If you’ve ever opened a blank doc and felt that small wave of dread—welcome to my old mornings.

I run Hillflare, a small growth marketing agency in Monterrey. For years, “scale” meant adding people or buying software that looked great in demos and heavy in real life. We shipped good work, but it always required pushing a big, creaky machine. Then I tried a different approach: instead of hiring another layer, I gave myself a few reliable AI “teammates” that handle the first pass. I (and the team) still make the final call. The result isn’t sci-fi. It’s calmer days and faster loops.

That’s what I mean by AI-native. Not “we use AI sometimes,” but “we operate with AI by default.”

The morning it clicked

8:05 a.m. A client pings: “New DTC skincare bundle. Mexico, 25–34F. Need fresh angles today.”

Two years ago that message would have kicked off a mini circus: a briefing doc, a call to “align,” a copy deck, comments on comments, a follow-up call to align on the alignment. You know the dance.

Now I paste the brief into our workspace and press go.

At 8:07, my research buddy (yes, an agent) drops three directions on my desk: habit stacking, routine over miracle, and first-week progress. No drama. Just raw material.

At 8:12, the copy buddy drafts twenty hooks, six headlines, two short UGC scripts, and—this part makes me smile—suggests a clean naming convention so the files don’t disappear into neverland.

At 8:18, the reviewer buddy runs through our checklist: tone, claims, localization. It flags three lines that could get us in trouble and one phrase that sounds weird in Mexico. (It’s right.)

At 8:25, the analytics buddy peeks at last month’s results. It reminds me that “routine/progress” beat “derm-backed” for this audience. It proposes a simple test split.

At 8:40, I take my human pass: tighten two hooks, swap a CTA, approve the plan. The workflow opens the tasks, pings design for the storyboards, and logs the experiment in our playbook.

By 10:00, the first batch is live. Not because we cut corners. Because that heavy first 70%—the part that eats time and morale—is now quick and consistent, and we spend our energy on the last 30% where taste and judgment live.

That’s the edge.

What “AI-native” actually means (to me)

It’s simple: agents start, humans finish, workflows repeat.

Agents do research, first drafts, summaries, and options. Humans decide, rewrite, approve, and ship. The workflow makes the handoffs clear: who owns what, when it’s “good enough,” and where the result lives so the next person—and the next agent—can learn from it.

When you operate this way, meetings shrink. People stop arguing about hypotheticals because getting a first version is cheap. You replace “we should try X” with “we tried X this morning; here’s what happened.”

It’s a mindset before it’s a stack.

What changed for us (and felt great)

Less meeting, more making. We moved from “alignment” to “experiments.” Instead of debating angles for an hour, we test five by lunch. That single shift turned down the temperature across the team.

Small team, bigger output. We didn’t add a department to launch a new service. We added a workflow. One PM plus a few agents now handle what used to require a small squad. Nobody misses the scheduling Tetris.

No more tribal knowledge. Prompts, decisions, and results get logged. Last quarter’s winners are a click away. When someone new joins, they learn our voice by reading—and hearing—the best of what we’ve shipped.

What stayed human (on purpose)

Voice and taste. Models can write. They don’t know the brand like you do. The last pass belongs to us, and that’s where the soul lives.

Claims and context. Anything sensitive gets a human pass. Not bureaucracy—trust.

Direction. Someone still chooses why we’re doing the thing. AI is a fast car. It still needs a driver and a map.

A week in the life (real loops, not theory)

Monday – Lead triage.

Morning inbox is a mix of contact forms, referrals, and mystery emails. The agent reads them, tags ICP fit and urgency, drafts a reply, and proposes a time. I edit in sixty seconds and hit send. The difference is subtle: fewer “sorry we missed your note” moments, more qualified calls on the calendar.

Tuesday – Creative testing.

We’re building a new angle library. The agent proposes three directions and scripts for each. I pick five that feel human and local, then send them to design with tidy file names (a small miracle). We launch variants, review quietly in the afternoon, and log what worked.

Wednesday – Landing page tune-ups.

The agent pulls scroll maps, top exits, and queries from search and site search. It writes a short “fix first” list. We tweak two headlines, swap social proof, and move the CTA above the fold for mobile. The next morning brings a tiny bump, which we note and move on.

Thursday – Sales enablement.

For B2B clients, the agent turns last week’s best content into a short outreach pack for reps: two personalized intros, a case summary, and a “what to ask” list. Reps appreciate not starting from a blank screen. Wins compound when you stop reinventing the opener.

Friday – One-pager insights.

The agent compiles the last thirty days: spend, CPA, winners, losers. It writes a one-page “what to do next” in plain Spanish. I add three lines that clients actually care about. People read it because it answers the questions they already have.

None of this looks flashy. It looks like work that moves.

The stack we actually use (kept boring)

I “vibecoded” a small data warehouse for our paid media results. Nothing heroic—just clean names, clear owners, one place to check the truth. We have a few tiny connectors so agents can read the brief, pull past winners, open tasks, and peek at performance before making a suggestion. For glue code I use a coding copilot. If a tool adds friction or confusion, we drop it. Boring scales. Boring also sleeps at night.

The guardrails that saved us

We started with assist, not autonomy. Agents earned permissions. We set confidence gates: if the agent is unsure about a claim, a translation, or a sensitive topic, it must tag a human. There’s one owner per workflow—a name, not a committee. And we log everything. Not for compliance theater, but because future-us deserves an easier week.

Skills that matter now

You don’t need a PhD to work this way. You do need to frame problems clearly (“done looks like X, avoid Y”), keep your data tidy (agents mirror your mess), and develop taste. Taste sounds vague, but you know it when you read a line and think, “That actually sounds like us.” The rest is orchestration: giving the right agent the right input with the right guardrails.

Where to start (this week, not someday)

Pick one loop you touch every day—lead triage, creative testing, or monthly insights. Write a one-page SOP that names the goal, inputs, what agents do, what humans do, and the final owner. Run it on a real task. Fix what breaks. Track four numbers: time to first draft, percent auto-approved, errors caught at review, and cost per deliverable versus your old way. Do it again next week.

Do not try to “AI all the things.” Resist tool collecting. One connector, one spreadsheet, and two agents will get you 90% of the way. Call that progress and move on.

Mistakes I made (so you don’t)

I went full auto too early and had to roll back. I let messy data linger and watched the agents mirror the mess. I added too many tools and ended up giving demos instead of getting results. And I wrote “someone owns it,” which meant no one did. The fix was simple and a little humbling: fewer tools, cleaner data, one owner per workflow, and a clear definition of “done.”

The edge, in one sentence

Agents start, humans finish, workflows repeat. Do that across a few high-rep processes and a small team starts to feel big—without turning into a meeting factory.

If you try this, tell me what worked and what didn’t. I’m still learning too. And when I open a blank doc now, I don’t dread it. I press go, sip my coffee, and get to the part only I can do.