Making AI Work: Start With the Problem, Not the Tool
Apr 2026 · 9 min
Many brands are running AI backwards.
They hear about a tool, find a use case to fit it, and ship something that looks like progress - a dashboard, a workflow, a demo that looks impressive. Six months later, nothing has changed in the P&L.
I've seen this pattern enough times. It's what happens when you start from the solution instead of the problem.
The irony is that the real AI opportunities - the ones that compound over time and actually move revenue, margin, or speed - are hiding in plain sight. But you only find them when you start by asking: why are we here?
Start with the problem, not the tool.
The credible AI work maps to a business constraint or a clear business opportunity you can name. Everything else is just noise.
Here are a few examples of what working backwards from business problems looks like
In 2024–25, my team at Mejuri started investing in authoritative content as part of our SEO strategy. This wasn't a trend play - it was a thesis about how search was going to evolve. We used SEMrush data to identify long-tail search terms with high intent and low competition. We built blog content that was text-forward, semantically rich, and designed for both human readers and crawlers. Before AEO or GEO were mainstream terms, we signed on with Profound AI to track our presence in AI-generated search results.
Result
Over 50% improvement in Google non-branded search impressions once the first wave of posts went live.
The failure mode I see repeatedly
Brands outsource SEO and AEO entirely, so no one becomes an internal SME. You get reports, not reasoning. When the landscape shifts - as it did with AI-generated answers and chat-based discovery - teams that never owned the fundamentals end up reconciling vendor dashboards instead of steering strategy.
Treat AI search readiness as a P&L lever: assign an owner, establish a baseline, run a pilot, measure, and iterate.
SEO isn't a black box. I know this because even in a leadership role at MakeSpace, I just decided to learn it - a few nights, a MozPro account, and some consistent execution. Became an SME quickly and unlocked meaningful value. Brought the same mindset to Mejuri, now applied to AI search.
At Mejuri we had a thesis that text-forward emails - the kind that read like a personal note — could earn better inbox placement than heavily designed HTML marketing assets. The creative looked quieter; the intent was clearer.
Result
50% improvement in CTR, with an experience that felt like a relationship rather than a broadcast.
The constraint was volume. You can't hand-write thousands of variants a week. LLMs made personalized content at scale tractable - with a human in the loop for strategy, tone, and QA so the brand voice stayed intact.
The problem statement came first. We used LLMs specifically to remove a bandwidth constraint - not to invent work for a new tool.
If you've worked with NetSuite, you know the pain. One of the most time-consuming workflows in any mid-size company is coding bank transactions: categorizing financial data into the right accounts, cost centers, and GL codes. We used semantic mapping - training a model on historical transaction patterns and account structures - to automate a substantial portion of that work. The model learned the patterns. Humans reviewed exceptions. The time savings were real.
The opportunity wasn't on any AI trend report. It was sitting in a painful, manual, rules-based workflow that consumed real headcount. That's where most of the non-obvious applications live.
Growth marketers spend hours manually scrolling ad libraries, screenshotting creative, and trying to pattern-match what competitors are doing. It's slow, incomplete, and almost impossible to do consistently across multiple brands and markets.
I built a workflow in Gumloop that pulls publicly available ads per brand from Foreplay's ad library and runs AI analysis against assumed goals (e.g. new user acquisition). No internal data required - just what's running in the wild. In one teardown, the gap between two competing luggage brands was immediately visible: one had gone UGC-heavy with almost zero discounting, the other had coordinated a 20% off promotion across four global markets simultaneously. Two completely different strategic bets - surfaced in under a day, from public data alone.
What AI uniquely enabled here wasn't automation of a manual task. It was synthesis at a scale no analyst team would attempt. Layer your own internal performance data on top and the signal gets sharper. The workflow is now a public template anyone can run.
Public template — run the workflow on Gumloop.
The opportunity wasn't in some proprietary dataset. It was sitting in plain sight - in the ads your competitors are already paying to run.
Practical Framework for AI Evaluation
The brands and operators who will build durable AI advantages share a few traits. They start from the business problem, not the tool. They establish baselines before they run pilots. They assign real ownership, not just enthusiasm. And they measure aggressively before they scale.
Work through each step — the durable AI advantage is in the discipline.
AI is not a strategy. It's a capability. The strategy is the business problem you're solving. Start there, and the right applications will become obvious. Skip that step, and you'll spend a lot of time and money on demos that look like progress.
Name the problem first. Then let the tool earn its place.