Agentifying Your Business
Corporate mandates, strategy decks, and who actually ships agents
Somewhere in your company right now, someone is drafting a memo about “agentifying the business.” It will mention productivity gains, competitive advantage, and a phased rollout. It will not mention infrastructure. It will fail.
A decade ago it was “digital transformation.” Same slide decks. Same gap between mandate and execution.
The Strategy Deck
I recently saw an AI strategy deck intended for distribution across a portfolio of companies. Polished. Structured. Clearly AI-generated itself. Three pillars, a maturity model, phased rollouts by quarter.
The underlying argument had substance. The people behind it understood what they wanted. But the presentation undermined the message. You could see the skeleton of a real strategy buried under paragraphs of filler, every section following the same rhythm, the same slightly-too-perfect transitions. The irony was right there: they used AI to write about the strategy instead of to execute it. The deck described what agents could do for operations. Nobody had built one yet.
Strategy without tactics.
Why Mandates Fail
Most companies try to buy their way in. “Get us an AI platform, plug in our data, agents appear.” The vendor demo used clean data and a curated scenario. Your data is messy. Your scenarios are weird. Your ticket routing logic, escalation rules, attribution model: all custom. The platform handles none of it. The result is expensive shelfware, or worse, a “customization project” that costs more than building from scratch.
The next instinct is to hire. Five AI engineers, a new team, a mandate to “agentify operations.” But they don’t know the operations. The ops people don’t trust them. The agents work in demo, break in production, and the ops team goes back to spreadsheets. Domain knowledge matters more than AI knowledge, and mandates skip the domain knowledge entirely. “Agentify customer onboarding” assumes someone has actually mapped how onboarding works today. Usually nobody has.
And even when an agent ships, nobody budgets for what comes after. The agent drifts. Nobody corrects it when it’s wrong. Trust erodes. People go back to the old way and the initiative quietly dies. Human-in-the-loop isn’t a nice-to-have. It’s how agents learn your domain. Mandates treat agents as “deploy and done.” Working agents are “deploy and iterate forever.”
The Guy on the Factory Floor
The most interesting AI adoption I’ve seen has nothing to do with mandates.
A head of manufacturing engineering I know built his own purchase order system. Not a developer. A guy who runs a factory floor. He used ChatGPT to write it in Google Apps Script, connected it to his bill of materials so he could send accurate invoices. No AI team. No strategy deck. He had a problem and the tool was right there.
A friend in IT is building a knowledgebase for his team the same way. Operations teams are doing their own context engineering with Claude. None of these people waited for permission or a “phased rollout.” They needed a problem worth solving and a tool that was accessible.
People closest to the problem, building for themselves. Not a centralized team building for someone else based on a requirements document that was already outdated when it was written.
Two Years, Not Two Quarters
Two years ago, in the GPT-3 era, we were heavily engineering every workflow. Hand-crafted prompts, brittle pipelines, constant babysitting. I burned plenty of tokens on workflows that went nowhere before anything stuck. Over time the ones that survived became self-improving: agents that corrected their own output, harnesses that learned from human feedback. Now we’re building with agent SDKs and composable skills.
That progression took two years of daily iteration. You can’t skip it with a memo.
Nobody declared “we’re agentifying.” An engineer was drowning in 2,000 daily errors and built a triage agent. Someone asked “how much does it cost to serve each client?” and a BI pipeline appeared in a day. A security rollout was painfully manual, so the API got wrapped in skills. Same harness, same patterns, different domains. The infrastructure made each new agent cheap. By the time someone asked “when are we going to agentify?” we already had.
The thing mandates skip is that the first agent is a bespoke project no matter what. The value isn’t in agent number one. It’s in the harness that lets you build agent number two in a day instead of a month. Without that, every agent is a standalone effort that doesn’t compound. If someone needs a business case and ROI projection to build a 200-line script that saves two hours a week, you’ve already lost. (The economics always work at the small scale. The compound effect is what works at the large scale.)
The Roads Come First
Tobi Lütke told Shopify employees to use AI before requesting headcount. That sounds like the kind of mandate I’ve been arguing against. But look at what came before the memo: no cost limits on AI tokens, 24+ MCP servers wrapping internal systems, prompt libraries, usage leaderboards, leaders visibly coding with AI. They didn’t mandate agents. They mandated access and removed friction. Agents emerged. Non-engineering teams ended up using AI tools more than engineering.
The memo was the last step, not the first. Shopify spent years building the roads. The memo just said “drive.”
Most companies are issuing the memo without the roads. Build one agent that solves one real problem. Let the person closest to the problem build it. Make the second one cheaper than the first. The strategy will emerge from what works.
What’s actually working in your org: the mandate or the people ignoring it?