Blog

Substack told its writers this year that it will start demoting and removing “AI slop” from its recommendation engine. No warning shot, no grace period. If a post reads like generic filler with no point of view behind it, the algorithm now buries it, even when a human technically wrote it. That’s the moment content marketing changed, and most agencies haven’t clocked it yet.

The gap nobody’s talking about

Marketing teams have gone all-in on AI for content. HubSpot’s 2026 State of Marketing report puts AI use at 91% of teams, up from 63% the year before, which is fast adoption for anything in this industry. Idea generation, first drafts, multi-asset campaigns, product copy: all largely AI-assisted now.

Here’s the part that should worry you more than the adoption number. Only 41% of marketers say they can confidently prove that AI is delivering ROI, down from 49% the year before. Adoption went up. Confidence in the results went down. That’s not a rounding error, it’s a trend line pointing the wrong way, and it’s exactly the pattern you’d expect if a lot of teams adopted the tool without adopting the discipline that makes it pay off.

The UK numbers tell the same “everyone’s doing it, few are doing it well” story. The ONS’s June 2026 survey of AI in UK businesses found 35% of businesses with ten or more staff now use at least one AI technology, up from around 12% in late 2023. Within that, 18% are using large language models specifically to generate text, the single most common use case. But adoption is shallow: the average number of AI tools an adopting business actually uses has crept from 1.4 to just 1.6 over the same period. Widespread use. Thin evidence it’s actually working, and barely any sign firms are going deeper once they start.

Why the ROI number is falling, not rising

Two things are happening at once. First, governance hasn’t kept pace with output speed. Concerns about legal, compliance and brand review grew 3.4 times year over year in the same report, the fastest-growing worry in the whole survey. AI can produce a week’s worth of content in an afternoon. Most review processes were built for a much slower publishing cadence, so a lot of that output goes out unchecked simply because nobody had time to look at all of it.

Second, and more basic: a lot of what’s going out is genuinely bad. Not wrong, just forgettable. It reads like the median of everything ever written, smoothed into a paste and published at volume. Readers can feel it even when they can’t say why, and platforms are starting to detect it algorithmically too. Substack’s policy shift is specific about the target: generic listicles, recycled takes, posts with no discernible point of view. Using AI for research or editing is fine. Publishing its unedited median opinion is what gets punished, and that distinction is the one most agencies still haven’t built into their workflow.

Adoption solved the speed problem. It didn’t solve the taste problem, and taste is the part a platform can now detect.

What “doing it well” actually looks like

The firms in that 41% who can prove ROI aren’t using less AI than everyone else. They’re using it with a visible layer of human judgement sitting on top of the output, applied at two specific points: before a draft gets written, and before it goes live. Neither point requires slowing production down much. Both require someone to actually own the decision, rather than letting a tool’s default output become the published version by default.

Ruben Hassid, whose “How to AI” newsletter has become one of the more widely read practical AI guides for non-technical readers, has a method worth stealing for the first point. Before you generate anything, build a running document (he calls it an anti-ai-writing-style file) that lists every word and phrase you’ve come to recognise as an AI tell: “unlock,” “let’s dive in,” the “it’s not X, it’s Y” sentence pattern, overused dashes, whatever your own ear catches. Feed that file into your prompt every time. It’s a five-minute setup that pays off on every piece you generate afterwards, because it forces the model to write in a voice instead of the average voice.

Try this

Two changes you can make this week, no new tools required:

1. Build the ban list. Open a doc, name it whatever you like, and every time an AI draft uses a word or construction that makes it read as machine-written, add it. Paste the list into your system prompt or custom instructions before every generation. It gets more useful every week you keep it, and it’s the single cheapest fix for the “slop” problem, because most AI tells repeat across dozens of pieces once you start noticing them.

2. Add one gate before publish, not before drafting. Governance concerns are spiking because review hasn’t scaled with output, so don’t try to review everything equally hard. Pick the one question that actually catches bad content: “could a named person on our team defend every claim in this piece if a client asked them to?” If the honest answer is no, it doesn’t go out under your brand, however fast it was to produce. Assign that gate to one person by name, not to “the team”, or it quietly stops happening within a month.

The opinion part

AI content tooling is not the differentiator anymore. At 91% adoption it was never going to stay one for long. The differentiator is whether a firm has built the fifteen minutes of human judgement around each piece that keeps it out of the “slop” bucket Substack is now actively burying. That’s a process problem, not a technology problem, and it shows up first in a team’s skills and habits rather than in its toolkit. If you’re not sure whether your team’s AI habits would survive that kind of audit, that’s what an AI skills and adoption review is for: not to add more tools, but to find out where the judgement layer is thin before a platform, or a client, finds it for you.

Everyone bolted AI onto their cold email this year. Reply rates, on average, did not move.

Instantly’s 2026 Cold Email Benchmark Report, built from billions of interactions across thousands of active workspaces over 2025, puts the overall average reply rate at 3.43%. That number is barely different from where cold email sat five years ago, before every outbound tool shipped an AI writer. Meanwhile the British Chambers of Commerce’s March 2026 research shows 54% of UK firms now actively using AI, up from 35% in 2025 and 25% in 2024. Adoption tripled in two years. The metric that actually pays the bills stayed flat.

That gap is the real story, and it’s a useful one if you run outreach for an agency or a small service firm. AI didn’t fail at lead generation. Most people used it to do the same thing faster, not to do a different thing.

Where the reply rate actually moved

The BCC data has a second number worth sitting with: of the 54% of firms using AI, only 11% use it extensively to automate or streamline operations. The rest are dabbling: a chatbot here, an AI-drafted email there, without changing the underlying process. Same list, same generic template, same “Hi {firstName}, hope you’re well” opener, just typed by a model instead of a person.

Instantly’s own benchmark shows what the other tier looks like. Their top 10% of senders hit 10.7%+ reply rates, three times the average. What separates them isn’t a better AI model. It’s what the AI is pointed at. The consistent thread across the top performers, and across separate research from Sendr, Unify GTM and Martal’s 2026 cold email data, is signal-based targeting: emailing people because something specific just happened (a hire, a funding round, a tool switch, a website visit) rather than because they match a firmographic filter.

The mechanism is simple. A generic list gets a generic message, and the prospect can tell in one line. A signal-triggered message references something true and current about their business, which is the entire definition of relevance. AI is very good at drafting that message once you hand it the signal. It is not good at inventing the signal for you, and most tools people bought this year only automate the drafting.

Adding AI to a bad list gets you a faster bad list.

The maths behind fewer, sharper emails

Volume was the old lever: send more, book more. Signal-based sending inverts that. A team sending 200 emails triggered by a real signal at a 20% reply rate books roughly the same number of conversations as a team sending 1,000 generic emails at 3%, using a fifth of the volume, a fraction of the domain-warmup risk, and far less time spent by whoever’s writing follow-ups. Fewer, better-targeted emails compound: deliverability holds up longer because you’re not burning domains on blast volume, and every reply is a genuinely interested prospect rather than someone confused about why they got the email.

This is also why “AI for lead gen” as a category gets a mixed reputation. Bought as a volume multiplier, it multiplies whatever was already mediocre about the list and the message. Bought as a targeting and drafting layer on top of real signals, it does what people hoped AI would do for outreach in the first place.

Try this

Two changes you can make this week without buying anything new:

1. Benchmark your own reply rate against 3.43% before you touch AI copy. Pull your last 90 days of outbound. If you’re below 3.43%, the problem is almost never the writing; it’s the list. No amount of AI polish fixes a list built on industry and headcount filters alone. Fix the list first.

2. Pick three trigger events and test a small signal-based batch against your usual send. New hire in a relevant role, a funding announcement, a tool or vendor change you can detect (job posts mentioning a competitor’s product are an easy one). Send 50–100 emails against real signals within 48 hours of the trigger (timeliness matters roughly as much as the signal itself, since relevance decays fast) and run it against your normal batch as a side-by-side. Most firms that do this find the signal batch outperforms 3–5x, which matches what the wider research above shows. You don’t need a fancy intent-data platform to start: a saved LinkedIn search and a Google Alert cover the first three signals for free.

Once that’s proven at small scale, the actual bottleneck becomes catching the signals reliably and getting the drafted email into a rep’s outbox within the 48-hour window — which is a process problem, not a copywriting one. That’s the part worth automating properly rather than doing by hand every morning: it’s exactly the kind of repetitive, time-sensitive workflow we build for clients at Ishigai’s process automation service — watching for the signal, drafting against it, and queuing it for a human to approve, so nothing sits stale past the window that actually matters.

The opinion part

The 3.43% average isn’t a ceiling on what AI can do for outreach. It’s a snapshot of what happens when a fast tool gets pointed at a slow, generic process and everyone assumes the tool did the work. It didn’t. The list, the signal and the timing did the work; the AI just typed it up quicker. Firms chasing the 10.7% tier aren’t using smarter AI. They’re using AI on a sharper input. Fix what goes in before you worry about what writes it.

Nine in ten UK businesses raised their AI budgets this year. Fewer than a third can point to a return on it. That gap between rising spend and flat results is the real AI story for UK agencies right now, and it isn’t a technology problem.

According to Spicy Advisory’s 2026 UK SMB AI adoption guide, 85-91% of UK organisations increased AI spend this year, yet only 31% report positive ROI. Seventy-seven percent saw no measurable revenue change at all. Where AI does pay off, it shows up as time saved rather than top-line growth, which would be fine, except most firms aren’t measuring time saved either, so the return is invisible even when it’s real.

ONS figures from June 2026 point at why. The average AI-adopting UK business uses just 1.6 AI tools, and only 10% report extensive use of any of them. Adoption has tripled since 2023, but depth hasn’t moved much at all. Only 15% of businesses in the 0-9 employee bracket — where most agencies sit — use AI at all, against 68% of large firms. UK firms aren’t short of AI. They’re short of depth.

There’s a skills story sitting underneath both numbers. Over 60% of firms name the skills gap as their primary barrier, and only 11% have trained more than half their workforce in AI skills. Buying a licence isn’t the same as building the muscle to use it well, and most firms are stopping at the purchase.

More tools, same problem

Tool fragmentation is Spicy Advisory’s second-named barrier to AI ROI, right behind the skills gap. It’s a pattern I run into constantly with clients: a Copilot licence bought for the whole team, a ChatGPT subscription three people actually open, a Claude account someone on the leadership team set up after a conference, and none of them wired into an actual workflow. Everyone in the building is “using AI.” Nobody’s using it for anything specific enough to measure.

Success in 2026 isn’t about using every AI tool. It’s about choosing the right tool for the right task and going deep on it, as the Stop Chasing Every New Tool newsletter put it this year.

That’s the uncomfortable part for a lot of firms: the fix isn’t procurement, it’s subtraction. Before evaluating a fourth or fifth tool, work out what the first three are actually doing for you, if anything.

Try this: a 20-minute tool audit

You don’t need a consultant for the first pass. Block 20 minutes and do this:

1. List every AI tool with an active licence. Finance can usually pull this off the card statement faster than IT can list it from memory.
2. Write down the one workflow each tool touches — not “content” or “admin”, but the actual task: “drafting first-pass client proposal decks”, “summarising onboarding calls”.
3. Cut or pause anything without a named workflow attached. If nobody in the room can say what a tool is for, it isn’t producing ROI, measured or not, and it’s a licence you can reclaim this afternoon.
4. Pick the tool touching your highest-volume repetitive task and go deep on it. Build the prompt library, the template, the checklist, whatever makes the fiftieth use faster than the first. That’s where the actual time saving lives, and it compounds; a tool used once a week never gets good enough to notice.

Most firms find step 3 uncomfortable, because “we’re paying for it, so we must be using it” feels true even when the usage logs say otherwise. Check the logs. Most seat-based AI subscriptions show login frequency somewhere in the admin panel, and it’s usually a harder number than anyone expects.

Then close the loop: whichever tool you keep going deep on, write down what you’re measuring before you start: hours saved on that one workflow, turnaround time, or output volume. Not “productivity”, a specific number tied to the specific task from step 2. That’s the difference between joining the 31% who can show ROI and staying in the 69% who are pretty sure it’s helping but can’t prove it.

The standard stack, plus one

Spicy Advisory’s guidance for UK SMBs is blunt and, in my experience, correct: standardise on Microsoft 365 + Copilot or Google Workspace + Gemini for the baseline, add one external assistant — Claude or ChatGPT — for the work your core suite doesn’t cover well, and stop there until you can show the first two are earning their keep.

That’s a smaller stack than most firms are currently running, and that’s the point. Depth over breadth isn’t a slogan here: it’s the difference between the 31% seeing ROI and the 69% who aren’t. The 31% aren’t using more tools. They’re using fewer tools for longer, on narrower, higher-volume tasks, until the tool actually knows the job.

If the audit above turns up five half-used subscriptions and nobody in the business owns the decision of what to cut or what to go deep on next, that’s exactly the kind of prioritisation call worth getting an outside, structured view on. That’s the whole point of an AI opportunity audit: not another tool recommendation, but a ranked list of where AI actually pays off in your specific workflows, and where it’s just spend.

The opinion part

AI budgets will keep rising into 2027 whether or not ROI follows: the EY and BCC data both point that way, and nobody wants to be the firm that “isn’t doing AI.” But the 69% gap between spend and return isn’t going to close by adding tool number four. It closes by someone in the business deciding which one tool gets used until it’s actually good, and which three get switched off. That decision is boring, it takes 20 minutes, and almost nobody is making it.