Most companies that call themselves AI-native are burning money on AI infrastructure. We are not. Here is what actually changed when we rebuilt operations around AI — and what surprised us.

BACKGROUND

AC Solutions has been profitable since 2019. We bootstrapped from zero, never raised a round, and built to $115–135K MRR serving 200+ enterprise clients across 30+ countries. In 2024 we made a deliberate decision to rebuild internal operations around AI — not add AI features to our products, but actually run the company on AI.

This is what we learned.

DECISION 1: DEEPSEEK OVER GPT-4 FOR INTERNAL REASONING

The first choice that made us capital-efficient was model selection. We use DeepSeek Chat as our primary agent model for internal work — strategic analysis, document drafting, market research, code review. It is not the most capable model on every benchmark. It is good enough for 80% of internal tasks at roughly one-tenth the cost.

The remaining 20% — synthesis, complex multi-step reasoning, final decision-making — goes to Gemini 2.5 Pro with extended thinking, and occasionally to DeepSeek Reasoner for high-stakes checks.

We stopped using GPT-4 for anything internal. The quality difference in our use cases is negligible. The cost difference is not.

DECISION 2: BATCH PROCESSING OVERNIGHT

We have a daily batch job that runs every night at 2:00 AM. It processes all the non-urgent AI tasks queued during the day: competitive monitoring, client health score updates, weekly brief generation, contract renewal risk flagging.

This sounds obvious. It is not how most teams operate. Most teams run AI tasks in real-time because it feels faster. In practice, the results of overnight batch analysis — available fresh at 8 AM — are more useful than real-time results that nobody reads because they arrive in the middle of a meeting.

Batch processing also allows us to use cheaper compute windows and avoid rate limits by spreading load across off-peak hours. Our overnight AI spend is about 40% of what it would be if we ran the same tasks in real-time.

DECISION 3: AI FOR INTERNAL PROCESSES FIRST

This is the one that matters most, and the one most companies get backwards.

We deployed AI internally — in our own operations — before we deployed it in any client-facing product. Our Hiring Assistant processed 200+ candidate applications in January without a dedicated HR manager. Our Financial Audit Bot flags P&L anomalies before the monthly close. Our CEO Coach runs decision frameworks before I commit to anything significant.

The reason this order matters: when you use your own tools, you find the failure modes before your clients do. We discovered that autonomous AI agents without human checkpoints produce confident-sounding nonsense at a rate that would be embarrassing in a client environment. We fixed this internally. Our clients never saw the broken version.

THE CPM NUMBER

From Q3 to Q4 2025, our blended CPM across client campaigns went from $491 to $847 — a 72% increase with the same team size and roughly the same campaign volume.

Three changes drove this: 1. We stopped running broad inventory buys and switched entirely to regulated-market-specific placements where competition is lower and intent is higher 2. We added frequency controls driven by player risk scores — players flagged as high-churn risk receive fewer but higher-value touchpoints 3. Overnight batch analysis identified underperforming creative variants by 2 AM; by 9 AM the next day, underperformers were paused

None of this required more headcount. All of it required better tooling and the discipline to actually use the output.

WHAT AI-NATIVE ACTUALLY MEANS

It does not mean every process is automated. It means every decision has an AI-augmented input before a human commits to it.

We still have humans making every significant call — pricing, hiring, market entry, client escalations. What changed is the quality and speed of the information those humans are working from.

The mistake we see in other companies: they automate execution and leave reasoning unchanged. We did the opposite — we augmented reasoning and kept humans on execution where stakes are high.

THE FAILURES WE ARE HONEST ABOUT

AutoGPT-style autonomous agents: we tried them in Q2 2024. They produced elaborate plans and executed none of them correctly without constant supervision. We abandoned the experiment after three weeks and two near-misses that would have sent incorrect client communications.

Knowledge freshness: our AI systems are only as good as the context we feed them. We are still solving the problem of keeping AI agents updated on internal state — which clients are churning, which team members are stretched, what the actual pipeline looks like this week.

We are building structured knowledge base pipelines. It is not elegant yet.

THE BOTTOM LINE

AI-native means accepting that most of the value comes from boring infrastructure decisions — model choice, batch scheduling, internal deployment before external — not from the most impressive demo.

We are profitable. We are smaller than companies that raised $10M to do what we do. We are faster on strategic decisions than we were two years ago.

That is what actually works.

→ 200+ enterprise clients run on the AC Solutions platform. acsolutions.ai