Home
Dubb

The AI Agent Settings Most People Skip

Rob

The AI Agent Settings Most People Skip

Most people set up an AI agent the way they set up a new phone. They sign in, poke at the defaults, and start using it. Then they wonder why the output is generic and why they still do everything manually.

The gap is almost never the model. It is a handful of settings that take an hour to configure and change the output for every task that follows. Here are the ones that matter, in the order they matter.

Start by understanding the two sides

Before any setting, understand that the platform has two halves and they do different jobs. One side records and hosts video. The other side is the agent that thinks, writes, connects tools, and runs scheduled work.

Confusing them wastes days. People look for automation inside the recording interface and for recording inside the agent. Once the split is clear, every menu makes sense.

Step one is access, not prompts

An agent cannot triage an inbox it cannot read. It cannot book a meeting if it has no calendar. It cannot clean a CRM it was never connected to.

So the first session is a connection session, not a conversation. Connect the email provider you actually send from, Gmail or Outlook, not a secondary address you rarely check. Connect the CRM where the real records live. Connect the calendar link you use for booking. Add the two or three other tools that hold the work.

There is a specific detail people miss here. The connection needs permission to act, not just to read. Read-only access produces an agent that can tell you what you should do and cannot do it.

Memories are the setting nobody fills in

This is the highest return setting on the list and it is almost always empty. Memories are the standing facts the agent should apply to everything: how you write, what you sell, who you sell to, what you refuse to claim, how you want prices discussed.

Fill it in deliberately. Ten good entries beat a hundred vague ones. Useful entries sound like: we never discount more than ten percent, our buyers are operations managers at companies between fifty and five hundred people, we do not use the word solution.

Then add to it as you go. When you correct the agent, ask it to remember the correction. The corrections are the most valuable data you will ever give it, because they are specific to you and no competitor has them.

Routines beat clever one-off prompts

A prompt is a thing you remember to do. A routine is a thing that happens. The difference is the whole reason to use an agent instead of a chat window.

The routines worth setting up first are the boring ones. Every weekday morning, summarize overnight replies and draft responses. Every Monday, list this week's stalled opportunities. Every afternoon, turn the day's recording into two clips.

Set the schedule, write the instruction precisely, and give it a defined output. A routine that ends with a summary you can skim is worth ten that end with a document you have to read.

Use the skill library before you write your own

There is a temptation to handcraft the perfect prompt for each task. That instinct produces one excellent prompt that lives in a notes app and is never used again.

Start from the library instead. Pick the skill closest to the job, run it once, then adjust. The advantage is not that the library is perfect. It is that a shared skill produces consistent output across everyone using it, and consistency compounds.

Only write your own when the job is genuinely unusual. Then save it as a skill so the work is not repeated.

The workflow step the agent cannot finish for you

This is the honest limit. Sequencing a multi-step campaign, deciding how long to wait between touches, and choosing branch conditions still needs a human decision. The agent can execute the steps, but someone has to decide the logic.

So write the logic first, on paper. How many days between emails. What happens if there is no reply. What happens if there is a reply. What counts as a qualified conversation. Once those are decided, the build takes minutes.

The mistake is building the flow before knowing the policy. You end up with an automated version of a process nobody agreed on.

Build a review step on purpose

Early on, have the agent prepare work and hold it. Read the drafts, correct them, and let it send. Do that for a week or two, then loosen the gate on the tasks it has earned trust on.

The review is not a permanent brake. It is how the agent learns your standard, and it is cheaper than the cleanup after a bad send.

The two settings that carry the most weight have their own walkthroughs: company information and building routines.

Frequently asked questions

How long does a proper setup take?

An hour for connections and memories, another hour for the first two routines. The rest grows with use, which is fine, because the value arrives with the first routine rather than the last.

What if the agent writes in a voice that sounds wrong?

Write three sentences in your own voice, put them in memories with an instruction to match that register, and stop describing yourself with adjectives. Examples work better than instructions about tone.

What if it gets something factually wrong?

Give it a source of truth and a rule about refusing. A knowledge base it can cite plus an instruction to say I do not have that rather than guess removes most of the risk. Then spot check the output that goes to clients.

If you only do one thing after reading this, fill in memories. Everything else gets better once the agent knows what business it is working for.

If you would rather spend your time selling than configuring, start a free 7-day Dubb trial.

The settings are the product

Company information, routines and integrations decide what the agent can do. Set them once and the answers stop sounding generic.

See Dubb AI

Related Posts

View all posts