Build it

Mostly script, a little AI: the hybrid workflows that actually ship

July 30, 2026 · 7 min read

There's a false choice in most automation advice: either write a rigid script or hand everything to AI. The workflows that actually survive contact with real business data are hybrids — a deterministic pipeline doing ninety percent of the work, with AI invited in at exactly the joints where judgment lives.

Why pure approaches fail

Pure scripts break on mess. A script is perfect while the data behaves. But real workflows have a spot where a human used to squint: the free-text field, the inconsistent product name, the email that could mean two things. Write rules for that and the rules become a haunted house of special cases.

Pure AI breaks on repetition. Hand the whole workflow to a model and you pay the token tax on every step — including the ninety percent that never needed judgment — while accepting probabilistic answers for arithmetic and file-moving, which is exactly where you want boring certainty.

The hybrid shape

Picture the workflow as a pipeline. Most segments are pipes: read the file, match the columns, compute the totals, build the output. Pipes should be script — free, instant, identical every run. But somewhere there's a joint: a point where the input is messy human language, or a call has to be made. Put a single AI call there, with a narrow question and a defined answer format, and the pipeline keeps flowing.

• Script: gather this week's incoming requests from the inbox export.
• AI joint: "Classify each request: quote, complaint, order change, or other. Answer with one word."
• Script: route each class to the right folder, update the tracker, draft the boilerplate parts of the response.
• Human: review the drafts that matter, handle the "other" pile.

A real one: prospecting decks that build themselves

The clearest hybrid I run is prospecting. Finding gaps in a customer's catalog — products they could carry but don't — used to be hours of manual comparison, followed by hours of building a PowerPoint to present the findings. Now it's a pipeline: scripts collect and compare the catalog data; an AI joint helps judge which gaps are genuinely worth a conversation (messy, context-dependent — real judgment); and scripts again generate the presentation deck automatically from the results. A day of prep became minutes, and every deck arrives grounded in evidence. It's one of the everyday wins on the Proof page.

Design rules for the AI joint

The hybrid pattern works precisely because the AI's job is kept small and inspectable:

• Narrow question, fixed format. "Answer with one of these four words" beats "what do you think?" — narrow answers can be checked by the script downstream. This is asking precisely, industrialized.
• Right-size the model. A narrow classification joint usually runs fine on a small, cheap model. Save flagships for joints with real depth — the token math rewards it.
• Let "unsure" exist. Give the AI an explicit escape hatch ("if unclear, answer REVIEW") and route those to a human. A joint that's allowed to say "I don't know" is a joint you can trust.
• Log everything. The script should record what went into the joint and what came out. When something looks off, you can see exactly what the model was asked and what it said.

Give the rules to the script, the judgment to the model, and the responsibility to the human. Every workflow I trust has all three, in that proportion.

Why this matters for a small business

Because it changes the economics of "we're too small for automation." Hybrid workflows are cheap to run (the AI touches only a sliver of the volume), cheap to build (AI writes the scripts too), and safe to adopt incrementally — you can automate the pipes first and keep a human at every joint, then delegate joints one at a time as trust builds. That incremental path is exactly how I went from knowing next to nothing about coding to running a stack of these in about a year.