You don't install an agent. You apprentice it.
A couple weeks ago I showed someone a workflow I'd built. Call him Joe. He was stoked to try it, until I started asking questions. What happens first? What counts as an exception? Who makes the final call? How do you know when the work is done?
He was surprised. He thought he was trying software, as it used to exist. He didn't expect to explain his job first.
Well that’s fair. Twenty years of software trained him that way. Install Office, open a blank doc, start typing. The software works; you do the work. Type, copy, paste, repeat. Software as a Service (SaaS) polished the same deal: sign up, connect a few accounts, invite the team, then learn the quirks of someone else's UI. The product stays fixed. You morph around it. (Ask my wife, learning D365 right now. It's painful.)
Great margins. Mediocre productivity.
Agents break that covenant. The moment software starts doing the work, it has to understand the work. Not just the happy path. The weird exceptions. The source of truth everyone ignores. Who decides when the data conflicts. When a human takes back the wheel.
The model is the easy decision.
You rent cognition by the token. The value is in the harness around it: one workflow, end to end, wired to the right context and permissions, with checks, escalation paths, a cost ceiling, and a definition of done. Some of that is code. Some of it is judgment. None of it comes in the model box.
Look at any AI company's org chart and the same admission shows up. AI Success Engineer. Solutions Architect. Forward-Deployed Engineer. Different names for one truth: the product can't finish the job from across the internet.
Services came back with a vengeance. Two years ago the fashionable take was that McKinsey was finished. Cognition was a commodity now, so who needs the advice? Turns out everyone. An answer was never the end game. Getting a messy organization to agree on the work, encode it, and keep it running is the challenge. Palantir understood this a decade ago and got called a "services company" like it was an insult. Who's laughing now?
Customers stopped buying software. They started buying outcomes. And someone has to own the last mile.
The economics break again. Tokens already gave software a marginal cost again. Agents add deployment labor on top. That doesn't just change the cost per run. It changes the business model. Self-serve onboarding, effortless seat expansion, 90% margins: none of it fits cleanly anymore. The companies delivering value have accepted what they are.
Half software. Half services. Zero drop-in.
Prompting is straightforward. Picking a model is a commodity decision now. The scarce resource is turning a messy human process into something a workflow can run reliably, again and again. You have to know where the human breaks, where the machine breaks, what each run costs, and hardest of all: what the result is actually worth. Automating low-value work gets expensive fast. Automating high-value work compounds. Knowing which is which is the true job.
So the last mile to revenue is now an engineering problem. Engineering leadership just became a go-to-market function. Someone technical has to decide what deserves automation, how to measure it, what's allowed to fail, and who owns it when it breaks.
That was the turn with Joe. Not a better model. Not a cleverer prompt. A real conversation about what his job actually took.
Software drops in. Agents move in.
So if you've put agents into a real business: what took more work, the model or everything around it? I keep hearing the same answer. Everything around it.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
