An AI marketing agency sells tools. Most go unused.
Capability stopped being the hard part somewhere around 2024. Everything difficult about this now happens after the purchase order, in the gap between a working build and a habit.
So the first fortnight here usually cancels something rather than recommending something.
Illustrative, and the shape is always the same. The two that work are the two somebody built a workflow around.
Four ways a tool becomes a line on a renewal notice.
None of them is that the software was bad. Most ai marketing tools work approximately as advertised, which is why explaining the failure rate needs a different story from the one the category tells.
- 01
Nobody owned the workflow it was supposed to replace
The tool was bought to fix a process that was never written down. With nothing to replace, it becomes an extra step beside the old one, and the old one wins because it already works.
- 02
The output was right often enough to be dangerous
Eighty per cent accuracy sounds excellent and is unusable without a review step nobody scoped. The team checks everything for a month, realises checking costs more than doing, and quietly stops opening it.
- 03
It was bought by someone who would never use it
A decision made a level above the daily work, evaluated on a demo rather than a Tuesday. The people who would have to change their habits were consulted after the contract was signed, if at all.
- 04
It solved a step that was not the bottleneck
Drafting got four times faster in a team whose actual constraint was legal approval. The saved hours went nowhere, the queue did not move, and the tool gets blamed for a problem it was never pointed at.
What earns its keep, what is oversold, and what needs a human.
An ai marketing strategy is mostly this list, argued out for your situation. Most ai marketing services will not publish a middle column, because the middle column costs them upsell.
- Research and synthesis before a brief is written
- First drafts against a real brief, then edited hard
- Repurposing one asset across every surface
- Reporting assembly, so the monthly read takes an hour
- Summarising calls and tickets into usable evidence
- Predictive lead scoring below a few thousand records a month
- Deep personalisation before the generic version converts
- Autonomous agents left to act without a review step
- Anything requiring clean CRM data you do not have
- Publishing anything under your name unreviewed
- Replying to customers as you
- Deciding what is worth making
- Any claim somebody could sue you over
The middle column is where most budgets go and it is not a permanent judgement. Several of those become sensible at a scale most teams have not reached yet.
What an AI marketing agency cannot promise you.
Every agency in the category rebranded within about eighteen months, which makes it unusually hard to tell who has changed how they work from who has changed their homepage. An AI marketing agency runs into the same four walls regardless, and they are worth agreeing on before a contract.
- 01
The constraint is almost never capability.
The models are good enough for most of what marketing teams actually need, and have been for a while. What stops the value arriving is that nobody owns the workflow, nobody scoped the review step, and the person whose habits had to change was not in the room. Buying more capability does not touch any of that.
- 02
Automation makes a bad process worse, faster.
If your brief quality is poor, automating production gets you more pieces built on poor briefs. Every efficiency gain multiplies whatever sits in front of it, including the mistakes. The unglamorous first move is usually fixing the process, and only then pointing anything automated at it.
- 03
Most of the savings will not show up as money.
The realistic outcome is a team that does more with the same headcount, not a smaller team. Anyone selling this as a headcount reduction is making a promise about your organisation that they are in no position to make, and it tends to poison adoption among the people who have to make it work.
- 04
Anything genuinely autonomous will eventually be wrong in public.
Agents acting without review will, given enough runs, send something incorrect to a customer or publish something you would not have approved. That is a question of when rather than whether, so the honest design keeps a human in front of anything with your name on it and accepts the cost.
What is genuinely achievable is a smaller stack, three workflows people actually use, and a team that produces noticeably more. That is a good outcome and it is not the one the category advertises.
Six workstreams, and the first one cancels things.
Marketing automation projects fail at adoption far more often than at implementation, so four of these six are about people rather than about software.
- 01
Weeks 1–2
Audit what you already bought
Every licence, what it was bought for, who logs in, and what happens if it is switched off tomorrow. This is the least popular fortnight of the engagement and it frequently pays for the whole thing, because the answer is usually that several tools can simply be cancelled.
Stack audit3 of 5 dormantUsed weekly by a named personkeepLogged into twice this quartercancelNever onboardedcancelDuplicates something you owncancelThe first saving is usually a subtraction, not a purchase.
- 02
Weeks 2–3
Find the actual bottleneck
Map how work really moves from idea to published, with the waiting time included. Almost every team automates the step that was easy to automate rather than the step that was slow, and the two are rarely the same. Marketing operations improves when the queue moves, not when a task gets faster.
Where the time goesthe queue, not the taskWaiting for approval21 daysWaiting for a brief9 daysActually drafting2 daysActually editing1 dayIllustrative, and the shape is common. Automating drafting saves two of thirty-three days.
- 03
Weeks 3–8
Automate three things, properly
Not thirty. Three workflows, each with a named owner, a defined review step and a written fallback for when the output is wrong. Workflow automation that nobody is accountable for is a tool purchase with extra steps, and it will be dormant by the next quarter.
Build criteriaper workflowNamed ownerrequiredDefined review steprequiredWritten fallback when wrongrequiredReplaces an existing steprequiredA build failing any row does not ship. It becomes shelfware otherwise.
- 04
Weeks 4–10
Get it adopted, which is the hard part
Sit with the people who have to change how they work, watch them use it, and fix the friction rather than writing documentation about it. Ai adoption fails in the gap between a working build and a habit, and that gap is closed by presence rather than by training decks.
Adoption checkweek 6Used without being remindedthe testOld process switched offrequiredSomeone can fix it without merequiredDocumented but unusedfailedIf the old way still runs in parallel, the new one has not landed.
- 05
Ongoing
Keep humans in front of anything named
A standing rule rather than a preference. Nothing publishes, replies to a customer or makes a claim without a person approving it. That constraint costs some throughput and removes the entire category of failure where a system embarrasses you at scale while everyone is asleep.
The standing rulenon-negotiableResearch and synthesisautomatedDrafting and repurposingautomatedAnything customer-facingreviewedAnything under your namereviewedThe review step is the product. The automation behind it is infrastructure.
- 06
Month 2+
Report on usage, not on capability
Which workflows ran, how often somebody overrode the output, how much queue time actually came out, and what got cancelled. A stack review every quarter where something is switched off, because software spend only ever grows when nobody is tasked with reducing it.
Monthly readleading · businessLeading
Workflows in daily use, override rate, queue time
Business
Output per person, licence spend
Shape of a report, not a forecast. Tools bought is not a line on it.
If the question is specifically about producing content rather than running the function, the pipeline has its own page.
Four ways to buy this. One of them starts by subtracting.
An implementation partner will configure a platform well. A vendor will sell you seats. A hire will eventually know your business better than any of us. The question is who is measured on whether the thing gets used in six months.
If you have already identified the three workflows and simply need them built, an implementation partner is faster and cheaper. This is for when nobody has yet decided which three, and tailored experiences, agents and scoring are all competing for the same budget.
Audit first. It frequently pays for itself.
The diagnosis starts by counting what you already own and who logs in. That is not a sales tactic, it is just where the money usually is.
Start here
Discovery
$2,500one-off · two weeks
The stack audit with a cancel list, a map of where time actually goes including waiting, the three workflows worth building, and an honest verdict on what should stay manual.
- Every licence, its owner, and whether anybody logs in
- Where the time really goes, with queue time separated out
- The three workflows worth building first, and why not the others
- A cancel list, which often exceeds the cost of this engagement
Most common
Ownership
From $5,000per month
The builds, the adoption work that makes them stick, the standing rule about human review, and a quarterly stack review where something gets switched off rather than added.
- Three workflows with named owners and defined review steps
- Adoption driven by sitting with the team, not by documentation
- A human in front of anything carrying your name
- Reporting on usage and override rate, never on capability
Automations built this way are the same ones behind the rest of the work here. The library is public.
One call. Real plan, not a pitch.
30 minutes. We talk about what's already working, who owns content today, and whether a fractional Head of Content is actually the right move. If it isn't, I'll say so.
Direct calendar
Book a 30-min intro call
No sales rep, no qualification form. You pick a slot, we talk.
Calendar busy?
Send a note instead.
One sentence on the bottleneck. I'll reply within 24h with a sharper next step.
Questions, ahead of time.
Which AI tools should we actually be paying for?+
Fewer than you are, in most cases. The audit usually finds two or three genuinely load-bearing tools and several that nobody has opened this quarter. I have no reseller relationships and no incentive to recommend a platform, so the answer often starts with cancelling something rather than buying it.Will this let us reduce headcount?+
Realistically no, and I would be cautious about anyone promising it. What happens in practice is the same team producing considerably more, with the tedious parts removed. Framing it as a headcount play also makes adoption much harder, because you are asking the people whose jobs are implicitly at risk to make the project succeed.Can AI agents run our campaigns without supervision?+
They can run parts of the mechanics reliably. What they should not do is anything carrying your name to a customer without a person approving it, because over enough runs something will be wrong and it will be wrong publicly. The design here keeps automation on research, assembly and drafting, and keeps a human in front of anything outward-facing.How do you decide what to automate first?+
By mapping where the time actually goes, including waiting time, and automating the queue rather than the task. Teams usually automate drafting because it is easy and visible, when the real delay is a three-week approval step. Speeding up two days of a thirty-three day cycle is a rounding error that feels like progress.How is this different from your AI content page?+
That one is about the content pipeline: what a model may draft and where the line falls inside publishing. This one is about the whole function, where the recurring failure is not capability but adoption, and most of what was purchased is dormant. If your problem is a publishing process, read that page. If it is a stack nobody uses, this one.We already bought a lot of this. Is that wasted?+
Some of it, and it is better to establish which quickly. The common pattern is two tools doing real work, one that would work with an owner and a review step, and two that should be cancelled at renewal. Working out which is which takes about a fortnight and frequently covers the cost of the diagnosis on its own.Do we need clean data before any of this works?+
For the predictive things, yes, and that is the honest reason predictive scoring disappoints so often at mid-market scale. For research, drafting, repurposing and reporting assembly, no. That is why the sequence here starts with the workflows that do not depend on your CRM being tidy, and treats the data work as a separate decision.What does this cost, and does it pay for itself?+
A two-week Discovery is $2,500 and includes the stack audit, which regularly identifies more in cancellable licences than it costs. Ongoing ownership starts at $5,000 a month. I will not model a return for you, because the honest inputs are how much queue time comes out and whether your team actually changes how it works.