Nick QuirkEvery agency dashboard now has an "insights" panel. Most of them are wrong in the same way: they tell you traffic dropped 14% week over week without knowing that the client pushed a site migration on Tuesday, that the drop is concentrated in a branded query set, or that the CMO already knows and doesn't care because the migration was her idea.
The question isn't automated insights vs human strategy as a competition. It's a division of labour. Automation is very good at noticing things. It is very bad at deciding whether noticing matters. Get that split right and a three-person account team can run twelve retainers without drowning. Get it wrong and you either burn 40 hours a month on data entry or send clients confident nonsense.
What automated insights are actually good at
Strip away the marketing language and automated insight generation does four things reliably:
- Detection at scale. Scanning 400 tracked keywords across 9 clients daily and flagging the 6 that moved more than two standard deviations. No human does this consistently by week five.
- Aggregation. Pulling GSC, GA4, rank tracking, time logs and task completion into one place so nobody spends Thursday afternoon in spreadsheet copy-paste.
- First drafts. Turning a month of project data into a 400-word narrative summary that's 70% right and takes 10 minutes to fix, versus 50 minutes to write from scratch.
- Consistency. Applying the same audit checklist to client 1 and client 30 at 4pm on a Friday.
That's genuinely valuable. On a typical £5k/month retainer, reporting and status admin eats 6-9 hours. Automating the data and draft layer reliably pulls that to 2-3 hours. Across ten clients that's most of a full-time salary recovered — and it's recovered from the least strategic work you do.
What automation still can't do
The gap is narrower than it was two years ago but it's stable, and it's always in the same place: context that never got written down.
- The client's dev team has a two-week release freeze in December, so your Q4 technical recommendations are theoretical.
- The marketing director is being performance-reviewed on demo requests, not sessions, so a traffic-focused report reads as tone-deaf.
- You lost a similar client last year by pushing a content refresh too hard; you're deliberately going slower this time.
- The 14% drop is real, but the correct response is "wait ten days" because you've seen this exact volatility pattern after three previous core updates.
None of that lives in your analytics. All of it determines the right decision. That is human strategy, and it isn't going anywhere.
The three-layer model: data, interpretation, decision
The cleanest way to think about the balance is as three layers, and to be ruthless about which layer each task belongs to.
Layer 1 — Data (automate 100%)
Collection, normalisation, anomaly flagging, threshold alerts, crawl comparisons, budget burn against retainer hours. There is no strategic value in a human assembling this. If someone on your team is manually exporting rank data into Sheets, that's a process bug, not a job.
Layer 2 — Interpretation (automate the draft, human-edit)
"Organic traffic to the /guides/ folder fell 22% since 14 March, concentrated in 8 URLs, all of which lost featured snippets." An automated system can write that sentence accurately. It can even hypothesise a cause. What it can't do is rank that finding against the other five findings by commercial importance to this specific client. So: let the machine draft, let a human reorder and cut. Our rule of thumb is that a strategist should delete 30-40% of an AI-drafted summary. If they're deleting nothing, they're not reading it. There's more on structuring that workflow in our guide to AI client summaries from project data.
Layer 3 — Decision (human, always)
What goes into next month's sprint. Whether to spend 12 hours on internal linking or 12 hours on a landing page rebuild. Whether to tell the client their site architecture is the problem when you know they paid another agency £30k for it eighteen months ago. This layer is where your fee actually comes from.
Three ways agencies get the balance wrong
1. Shipping raw automated output to clients
The tell is language like "consider optimising your meta descriptions to improve CTR." A client paying a retainer can generate that themselves in nine seconds. When automated recommendations reach the client unfiltered, perceived value collapses — and it collapses quietly, showing up four months later as a non-renewal, not as a complaint.
Fix: nothing generated goes out without a named human owner and at least one sentence that could only have been written by someone who has been in a call with this client.
2. Refusing to automate commodity work
The opposite failure. Senior people hand-building the same status update every Monday because "clients want a personal touch." Clients want accurate, on-time, and insightful. They do not care whether the chart was assembled by hand. If a task has no judgment in it, protecting it as human work is just expensive nostalgia.
3. Automating insight without automating context
This is the subtle one. Most tools generate insights from analytics data alone, with no idea what your team actually did. An insight that says "rankings improved for 40 terms" is thin. An insight that says "rankings improved for 40 terms in the cluster we published against in February, 6 weeks after go-live" is a strategy input, because it links output to effort.
That link only exists if your project data and your search data live in the same system. It's the core reason we built search data connected to project milestones into PeakKR rather than treating reporting as a separate bolt-on — and it's a genuine difference between a project tool and a marketing-agency project tool. If you're evaluating options, the tool comparison hub breaks down where each platform sits on this.
Two worked examples
Monthly reporting on a 25-hour retainer
Automated (about 90 minutes of machine time, 0 human): data pull across GA4, GSC, rank tracker and time logs; deliverables completed vs planned; hours used vs contracted; anomaly flags; a 500-word draft narrative.
Human (35-45 minutes): cut the draft to the three things that matter; add the causal story the data can't see ("the drop is the migration, we expected 3-4 weeks of noise, we're at week 2"); write the next-month plan; decide what to say about the thing that went badly.
Total: under an hour of senior time for a report that used to take three. The strategy content went up, because the strategist spent all of their time on strategy.
Technical audit triage
A crawl returns 2,340 issues. Automation clusters them: 1,800 are low-value pagination noise, 400 are legacy redirect chains, 140 are duplicate title tags. That clustering is pure machine work and it's excellent.
Then a human asks the question no crawler asks: which of these will the client's two-person dev team actually ship this quarter? The answer might be "the 140 title tags, because we can do them ourselves in the CMS, and one redirect fix, because it's a single .htaccess change." That's a strategy decision made in 90 seconds by someone who knows the account — and it turns
