When should a workflow stay human?
Five signs a workflow should stay human, and why a good implementer says it out loud
Five signals that a workflow should not be given to an agent — and why an implementer who never says so is telling you something about themselves.
Almost everything written about enterprise AI is about what to automate. The more useful question in most discovery calls is the reverse one, because getting it wrong is expensive in a way that doesn’t show up for months.
Here are the signals we look for.
The judgment is real, and nobody wrote it down
Some decisions depend on context that exists only in someone’s head: whether this client relationship can absorb bad news this week, which of two equally valid escalations to run first, how hard to push a renewal with a customer who’s quietly unhappy.
An agent asked to make that call will make it, confidently, against a rule nobody ever defined. The output will look like the work. That’s the problem: a wrong answer that looks right is worse than no answer, because nothing prompts anyone to check.
Being wrong is expensive, and being wrong is quiet
Two conditions, and it’s the combination that matters. High-stakes work with immediate feedback is a good candidate for an agent with an approval gate — the mistake gets caught before it lands. High-stakes work where the error surfaces a quarter later is not, because by then the agent has made the same mistake four hundred times.
The test is not “how bad is a mistake”. It is “how quickly would we know”.
It runs eleven times a month
Placing approval gates, defining confidence thresholds and designing an escalation path that works at 2:00 in the morning is most of the labour in an agent deployment. For a workflow that runs eleven times a month, that design work will never be repaid. You’ll still be maintaining it in three years, when the person who understood it has moved on.
The process is about to change anyway
Automating a workflow that’s under review is building a bridge to a road that’s being moved. Wait. The agent will be cheaper to build against the process you’re about to have than against the one you’re about to retire.
The human contact is the product
Some steps exist because a person doing them is the point — the call where a client feels heard, the negotiation where a relationship gets built, the difficult conversation somebody has to own. Handing these to an agent doesn’t save cost. It removes the thing the client is paying for.
Why an implementer should tell you this
Every one of these findings shrinks the engagement. That’s precisely why it’s worth telling you.
An implementer whose assessment concludes that AI pays everywhere has either not looked or is selling. In practice, most organizations we assess have two or three candidate workflows that should stay human, another handful that belong in a hybrid arrangement — the agent prepares the work, a person releases it — and a smaller set that are genuinely ready for autonomy.
That last group is where the return is. The other two are what makes the return believable, and they’re the reason the deployment is still running two years later instead of having been switched off in six months.
The middle is bigger than either end
Worth saying plainly, because the market frames this as a binary: the answer for most workflows is neither “give it to an agent” nor “leave it alone”. It’s one process with human steps and agent steps in it, and the handoffs designed rather than improvised.
Anything a customer sees, and anything with money attached, generally belongs there — the reasoning is in what an AI agent should own, and it comes down to the review being cheap and the mistake not.
Designing the handoffs in that middle is most of our workflow automation and AI work.
Want this applied to your business?
Thirty minutes with a principal turns the general case into your specific list: which workflows, what they’re worth, and in what order. The article can only tell you what we would look at; the call tells you what we found.