitely. Anyone who tells you a deployed agent simply runs is describing something they have not deployed.
What is the most important thing automation changes in trade account management?
The flow of information. With a CRM agent handling the data management layer, the sales team spends less time looking at spreadsheets and more time on phone calls and in-person visits. That shift matters more than the efficiency gain on any individual task. The relationship work cannot be automated. What automation can do is clear the path to that work.
How do you evaluate whether an AI agent is worth the build cost for a small producer?
The honest measure is whether it changes where human time goes, not whether it saves time in the abstract. If a six-person sales team is spending two hours a week on data housekeeping tasks that an agent now handles, that is two hours redirected to the relationship work that actually compounds. The build effort is front-loaded and significant. The payoff shows up gradually and is hard to quantify precisely.
Can you build agents to manage other agents?
Yes, and at some point you need to. When you have more than one agentic workflow running, the coordination and monitoring overhead starts to resemble the data-management overhead you automated in the first place. Meta-agents that check on other agents' outputs, flag anomalies, and surface decisions that need human review are a natural next step once you have a few production agents running.