We score every workflow stage against five factors — volume, repeatability, verifiability, cost-of-error, and reversibility — to decide where AI belongs in a workflow at all. That’s our AI Workflow-Placement Model, not an industry law: it’s the policy AlexDesigns runs, and applying it to a typical marketing workflow currently puts AI in six stages (research, synthesis, prioritization, content operations, QA, reporting) and keeps three off-limits (strategy, claims, final approval). A stage only earns AI budget once it clears a four-question gate that checks the decision it speeds, the data it reads, and who’s still accountable for the result.
This is a different question from whether a capability already living in a stage should be allowed to act without a person checking its output first — that’s a separate, narrower scoring model (reversibility, cost of error, brand risk) covered in who gets to decide: drawing the line between AI and human authority in marketing. This article decides where AI gets to operate; that one decides how much it’s trusted to do on its own once it’s there.
Most teams get this backward. They pick an AI platform first, then go looking for somewhere to use it, and that order almost guarantees the tool ends up bolted onto the workflow instead of built into it: a chatbot nobody asked for, a content generator nobody reviews closely enough. The fix isn’t a better platform search. It’s starting from the workflow you already run, finding the step that’s actually slow, and only then asking whether AI belongs there and whether it’s earned the budget to try.
Is this actually where your team is right now?
Mapping AI onto a workflow only helps once you have a real, repeated workflow to map it onto, so check that before evaluating any platform.
This isn’t your priority yet when:
- Your team doesn’t have a consistent, repeatable process for the work in question (content, reporting, prioritization). Standardizing the process comes before automating any part of it.
- You’re evaluating AI to replace a judgment call: strategy, a public claim, final approval. No workflow map changes that answer; see the always-human stage below.
- Nobody on the team can be named as the reviewer of AI output for the step you’re considering. That gap needs to close before a pilot starts, not after it’s already running.
If none of those apply, mapping the workflow below is the right next step.
What decides where AI belongs? AlexDesigns’ five-factor placement model
Before mapping stages, name the model doing the deciding. This is our AI Workflow-Placement Model — a policy we apply, not a rule handed down from outside. We score any workflow decision against five factors:
- Volume — how many decisions of this type happen. High-volume, repetitive calls are the ones worth automating first; a decision made twice a year rarely justifies the review overhead.
- Repeatability — how similar the decisions are to each other. AI is reliable on a pattern it’s seen before and unreliable on a one-off judgment call with no precedent.
- Verifiability — whether a human can check the output cheaply. If confirming the answer takes as long as producing it yourself, AI hasn’t saved anything.
- Cost-of-error — how bad a wrong call is. A miscategorized backlog item is cheap to fix; a false public claim is not.
- Reversibility — how easily a bad call gets undone. A draft nobody’s read yet is fully reversible. A published claim, a client-facing recommendation, or final sign-off often isn’t.
Score a decision high on volume, repeatability, and verifiability, and low on cost-of-error — reversible mistakes — and it’s a strong AI candidate. Score it low on volume or repeatability, and high on cost-of-error and low on reversibility, and it stays human, regardless of how capable the tool is. That scoring is what produced the six-stage / three-stage split below for a typical marketing workflow — it isn’t the split itself that’s the rule, it’s the five factors underneath it, and another team could legitimately score its own decisions differently.
Which stages does that scoring put AI in, and which stay human?
Running a typical workflow through the five factors lands AI in six stages, consistently, across the workflows we’ve built it into: research, synthesis, prioritization, content operations, QA, and reporting. Three decisions score low on reversibility and high on cost-of-error regardless of volume, so AlexDesigns keeps them human: strategy, claims, and final approval.
A strategist deciding which segment to prioritize this quarter is making a call that should reflect the business’s real risk tolerance and relationships, not a model’s best guess. A claim about a result needs someone who knows whether it’s true and sourced. Automating either one doesn’t save time; it moves the risk downstream to whoever notices the mistake later — and a mistake at that level is expensive and hard to walk back, which is exactly what the cost-of-error and reversibility factors are built to catch.
What’s the difference between AI built into the system and AI bolted on top?
A two-column paired comparison across three named checks, so a reader can tell whether a proposed AI use case is genuinely built into the workflow or just added on top of it.
| Row label | Built into the system | Bolted on top |
|---|---|---|
| Where it lives | inside an existing workflow step | a separate tool nobody's process touches |
| What it reads | the team's own data and prior decisions | a generic prompt with no context |
| Who's accountable | a named owner | "the AI's answer" |
Built into the system
Bolted on top
Three checks separate the two, every time: where the tool lives, what it reads, and who’s accountable for the output. Where it lives: inside an existing workflow step, not a separate tool nobody’s process touches. What it reads: your own data and prior decisions, not a generic prompt with no context. Who’s accountable: a named owner on your team, not “the AI’s answer.” A proposed use case that fails more than one of these isn’t ready to build, not because the technology can’t do it, but because the accountability and the data connection aren’t there yet.
Optimizely’s own Opal is a useful reference point for what “built into the system” actually looks like inside one of these six stages. Per Optimizely’s own documentation, Opal carries program memory of a team’s existing experiments, metrics, feature flags, and history, and its agents can run autonomously from idea through build, run, pick-a-winner, and push to production. That’s the built-in pattern: the tool reads what the team already knows and sits inside a step (experiment prioritization and QA) the team already owns, rather than starting from a blank prompt with no context. Opal is a fast-evolving product category as of this writing, so treat the capability set as current, not fixed.
How do you decide whether a stage is ready for an AI pilot?
The article's own four-question budget gate, walked through in order.
- What decision does this actually speed up?
- Does it connect to data you already have, or does it want a new data set fed to it?
- Who reviews the output, and how long does that take?
- What happens to output quality if you turn the volume way up?
A stage clears only when all four have a specific answer; "It depends" on any one means the stage isn't ready.
Passing the built-in-vs-bolted-on checks tells you a stage is a plausible fit. It doesn’t tell you the tool is worth the budget. Before any AI tool gets funded for a stage, run it through four questions, in order:
- What decision does this actually speed up? If you can’t name one, it’s a feature looking for a use case, not a pilot worth funding.
- Does it connect to data you already have, or does it want a new data set fed to it? A tool that requires a parallel data set is adding a system, not simplifying one, and it fails the “reads your own data” check above before it fails this one.
- Who reviews the output, and how long does that take? If the honest answer is “nobody,” that’s the same accountability gap the three checks already flagged. Name the reviewer or don’t fund it.
- What happens to output quality if you turn the volume way up? If quality holds, the tool is a real accelerant for that stage. If it doesn’t, you’ve found the ceiling before you paid for it.
A stage clears the gate only when all four have a specific answer. “It depends” on any of the four means the stage isn’t ready yet, no matter how capable the platform demo looked.
Alex’s Perspective
Across 100+ optimization and experience programs, and hands-on work with 25+ marketing, analytics, experimentation, personalization, and AI technologies, the AI pilots that actually stuck were never the ones bought first and mapped to a stage later. They were the ones built to speed a step a team was already doing by hand, reading the team’s own data, with a named person still signing off on the result. The ones that got quietly abandoned within a quarter were almost always bought backward: platform first, workflow fit and the four-question gate treated as paperwork to fill in after the contract, not before it.
Should we pilot AI on one workflow stage at a time, or several at once?
Just one. Piloting a single stage lets you tell clearly whether the tool actually helped that specific decision and whether the review burden dropped or just moved somewhere else. Piloting several stages at once makes it much harder to attribute a result, or a failure, to any one change, which is exactly the ambiguity the four-question gate exists to remove before budget moves.
Frequently Asked Questions
What if we don’t have anyone who can own the review step?
That’s the actual gap to close before piloting anything, not a reason to skip the review step. If no one on the team can realistically review an AI tool’s output, the honest move is to name that person first, even part-time, or wait, rather than run AI output through a stage with no accountable reviewer.
How do we know if an AI pilot actually succeeded?
Check whether the specific decision it targeted got faster or better, and whether the named reviewer’s workload actually dropped, not whether the tool “feels” more efficient. A pilot that speeds up drafting but adds an equal amount of new review time hasn’t reduced the bottleneck it was meant to fix.
What’s the earliest sign a stage is drifting toward hype instead of leverage?
Output volume increasing without a proportional increase in who reviews it. If a draft queue or a recommendation list is growing faster than the named reviewer’s capacity to check it, that’s the same gap question three already tests for, and it shows up well before anyone notices a quality drop.
What to do next
- Map your current content or campaign workflow end to end, and mark the one stage that’s genuinely the bottleneck. That’s where to look for AI first, not the flashiest use case.
- For that stage, run the three built-in-vs-bolted-on checks, then the four-question budget gate, in that order, before you evaluate a specific platform.
- Write down who reviews the output and how long that review takes. If the honest answer is “nobody” or “we haven’t decided,” that’s the gap to close before piloting anything.
Where this fits
This spans three stages of the framework, not one. Prioritize decides which stage is actually worth automating first and whether it clears the budget gate. Experiment pilots AI on one stage before rolling it out everywhere. Learn checks whether the review burden actually dropped, or just moved. Treat the gate as a standing practice across those three stages, not a one-time platform decision made once and never revisited.
Alex Harris leads AlexDesigns, working with a specialist network across strategy, engineering, and AI to build systems clients own rather than rent. This one comes from watching more AI platform purchases get shelved from a bad workflow fit and a skipped budget gate than from the AI itself falling short. Map the bottleneck and run the four questions before you shop. Last reviewed 2026-08-22.
If you’ve mapped your workflow and want a second opinion on where AI actually fits, and whether a stage is ready to clear the budget gate, before you shop for a platform, start with a free assessment and we’ll help you find the real bottleneck.
