I went looking for a receipt last spring and ended up counting subscriptions instead. Eleven line items, all of them some flavour of AI tools for agencies, all of them charging us every month. Two I could not remember signing up for. One nobody on the team had opened since the trial.

That was the whole moment. No epiphany, no whiteboard session. Just a slightly embarrassing number on a card statement and the realisation that I could not point to a single thing in the business that had actually changed because of it. If you run or work inside an agency, that gap will feel familiar: plenty of AI apps helping with isolated tasks, very little showing up in margins, throughput, or the way the agency actually operates. This piece is about that difference between buying AI tools and getting AI-driven change, where agency workflows quietly absorb hidden costs, why tools without retained client and company knowledge keep resetting to zero, and what to look for if you want systems that compound instead of another subscription.

Buying AI tools and being changed by AI are two different purchases

Here is the part that took me a while to admit. The tools were fine. They all worked. The writing assistant wrote. The transcription tool transcribed. The research thing found things. Nobody was being scammed.

What none of them did was know anything.

Every morning, every one of those tools woke up with no idea who we were. I would open the drafting tool and start by explaining the client. Then the brand voice. Then the thing we tried in March that did not work and should not be suggested again. Then I would get my draft, close the tab, and tomorrow do all of it again from scratch.

That is not a business running on AI. That is a business paying eleven monthly fees for the privilege of repeating itself.

Nobody logs that time. There is no line on a timesheet called “explaining the client to the software again”, and if there were, most of us would not want to see the total. It gets absorbed into the general texture of the day. You feel busy and productive and slightly behind, which is how agency people have felt since agencies were invented, so nothing about it registers as broken.

The thing marketing agencies actually own

Strip an agency back and what is left is not the deliverables. Anybody can produce deliverables. What is left is accumulated judgement about a specific set of clients.

You know that this client’s CFO reads every report and the CMO does not. You know that the last time you pitched a video budget it died in procurement, and why. You know which two people need to be on a call before a decision sticks. You know the phrase the founder hates. None of that is written down anywhere useful. It lives in four people’s heads, a Slack channel nobody searches, and about nine hundred emails. That gets even harder when you are handling multiple clients and need clear brand separation across client accounts.

That is the asset. It is also the thing that walks out the door when someone leaves, and the thing your AI tools cannot see. This is why agencies use Notion AI for knowledge management and client onboarding, so that internal context is at least organized.

So you end up in a strange position. The most valuable information in the business is the one input your expensive new software has no access to. Every session starts from zero, which means every session gets you a competent generic draft that somebody senior then spends forty minutes making specific.

You have bought speed on the easy half and changed nothing about the expensive half.

Why this shows up as flat margin

This is the bit that finally got my attention, because I like looking at margin more than I like looking at software.

Take monthly reporting. Say the writing takes two hours and the rest of it takes six: pulling numbers out of Google Analytics and three other platforms, doing client reporting, chasing the account manager for context on the weird week, formatting it, sense checking it, going back because a figure moved, and getting it approved. Cut the writing in half with a tool and you have saved an hour out of eight. AI marketing tools can analyze large datasets quickly and improve efficiency in data analysis and reporting tasks, but only when they are connected to the reporting workflow.

You have not touched the assembly. You have not touched the chasing. You have not touched the context rebuilding, which is the single biggest hidden cost in an agency and the one nobody puts on a timesheet because it does not feel like work. It feels like talking.

Databox automates client reporting and can save 3 to 5 hours weekly. Claude can help with analyzing data by drafting client reports from data exports and summarizing key trends and insights.

So the month closes and the margin looks the same as last year, and you have eleven subscriptions, and someone on the leadership call asks what the AI spend is doing for us, and the honest answer is that people quite like it.

The other thing I noticed, once I started counting properly, is how much of the tool spend was defensive. We bought some of it because a client mentioned it. We bought some of it because a competitor was posting about it. At least one we bought because I did not want to be the agency that had not tried it. None of those are terrible reasons to run a trial. They are terrible reasons to still be paying twelve months later without having asked what changed.

MIT’s NANDA initiative put a number on this that made me wince. Their GenAI Divide report found that around 95% of enterprise generative AI pilots produced no measurable impact on the P&L. Not that the tools failed. That the pilots did not change anything that showed up in the accounts. The reason they gave was a learning gap: the systems did not retain feedback, did not adapt to context, and did not improve with use. Which is a formal way of describing my eleven line items.

The Harvard and BCG field experiment on knowledge worker productivity found something similar from the other direction. Consultants using AI were meaningfully faster and better on tasks inside its capability, and measurably worse on tasks just outside it, without being able to tell which was which. Speed is not the problem. Knowing where to apply it is.

What compounding would actually look like

I am not going to pretend I had this figured out at the time. I did not. But I could describe the shape of what was missing, and it was not another tool.

It was one place that remembered.

If there were a single system that held every client conversation, every campaign we ran, every decision and the reasoning behind it, and every piece of feedback we ever received, then a drafting tool would not need to be briefed. It would already know. The reporting process would not need six hours of assembly, because the assembly is mostly retrieval and the retrieval is the part a machine is genuinely good at. That does not mean replacing everything with one AI layer; it means connecting software platforms and project management software so context can move between systems.

More to the point, that system would get better every month by doing nothing except being used. Month twelve would be better than month one, not because anyone upgraded it, but because there was more in it. Agencies should look for AI integration that streamlines client acquisition, project delivery, and reporting, because that is the sort of workflow change that compounds. That is what compounding means, and it is the thing eleven separate subscriptions structurally cannot do. They do not talk to each other. They are not supposed to. Each one is a complete product for a single job, which is exactly why none of them can be the memory. Zapier connects over 3,000 apps to automate workflows without coding, which makes it useful for moving information across different platforms. Gumloop lets teams build AI agents without coding, so an AI automation tool can sit between systems without turning the whole stack into an engineering project.

There is a version of this argument that gets made badly, so let me be careful. I am not saying the tools are worthless and the only real answer is a big infrastructure project. Plenty of agencies have got genuine value out of a single tool used well by one person who cared. What I am saying is that value of that kind stops where the person stops. Agencies should prioritize native integrations and focused stacks of AI tools for core jobs to reduce data errors and manual work. It does not accumulate, it does not survive them leaving, and it does not show up as a structural change in what the business costs to run. It is a good habit rather than an asset.

So what for client reporting

I do not think the answer is to cancel everything. The goal is a smaller stack of the best tools for marketing agencies, not simply more tools. We did cancel some of it, and I would recommend the counting exercise to anyone, if only for the fun of finding out what you are paying for.

But I would sit with the harder question first, because it is the one that determines whether any of this ever reaches the P&L. Start with the core jobs: content creation, content writing, keyword research, project management, client calls, and reporting.

If someone asked you today which of your business processes has genuinely changed shape in the last year, could you name one? For content tools and SEO work, Jasper supports content generation at scale for blog posts, blog articles, and long-form work, has more than 350,000 users, and its Brand Voice supports brand voice customization; Surfer SEO gives data-backed guidance for keyword research and helps optimize pages for search engines; ContentShake AI pulls from Semrush data; Writer keeps house style aligned with approved terminology; Grammarly offers a Style Guide feature that enforces brand guidelines; and Hemingway App scores readability using US educational grades. For creative output, Canva helps marketing agencies keep marketing materials and social media posts consistent across multiple clients. Not gone faster. Changed shape. Fewer steps, fewer handoffs, fewer people touching it.

If you can, you are further along than most agencies I talk to. If you cannot, the tools are not the problem and buying a twelfth one will not help.

I did not have a good answer to that question when I asked it of myself. That turned out to be the useful part. Many agencies now put roughly 3-7% of monthly revenue into AI and marketing automation, so each of these marketing tools should earn its place by changing workflow shape, improving customer experience and decision-making, and automating work like writing and data analysis rather than sitting there as a standalone writing tool or AI assistant.