Which Workflows Should You Point AI At First?

Somebody runs the prioritization session. A whiteboard, one representative from each function. The list that comes out is always recognizable: draft first-pass copy, summarize the research deck, triage the support inbox, clean the CRM data, speed up reporting. Everyone agrees. It goes into a slide, the slide becomes a roadmap, and the first three items get started within a month.
Every item on that list is a reasonable use of AI. What the list captures is the set of tasks that are easy to describe in a workshop to people from other departments, which overlaps only loosely with the set of tasks where AI would change the economics of the business.
Where is AI being pointed right now?
There is reasonable evidence on this, and it matches what many leaders would guess. Research from the US Census Bureau on how AI diffuses through firms found that among AI-using firms, the leading business functions are Sales and Marketing at 52%, Strategy and Business Development at 45%, and IT at 41%. At the task level, the leading uses of generative AI are writing, document analysis, and information search.
The same research found that 57% of AI-using firms confine it to three or fewer business functions.
Read those two findings together and a pattern comes out. AI is concentrated in the functions that produce documents, on the tasks that produce documents, in a small number of places per firm. That is a sensible starting point, because document-shaped work is where current AI models are strongest and where a demo is easy to build. It also has a property that nobody flags at the start: document-shaped work is not where the business is constrained.
Why do the visible workflows get chosen first?
Three forces push in the same direction, and none of them are stupid.
The first is demonstrability. A workflow you can show on a screen in a ten-minute meeting will beat a workflow you would have to explain, every time, regardless of value. The second is ownership. The tasks that come up in a cross-functional session are the ones people are comfortable volunteering, which in practice means tasks nobody is protective of and nobody is measured on. The third is risk. A pilot on a low-stakes workflow is easy to approve, because if it fails nothing happens, which is also the reason it does not matter much when it succeeds.
Underneath all three is a gap in who decides. Microsoft's 2024 Work Trend Index found that 60% of leaders worry their organization's leadership lacks a plan and vision for implementing AI. Direction, in the sense of choosing which work AI should touch and in what order, is a strategic decision that gets made by whoever was free on a Tuesday.
What separates a workflow worth starting from one that just demos well?
Whether AI can do the task is no longer the interesting question. For many tasks on many whiteboards, it can. The harder question is what happens to the time it gives back.
Four tests do most of the sorting. They take about fifteen minutes to run against a candidate workflow, and they can be run by the person who owns the work, with no consultant.
| Test | What you are asking | A candidate that fails it |
|---|---|---|
| The owner test | Does the hour it saves land in one identifiable person's week? | "Everyone spends less time on first drafts" |
| The critical path test | Is this step what the next step is waiting for? | Faster deck formatting on a deck that then sits three days with legal |
| The destination test | Is there somewhere for the extra output to go? | Tripling content volume when review capacity is unchanged |
| The consequence test | Does something downstream have to change if this works? | A summary nobody has agreed to read or act on |
A workflow that passes all four is less glamorous than the ones that came out of the session. It is a workflow with a queue in front of it, a deadline behind it, and one named person in the middle.
The critical path test is the one that eliminates the most candidates, and it is the one people resist. Saving forty minutes on a task that was never holding anything up produces a genuinely better day for the person doing it, and no change at all to when the work reaches the customer. That is a real benefit. It is just not a benefit that reaches the P&L, and if the program was funded on a return, someone is going to ask.
What about the workflows that look bad on paper?
The best candidates look unattractive at selection time, for structural reasons.
They sit inside one function, so they do not surface in a cross-functional workshop. They are frequently owned by someone senior enough that their time is expensive and protected, which makes them awkward to volunteer. They involve judgment more than production, so the AI model's role is narrower and the demo is less impressive. And they require something to change downstream, which means the work does not end when the AI is switched on.
Consider a marketing organization where the true constraint is the approval cycle. Campaign concepts wait on input from stakeholders who never see the brief at the same time. Pointing AI at concept generation doubles the number of concepts entering a nine-day queue. Pointing it at preparing decision-ready briefs, with options, tradeoffs and a recommendation, so that a stakeholder can respond in one pass, shortens the cycle. The second one is harder to pilot, less fun to present, and worth considerably more.
That second option also requires the three stakeholders to agree to a new way of reviewing. This is the part that gets left out of AI roadmaps, and it is why the sequence matters more than the vendor selection.
A useful habit when the list gets built: for each candidate, write one sentence describing what a person does differently on the Monday after it goes live, and who that person is. Candidates that cannot produce that sentence belong in a backlog, off the roadmap, and the sentence takes under a minute to attempt. The resources library has more on how the chain gets diagnosed in practice.
Direction is the third link, and it has no owner
Average Robot describes the distance between what an organization spends on AI and the commercial return it produces as the AI Profitability Gap™. The chain that explains it runs Investment, Usage, Direction, Skill, Reinvestment, Return.
Investment and Usage get measured, because finance tracks the spend and the vendor tracks the seats. Direction sits immediately after them and nobody measures it, because it is a decision and has no number attached. Every organization has made that decision, whether or not anyone recognizes it as a decision. The list on the whiteboard was the decision.
AR's working order is people before process before platform, and Direction is where the process question lands. Before you can redesign a workflow, you have to have chosen the right one, and the right one is defined by where a saved hour converts into capacity, revenue, or a cost that leaves.
If you want to pressure-test the list you already have, the AI Profit Ready Assessment will tell you whether Direction is where your return is stalling or whether the break is further along the chain. If the list needs rebuilding against evidence from your own workforce, that is what the AI Profit Partnership is built to do: it researches your own team and ends in a prioritized plan for the year ahead, with quick wins along the way. Either way, the useful move is to go back to the whiteboard list and run the four tests on it before the next AI product gets approved.
Take it with you
Download this as a PDF
A clean, branded version to read offline or share with your team.
For your role
For COOs
AI for COOs: Where the Saved Hours Go
AI was meant to take work out of the process and the steps look the same. Why saved time gets absorbed, and how operations leaders turn it into return.
Related reading
- AI Saved Your Team Hours. Where Did You Send Them?
AI frees up hours, then work expands to fill them and nothing reaches the P&L. Reinvestment is a leadership decision,…
- When The AI Workflow And The Real Workflow Don't Match
AI adoption stalls when new AI sits on top of old workflows. Here's why the real problem sits in the workflow itself,…
- Why Isn't AI Saving My Team Any Time?
Licenses are active and adoption looks fine, but the hours never show up anywhere. Here is where the time goes and ho…