AI Doesn't Give You Leverage. It Multiplies the Process You Already Have.
Nearly every agency now uses AI, and the most common operational complaint is still inefficient process. The two facts explain each other: the tools multiply the process a firm already has, and most firms have less of one than they think.
If you are tired of being told you are using AI wrong, that is a reasonable place to stand. The seats are bought. The team drafts with it every day. You have watched a demo turn a week of work into a morning, and watched your margin stay exactly where it was.
Between those two facts sits a suspicion that the missing piece is a better prompt, model or vendor. The evidence points somewhere else. AI does not create leverage on its own. It multiplies the process a firm already has, and most firms have less process than they think.
Agencies naming inefficient processes as their biggest operational challenge, the most common answer in the survey, in the same year AI use passed 99 percent of agencies. Siloed systems came second at 40.4 percent. Adoption stopped being the question. Process is.
The original version of this piece was built on a long interview with the owner of a done-for-you agency running about fifteen accounts in a single construction-adjacent vertical. His stack works. Onboarding produces a complete brand and campaign foundation from a fixed sequence of prompts run against an intake form, the client's site and their customer list. Building a campaign went from about a week to half an hour.
Then the part that matters. Before any of it, he spent six and a half years on the phone, averaging around 250 dials a day. He can describe the contact sequence he later automated hour by hour, because he ran it himself. The machine did not discover his process. It inherited one. As he put it, the AI did not make him good at follow-up; it let him stop personally being the follow-up. Only one of those is available to a firm that never wrote its process down.
The obvious objection is that well-run firms adopt technology well, so process discipline may be a symptom of competence rather than a cause. One clean test exists. In a randomized field experiment published in the Quarterly Journal of Economics in 2013, Nicholas Bloom and colleagues gave free management consulting to a randomly chosen group of large Indian textile plants and compared them with controls. Adopting structured practices raised productivity about 17 percent in the first year.
The detail worth keeping: computer use went up afterward, because the new processes demanded data that paper could not supply. The process came first and pulled the tools in behind it. Asked why the plants had not adopted profitable practices on their own, the researchers pointed mainly to information. Owners did not know what good looked like.
McKinsey's survey data points the same way at the scale of large companies. In its March 2025 State of AI report, fundamental workflow redesign had the biggest effect on whether organizations saw bottom-line impact from generative AI, out of 25 attributes tested. Only 21 percent of adopters had done it. By its November 2025 edition, just 39 percent of respondents attributed any enterprise-level profit impact to AI, and the roughly 6 percent it classed as high performers were far more likely to have redesigned workflows.
| Use | Share of agencies | What it takes to change anything |
|---|---|---|
| Ideation | 86% (2025) | Nothing. A strategist edits everything |
| Research | 72% (2025) | Nothing. Same workflow, faster input |
| Media planning | 29.1% (2026) | A written account of how the judgment is made |
| Media buying strategy | 22.1% (2026) | The same, for the decision closest to the client outcome |
Basis Advertising Agency Reports, 2025 and 2026, vendor-run surveys. Adoption is highest where no workflow has to change and lowest where automating would first require writing down how the work is done.
Agencies are not underusing AI. They are using it precisely where using it changes nothing.
One interview is an illustration, not proof. A successful operator's account of why he succeeded is filtered through survival and assembled after the fact. The owner and firm are kept anonymous by agreement.
The randomized evidence comes from textile plants, not agencies. It supports the order of operations; how far it carries into creative services is a judgment.
McKinsey's model is correlational. Workflow redesign was the strongest of 25 attributes, and together they explained about a fifth of the variation. It is the best single lever on the list, not a guarantee. Basis sells automation software to the agencies it surveys.
Writing a bad process down buys a machine that runs a bad process faithfully. Documentation is necessary, not sufficient. The Bloom result is about good practice, not any practice.
- If your best strategist left tomorrow, nobody could reconstruct how they work — tacit expertise transfers to nobody, not a hire and not a machine.
- Your AI use sits in drafting, and you cannot name one workflow it changed — layering on the old process rather than redesigning it.
- Two people do the same task differently and both count as correct — that is not a process; it is two habits sharing a job title.
- Client contact between deliverables happens when somebody remembers — the follow-up gap, pointed at accounts you already won.
- You cannot say what a tool changed in a number you tracked before buying it — without a baseline, no AI spend can be argued about in either direction.
This week, have the person who does it best write down the one process that makes your firm money, well enough for a competent stranger to follow.
Pick one: new business, onboarding, or account renewal. Not the operating manual, just the sequence that turns attention into revenue. Write it at the level where someone capable, but new, could run it without asking a question. Then hand it to a colleague who has never run it and have them try once while you watch and note every question they ask.
- What it costs
- A day of your best person's time, which is the most expensive day in the building, plus an hour of someone else's to test it. Expect the first draft to be shorter than the real process by half.
- How you'll know
- The number of questions the test run needs. Rewrite until it needs none. Only then is there a process for a tool to multiply, and only then is it worth measuring one number for thirty days before buying anything.
Being a good strategist is not hard because it is clever. It is hard because it has to be repeated, for the next account and again next quarter for all of them at once, by a tired person. That is exactly where a written process and a patient machine earn their keep, and exactly where most firms have neither.
Related from Sound Decisions: Creative Is Now a Line Item · You Don’t Need More Clients. You Need Longer Ones. · They’ll Pay for the Work. They Won’t Pay for the Thinking.
One conversation. The process that makes you money.
Not a tool recommendation. A second pair of eyes on the process you wrote down, and whether it is the right one to multiply.
Get a free positioning review Or read the frameworks →