How Should an Independent Ad Agency Adopt AI?

By Jeremy Swiller, ANA-Credentialed AI Training Instructor. Founder, Flux+Form. Published September 14, 2026.

Adopt AI in this order: name the one constraint actually costing you money or pitches, set a number that proves whether AI fixed it, redesign the workflow around that number instead of bolting a tool onto the old one, then govern what your team is already doing with AI whether you approved it or not. Skip the order and you get what most agencies have now: high AI usage and no measurable return.

Why does AI adoption fail even when everyone’s already using the tools?

Using AI and adopting AI are different projects, and most agencies only did the first one. RAND studied AI projects across industries and found more than 80% fail, roughly twice the failure rate of ordinary IT projects. The top cause, cited by 84% of leaders RAND interviewed, was leaders who never told their teams which problem they were trying to solve.

MIT’s NANDA initiative found the same gap from the other direction: 95% of organizations running generative AI pilots saw zero measurable return. The tools worked; the deployments didn’t, because most never adapted to how the team worked, and nobody measured whether they had.

McKinsey’s 2026 State of AI report puts a number on the divide: 44% of organizations say they’ve scaled AI, but only 37% report any measurable bottom-line effect. The organizations that saw results shared one habit: 73% had redesigned the workflow around AI, versus 25% who had simply dropped a tool into the process they already had.

It’s every industry buying AI in the wrong order: tool first, workflow second, measurement never. An independent agency needs a sequence, not a bigger AI budget.

Note: the RAND, MIT NANDA, and McKinsey figures above are about AI projects broadly, across industries, not agencies specifically.

What is the Constraint Sequence?

The Constraint Sequence is a four-step order of operations for deciding what to do with AI before you buy or build anything. It exists because the research above shows the same failure pattern every time: organizations skip straight to deployment and never define what success means. Skipping a step doesn’t save time. It just moves the failure downstream.

1. Name the Constraint. Before evaluating a single tool, name the specific thing costing you money, hours, or business right now: a thinning margin, a pitch you keep losing, a production step that eats a day it shouldn’t. “We should be using AI more” is not a constraint. “Concepting takes three days and clients won’t pay for day three” is one. RAND’s finding that 84% of failed AI projects trace back to leaders who never named the problem is why this step exists.

2. Set the Number. Decide, in writing, what result would prove the constraint is fixed, before anyone touches a tool: hours saved per project, pitches won, revision rounds cut, a margin percentage. McKinsey found organizations seeing real financial impact were twice as likely to have defined that measurement before scaling up. Most agencies do this backward, mandating usage first and looking for a metric afterward, if they look at all.

3. Redesign the Workflow. Rebuild the process around the number you just set, rather than inserting an AI step into the workflow you already have. This is the difference McKinsey found between its high performers and everyone else: the winners changed how the work moves, not just which software touches it. A concepting process that adds an AI draft but keeps every downstream review step unchanged has been decorated, not redesigned.

4. Guard the Edges. Write down who can use which tools on which client work, because your team is very likely already using AI without you. PagerDuty’s 2026 survey found 66% of office professionals had used AI at work despite believing policy restricted it, and more than a third had put client or financial information into public tools. This step is ongoing, and usually already overdue.

Run the four steps once per constraint, not once for the whole agency.

What should we tell clients who expect AI to shrink our fees?

Tell them what you’re actually pricing: the outcome, not the hours it took. Most agencies aren’t telling clients anything consistent, because they haven’t decided internally. Forrester and the 4As found 75% of US agencies absorb AI-driven efficiency as overhead, and just 6% bill it separately. The 4As’ own read: “This isn’t sustainable.”

Agencies making progress are changing what the fee is attached to, not just negotiating cuts. A Forrester Consulting study run with Dentsu Creative, surveying 356 US and Canadian marketing and procurement leaders, found 25% of North American agencies have moved fully to fixed-fee pricing, with 63% of those reporting satisfaction, and more than half of those not yet using it expressing real interest.

Set the Number before this conversation happens, not during it. An agency that can say “this workflow now delivers X in half the time, and here’s what we charge for X” is negotiating from a redesigned process. An agency still billing hours for an AI-assisted task is negotiating from the workflow AI already broke.

Should we worry about losing pitches to agencies that oversell their AI capability?

Less than you’d think, based on what pitch consultants are seeing. Rebecca Nunneley of AAR, a UK pitch consultancy, put it directly: marketers care “far more” about whether an agency uses AI responsibly and pragmatically than whether it claims to use AI at all. AI capability alone isn’t winning pitches. It’s becoming table stakes fast enough that claiming it has already stopped being a differentiator.

AI still matters to new business, just aimed at a different risk: a client asking a specific, informed question about your process and getting a vague answer, because you skipped Name the Constraint and never had to explain a workflow you actually redesigned. Digiday’s Q4 2025 survey found 73% of agency professionals believe their own clients don’t understand what agentic AI is. The agency that can explain its AI use in plain, specific terms is differentiating on clarity clients don’t get from anyone else.

What happens to the billable hour?

It keeps eroding, and the pressure already shows in the numbers. A 213-agency Basis survey, fielded in 2026, found 87.3% of agency professionals believe the traditional agency model is already broken or will be within three to five years, rising to 91.5% among VP-level respondents. The same survey found 99% of agencies now use AI, and 39.9% had conducted layoffs in the prior year.

A separate analysis, cited by MediaPost and drawing on a VoxComm and Lodestar Agency Consulting report, puts a number on the longer trend: average agency profit margins have fallen from roughly 30% to about 10% since advertising’s higher-margin era, even as the average creative produces close to five times the output for flat or lower pay. Agencies are selling more work for less money per hour, whether or not billing has caught up.

Redesign the Workflow exists to fix exactly this. An hourly rate assumes the hour is the unit of value. Once AI compresses the hour without compressing the value delivered, billing by the hour charges less for more, automatically. The fix is changing what you charge for, which requires the workflow to already reflect what AI changed.

How do we protect the junior talent pipeline?

Deliberately, because it won’t protect itself. A report cited by Forbes found 57% of agencies had slowed or paused entry-level hiring over the prior year as AI absorbed tasks once used to train juniors. Andrew Graff, CEO of Allen & Gerritsen, named the problem: “AI is in its infancy. We’re going to need a whole new set of skills,” meaning the job has to be redesigned, not preserved in its old form.

AI is in its infancy. We’re going to need a whole new set of skills.

Not every agency is pulling back. Salt XC, an experiential commerce agency, hired 100 people in two months, most entry-level, while other shops cut junior headcount. Both groups use AI. The difference is whether roles were redesigned around judgment work AI can’t do, or just shrunk because AI could do the old version faster.

The 4As has tracked agencies rebuilding this pipeline through apprenticeships, structured shadowing, and earlier real client responsibility, once the repetition that used to teach juniors is gone. Boris Dzhingarov of ESBO Ltd, writing for Forbes Agency Council, described new hires needing to “run a small real account within months,” not a year on execution tasks AI now handles. That belongs inside Redesign the Workflow: juniors learned the craft by doing the work, and if AI does that work now, the agency has to build a different way for judgment to get taught, on purpose.

What do we do about employees already using AI without a policy?

Assume it’s happening and write the policy around what’s true. PagerDuty’s 2026 Shadow AI Survey, fielded by Wakefield Research across 1,250 office professionals, found 66% had used AI at work despite believing company policy restricted it. Among that group, 34% had entered customer data into public AI tools, and 31% had disclosed financial or confidential strategy information. Thirty-nine percent said they’d keep using AI undisclosed, specifically to avoid restriction.

Shadow AI is employee use of AI tools that hasn’t been approved, reviewed, or disclosed to the organization. If your agency has no written AI policy, shadow AI is very likely your current, unmanaged state, on live client work.

AI governance is the set of decisions an organization makes about which AI tools are approved, what data can go into them, and who’s accountable for the output. Guard the Edges doesn’t mean banning tools your team already finds useful. It means writing down which tools are approved for which kinds of client data, who signs off on AI-assisted deliverables, and what happens when someone breaks the policy. Skip this step and the risk stays, only unmeasured.

How do we know if the sequence actually worked?

Check the number you set in step two, on the schedule you set it for. This sounds obvious, and it’s the step most agencies never reach, because most never wrote the number down. McKinsey found agencies seeing measurable financial results from AI were twice as likely to have leadership visibly committed and a defined process for tracking impact.

If concepting was the constraint and the number was “cut day three,” check whether day three is gone in three months, not whether the team likes the new tool. A number that wasn’t hit is a signal to go back to Name the Constraint and check whether you named the right one, not a reason to add more AI.

Where the pressure points show up, and what fixes them

Your next three moves

Pick one constraint this week, the one costing you the most right now, not the easiest to fix. Write down the single number that would prove it’s solved, and put a date on when you’ll check it. Then look honestly at who on your team is already using AI without a policy, because that part isn’t waiting for you to finish the other two.

The Creative Cadence Workshop walks a full creative team through this sequence live, if you want to run it together rather than alone. It won’t name your constraint. Only your P&L and pitch-win rate can do that.

Frequently asked questions