The agency processes worth automating first, ranked

Most agencies automate in the wrong order. They start with the thing that annoys them most, which is usually not the thing with the best return, and then wonder why six weeks of work produced a nicer version of the same week.

This is our ranking. It is based on what we did to our own shop and what we have since done for other agencies, and the ordering logic is the interesting part rather than the list. Two criteria decide position: how much of the process is retrieval and assembly rather than judgement, and how quickly you find out when it goes wrong.

That second one matters more than people expect. A process where errors surface immediately is safe to automate early, because you will catch the problems while you still have the appetite to fix them. A process where errors surface three months later, in front of a client, belongs at the bottom of the list regardless of how much time it would save.

One more thing before the list. Nothing here is ranked by how much time it saves, and that is deliberate. Time saved is the number everyone asks for and the least useful one for sequencing, because a process that saves forty hours a year but destroys team confidence when it fails is worse than one that saves ten and never does. Sequence for trust first. The hours arrive anyway.

1. Monthly client reporting

The flagship, and the one every agency should do first.

It is mostly retrieval: pull numbers from four or five platforms, compare to last month, note what shipped, assemble, format. The judgement is one section, usually a quarter of the document, and it stays with a person. Errors surface fast because a human reads the draft before it goes anywhere.

The saving is large because the work is monthly and repeated across every client, so a two-hour improvement multiplies by your client count twelve times a year. It is also the process most likely to be resented by the people doing it, which matters more than the arithmetic when you are trying to get a team on board with a change.

Start here. If it does not work here it will not work anywhere, and you will have found that out cheaply.

2. Client onboarding

The biggest single delta we measured, and the least controversial to change.

Onboarding is document collection, account access, kickoff scheduling, and the same eleven questions asked in four different formats. Almost none of it is judgement. The reason it takes weeks is not difficulty, it is handoffs: each step waits for a person who is busy with something else.

Automating it collapses the waiting, which is where the time actually goes. It also fixes a first impression problem. Clients form their view of an agency in the first fortnight, and most agencies are at their worst in the first fortnight.

There is a knock-on nobody mentions. When onboarding stops being painful, your sales team stops unconsciously slowing down deals to avoid it. Fix the misery and the pipeline moves faster, which is a revenue effect from an admin change and not one that shows up in any efficiency calculation.

A practical note on doing it. Do not automate your worst client’s report first, and do not automate your best client’s either. Pick a mid-sized account with a straightforward setup and an account lead who is willing to be critical. The first one is a test, and you want it run by someone who will tell you it is not good enough.

3. Meeting notes into actual next actions

Every agency owner has lost a commitment in a transcript. You said you would look into something, it was said at minute forty-one, nobody wrote it down, and you found out six weeks later when the client asked.

Transcription is solved. The useful part is the step after: extracting who committed to what, by when, and putting it where work actually gets tracked. That step is mechanical, low risk, and immediately visible when it is wrong, because someone gets a task that makes no sense and says so.

High value, low danger, and it builds trust in the wider project faster than anything else on this list because everybody feels it in the first week.

4. Deliverable QA before it reaches the client

This is the quality argument rather than the speed argument, and it is the one I would make hardest to a sceptical owner.

Every deliverable has a checkable layer: broken links, the wrong client name left in from the template, a figure in the summary that contradicts the figure in the table, a missing alt attribute, a claim with no source. None of that requires taste. All of it is embarrassing.

Automating the checkable layer does not replace review. It means the human reviewer spends their attention on whether the work is any good rather than on whether the client’s name is spelled correctly, which is a better use of a senior hour and a better outcome for the client.

Automation as a floor under the work, not a replacement for it.

It also gives you something you have never had, which is data on what goes wrong. Run automated checks for six months and you will have a list of the errors your agency actually makes, ranked by frequency. Most agencies have strong opinions about their quality problems and no evidence. This gives you the evidence, and it is usually not what anyone predicted.

5. Content briefs

Templates already exist for this, so the question is what a brain adds that a template does not.

Three things. It knows what this client has already published, so the brief does not duplicate. It knows what they rejected last time and why, so you stop proposing the same angle. It knows the house voice from the last two hundred pieces rather than from a paragraph in a style guide nobody reads.

The output is a brief a writer can start from rather than a form a strategist has to fill in. The judgement, which is deciding what is worth writing about at all, stays where it was.

One caution. A brief drafted from history is a brief anchored to history, and there is a version of this that quietly makes a client’s content more and more like their existing content until nothing new ever gets proposed. Someone should be reading the briefs monthly and asking whether anything on the list would have been on the list a year ago.

6. Rank-drop and anomaly monitoring

The one clients notice, and the reason it is not higher is the false positive problem.

Continuous monitoring means you find out about a ranking drop, a broken page or a budget overspend within about a day instead of at the next scheduled check. That changes the client relationship more than any other item here, because problems start arriving from you rather than from their boss.

The risk is alert fatigue. Tune it too sensitively and you will train your own team to ignore it inside a fortnight. Set thresholds high, review them quarterly, and put a person between the alert and the client. We got this wrong first time and it cost us some credibility with a client who received two non-events in a row.

7. Proposals and scoping

Here a brain has an advantage that is genuinely unfair, and it is nothing to do with writing.

It remembers what you quoted, to whom, when, whether you won it, and whether the delivery was profitable. Most agencies price the next proposal from a mixture of memory and optimism, because the data to do otherwise exists but is not reachable. When it is reachable, scoping stops being a guess.

The drafting help is fine. The pricing history is the actual value, and it is the thing most agencies have never once looked at systematically.

One thing to insist on: the proposal draft must show its sources. If it says a similar scope took ninety hours, you want to know which three projects that came from, because two of them will turn out to be nothing like the job in front of you. A number without its provenance is worse than no number, because people believe it.

There is a second-order benefit here that took us a while to see. Once scoping is drawn from real history, the arguments inside the agency change. “This will take three weeks” stops being a negotiation between an optimist and a pessimist and starts being a claim with evidence behind it. We spend noticeably less time arguing about estimates than we used to, and the estimates are better, which is a rare combination.

8. Invoice chasing and receivables

Unglamorous, immediately quantifiable, and it converts owners rather than practitioners.

Overdue invoices do not get chased because chasing is awkward and always less urgent than client work. A system that drafts the reminder, times it against how this specific client has historically paid, and escalates on a schedule removes the awkwardness by removing the decision.

Keep a human on the send for anyone you actually care about. There is a version of this that damages a good relationship over four days of lateness, and it is not worth the cash flow.

Worth being honest about the size of this one. Receivables automation is not going to change your business. It is on the list because it takes about two days to build, it pays for itself in one cycle, and it is a good early win to show a sceptical finance person. Not everything has to be transformational.

9. Client email triage

The highest anxiety item on the list, which is why it sits here rather than in the top five.

The guardrails come first, before any description of the benefit. It never sends. It only ever drafts. It never handles a complaint, a scope dispute, or anything with a legal edge, and it flags those to a person immediately rather than attempting them.

Inside those limits it is useful: sorting what needs a person today from what needs a person this week, and drafting the routine replies that make up most of an inbox. But if you are going to be nervous about one thing on this list, be nervous about this one, and start it last among the things you do start.

10. Link prospecting and outreach

Last, and I want to explain why rather than just ranking it.

This is the process most likely to produce confident rubbish. Prospecting at volume means evaluating whether a site is real, relevant and worth approaching, and that judgement is exactly the kind machines are currently bad at while sounding certain. We have watched an agent produce a long list of perfectly formatted prospects, a large share of which were nonsense, and the formatting made the nonsense harder to spot.

Outreach at scale also has a reputational floor. Bad automated outreach is not a neutral failure. It burns the relationship with a publisher permanently and it burns it for everyone at your agency, not just the person who sent it. There is a compliance edge too. The FTC’s guidance on substantiating what you claim about AI applies to what you tell publishers and clients about how your outreach is produced.

There is a version that works, and it is narrow: automate the research and the qualification, keep a human writing the actual approach. Automate the outreach itself and you will get volume, a worse response rate, and a slowly deteriorating sender reputation.

One general point that applies to everything below number five: the further down this list you go, the more the value depends on the guardrails rather than the capability. Nobody buys email triage because drafting is hard. They buy it because an inbox is exhausting, and they stop buying it the moment something goes out that should not have. Design those processes around what they are forbidden to do, and the useful parts take care of themselves.

What the ranking implies about sequencing

Read down the list and the logic is visible. Retrieval-heavy processes with fast error feedback come first. Judgement-heavy processes with slow, expensive error feedback come last, and two of them arguably should not be fully automated at all. If you want a more formal way of sorting this, a published framework for AI risk by use case does the same job with better vocabulary.

The mistake most agencies make is starting with the process that causes the most pain, which is almost always a judgement-heavy one, because judgement-heavy work is what makes people tired. Starting there means your first attempt is your hardest attempt, and if it fails, which it will, the whole project is finished politically before it has proved anything.

Start where the work is mechanical and the mistakes are cheap. Earn the trust. Then move up.

There is a political version of this argument too, and it is the one that actually decides whether a rewire survives. Every automation project inside an agency has an internal sceptic, usually someone senior and usually right about something. If your first three deliveries are boring, reliable and obviously useful, that person becomes an ally who tells you where the edges are. If your first delivery is ambitious and wrong, they become an opponent, and they will be quoting that failure back at you a year later. Sequencing is a people problem wearing a technical costume.

We did this in roughly this order at Shortlist, and got the order wrong twice on the way. The write-up of what we did to our own agency is the honest account, including the automation we deleted at week nine.