Outsource the Task, Never the Thinking

Harvard Business Review published a piece in June — AI Is Rewriting the Economics of Outsourcing by Abhinav Agrawal — that puts a macro frame around something I’ve been experiencing at the micro level for the past year. The argument: generative AI is automating the routine, rules-based work that companies spent three decades sending offshore for labor savings. The interesting part isn’t the automation claim. It’s the shift in the unit of analysis. The outsourcing decision used to be made at the function level — do we outsource IT, finance, HR? Agrawal argues it now has to be made at the task and workflow level: which activities can AI handle internally, which still require external expertise, and which just became more strategically valuable to keep in-house.

Agrawal compresses the whole shift into one line:

AI follows the work, not the org chart.

The org chart says who owns a function. The work says which tasks inside that function are routine enough to automate, rare enough to buy, or consequential enough to keep. AI only respects the second map.

I lead a web team at a large industrial company. I’ve spent twenty-plus years on both sides of this equation — as the in-house lead deciding what to send out, and earlier in my career as the agency resource that work got sent to. Reading the article, my reaction wasn’t “interesting prediction.” It was “yes, I have the receipts.”

The old math

For most of my career, the outsourcing calculation was genuinely simple. If work could be defined, standardized, and monitored, someone else could usually do it cheaper. That logic built the entire agency and offshore services industry, and it shaped how in-house teams like mine scoped our own work. The implicit org design question was always: what’s the smallest set of things we absolutely must do ourselves?

The tell was in how we bought services. When we needed a site replatformed, we didn’t hire for tasks — we ran an agency evaluation, compared statements of work, and handed over a function-sized chunk of the problem. The vendor’s margin lived in the gap between what the work cost them and what defined, standardized work was worth to us. That gap was the whole business model.

The migration that didn’t leave the building

Here’s my receipt. Over the past year, my team converted aging Drupal 7 sites to static Astro.js implementations — work I wrote about in Context Engineering: How I Learned to Stop Worrying and Manage the Smart Zone. Five years ago, that project had an obvious shape: scope it, bid it out, manage the vendor, pay well into six figures. It’s exactly the kind of work the old math said to outsource — well-defined, standardized, monitorable.

Instead, we did it in-house with AI-augmented workflows — the same shift from doing the work to directing it that I’d been practicing on personal projects, now at team scale — and saved over $100,000 against the outsourced alternative. Not because my team suddenly had spare capacity — we didn’t — but because AI collapsed the cost of the defined, standardized portion of the work. Template conversion, content migration scripts, redirect mapping, boilerplate. The exact category of task the labor-arbitrage model was built to absorb.

That’s the HBR thesis in one project. The work didn’t get cheaper because we found cheaper labor. It got cheaper because the routine layer of it stopped requiring much labor at all.

What the task-level view actually reveals

The uncomfortable follow-on — and the reason I think Agrawal’s framing matters more than the headline — is that once you analyze work at the task level, you can’t unsee it. Every function on your team decomposes into three piles:

Tasks AI now handles internally. The defined, repeatable layer. In our world: migration scaffolding, first-pass code review, documentation generation, ticket drafting. I’ve written about most of these individually — the daily Claude workflow, the self-organizing documentation system. Each one, in isolation, looked like a productivity tool. In aggregate, they’re a repricing of exactly the work the outsourcing industry was built on.

Tasks that still require external expertise. Not everything moves in-house. Deep specialization we touch twice a year — a security audit, a niche platform migration, legal review — still makes no sense to build internally. AI doesn’t change that math; if anything it sharpens it, because the vendors worth paying for are the ones selling judgment and accountability, not hours.

Tasks that just became more valuable to keep. This is the pile leaders underweight. When AI accelerates execution, the constraint moves upstream — to architecture decisions, context, institutional knowledge, and the ability to verify that fast output is correct output. That work was always in-house work. It’s now the in-house work, and treating it as overhead is how you end up with a team that can generate code quickly and can’t tell you whether it should ship.

The line I keep coming back to

In the context engineering post I quoted Dex Horthy’s rule for working with AI on hard problems: “Do not outsource the thinking.” I meant it at the scale of a single engineer and a context window. Agrawal’s article is the same rule at the scale of an org chart.

That’s the through-line worth naming. Whether the executor is a model or a vendor — and AI is increasingly both, the contractor nobody signed off on — the failure mode is identical: handing off work you haven’t decomposed, then discovering you no longer understand it well enough to evaluate what comes back. The dumb zone isn’t just a context-window phenomenon. Organizations have one too, and they enter it the same way — by outsourcing judgment along with execution and calling it efficiency.

The old outsourcing model got away with this because the feedback loop was slow and the work was stable. AI removes both excuses. When execution is fast and cheap, the quality of your task-level understanding is the only durable advantage left — and it’s the one thing you can’t buy back once you’ve let it atrophy.

What I’d actually do with this

If you lead a technical team, the practical move isn’t a reorg or a vendor purge. It’s an inventory. Take one function you currently outsource — or one you’ve never questioned keeping — and break it into tasks. For each: can AI do this internally today, does it require expertise we’ll never sustain in-house, or is it judgment work we’ve been undervaluing? Do this honestly and the function-level question dissolves. You stop asking “should we outsource X” and start asking “what is X actually made of.”

That’s a harder question. It’s also the only version of the question that still has a stable answer.

The economics changed. The rule didn’t: outsource the task — to a vendor, to a model, to whoever executes it cheapest and best. Never the thinking.