Start with the work that is repetitive, high volume and mechanically checkable, which in almost every B2B company means reporting first, then research and drafting, then lead handling. Start with data hygiene before any of it. The common failure is starting with a tool and looking for a use case afterwards.
Almost every B2B company we speak to has tried AI somewhere. Someone has a subscription, someone drafted a few blog posts with it, and there is usually a proof of concept that impressed people once and then quietly stopped being used. The tools are not the problem. The sequence is.
Those two figures together describe the situation well. Adoption is close to universal. Utilisation is not. Buying the thing is not the hard part, and it never was.
The test for what to automate
One question decides it: does this work compound?
Automating a task that runs once a quarter saves an afternoon a year and costs a week to build and maintain. Automating something that runs every week, consumes senior attention, and follows the same shape every time frees capacity permanently. The frequency matters more than the sophistication.
A second filter sits underneath it: can a mistake be caught cheaply? Reporting errors surface immediately, because numbers that do not reconcile are obvious. A wrong judgement call in a customer email surfaces later, in front of the customer. Automate where errors are cheap and visible first.
The sequence
Fix the inputs before automating anything
Automation applied to disorganised data produces wrong answers faster and with more confidence than a person would. Naming conventions, deduplicated records, working tracking and a written definition of a qualified lead all come first. This step is unglamorous and it is the one that decides whether the rest works.
Automate reporting
Collection, joining and refreshing of channel and CRM data. High frequency, entirely repetitive, and self-checking because numbers that do not reconcile announce themselves. This is the fastest reliable win in most companies and it buys back senior time every single week.
Automate research and assembly
Source gathering, competitor monitoring, query research, brief construction and first drafts. A person edits everything a customer will read. The output of this stage is a strong draft that a human finishes, not a published article.
Automate lead handling
Enrichment, deduplication, routing and follow-up sequencing. This is deliberately fourth, because it only works once qualification rules are written down, and writing them down is a conversation with sales rather than a technical task.
Review and retire on a schedule
Every automation gets a review date. Anything that has stopped earning its place is switched off. Unreviewed automation quietly accumulates into a system nobody understands and everybody is afraid to change.
What not to automate
The line we hold is straightforward: automate the assembly, keep the judgement.
- Positioning and messaging. These come from talking to customers and deciding what to stand for. A model can help you phrase a decision. It cannot make it.
- Anything published without review. Generic output earns neither rankings nor citations, and it is recognisable. The cost of publishing it is not zero; it is reputational.
- Qualification on a high-value lead. Automate the routing, not the decision about whether an unusual but promising enquiry is worth a senior conversation.
- Deciding what to test next. Models are good at generating variants and poor at knowing which hypothesis is worth the traffic.
- Explaining why a number moved. Attribution to a cause requires knowing what else happened in the business that month, which is context no tool has.
Why AI rollouts stall
The pilot has no owner
Someone builds something impressive, demonstrates it, and returns to their actual job. Without a named owner and a place in an existing workflow, a pilot is a demonstration rather than a change.
It was built on bad data
The output is wrong often enough that people stop trusting it, and once trust goes it does not come back without a visible rebuild. This is why step one is step one.
It replaced the wrong step
Automating the enjoyable creative part while leaving the tedious data work manual is common, because the creative part is the more interesting thing to build. It also saves the least time.
Nobody measured whether it helped
If you cannot say how many hours a week an automation returns, you cannot defend it or improve it, and it will be first out when priorities tighten.
A realistic first 90 days
For a company of 20 to 60 people with no dedicated operations resource, a sensible first quarter looks like this: weeks one to three on data hygiene and tracking; weeks four to six building automated reporting; weeks seven to ten on a research and drafting pipeline with a human editing step; weeks eleven to twelve writing qualification rules with sales and automating routing against them.
That sequence produces a compounding result rather than a demonstration, and every stage is independently useful if you stop there.
How do you know whether it is working?
Every automation should have a number attached to it before it is built, and the number is almost always hours returned per week. If you cannot state that figure, you are not measuring the automation, you are admiring it.
Three measures cover most cases. The first is time returned: how many hours a week the person who used to do this work has back, measured honestly rather than estimated generously. The second is error rate: how often the output needs correcting, tracked for the first month so you can see whether it is improving or whether you have quietly accepted a lower standard. The third is adoption: whether the people it was built for are still using it 60 days later, which is the measure that most often reveals a pilot nobody wanted.
That last one deserves particular attention. An automation that works technically and is used by nobody is a failure, and it is a common one, because tools are usually built by the person most enthusiastic about them rather than the person who does the work. Building alongside the person whose workflow changes is the single best predictor of whether it survives the quarter.
Set review dates when you build. A quarterly pass over everything running, asking whether each item still earns its place, prevents the slow accumulation of half-working automations that nobody remembers commissioning and everybody is nervous about switching off.
A note on what this is not
None of this is an AI strategy in the sense the phrase is usually sold. It is operations work that happens to use AI where AI is the cheapest way to do the job. The companies getting real returns are not the ones with the most sophisticated models. They are the ones whose data was clean enough that ordinary automation worked on the first attempt.
Common questions
Should we start with a tool or a process?
A process. Choosing a tool first commits you to its assumptions before you know your requirements, which is how companies end up with software that only fits after the workflow has been bent around it. Decide what should happen, then buy something that does it.
Will AI-written content hurt our search rankings?
Unedited generic content will underperform regardless of how it was produced, because it satisfies nobody completely and earns no citations. Content researched and assembled with AI and then genuinely edited by someone who knows the subject performs like any other well-made content.
How much should a company this size spend on AI tooling?
Less than most expect. The expensive part is the work of cleaning data, writing rules down and integrating systems, not the subscriptions. Companies that reverse that ratio usually end up with capable tools and nothing reliable running through them.
What is the first thing to automate if we only do one?
Reporting. It is repetitive, it runs weekly or monthly, the errors are self-announcing, and it returns senior time immediately. It also forces the data hygiene work that everything else depends on, which makes the next step cheaper.
The sequencing work above is part of a marketing operations audit, which produces a written roadmap rather than a tool recommendation. Where the operation needs building around it, that is marketing operations setup.