How we use AI
Principles are cheap. Practice is checkable.
Most organisations now have something to say about responsible AI. Far fewer can show you how responsible AI works in practice. This page is what we do, and what we will not do.
Our position
Humans use AI. Humans keep control. Humans are freed.
Three sentences. The rest of this page is the evidence for each.
Humans use AI
Marketing people work with AI on real work from the first week. Your team supervises the playbooks; the playbooks do not replace the team.
Humans keep control
Two named people on every deployment. A person approves first. Limits that are built in, not switched off. See what we do and what we refuse.
Humans are freed
The playbooks take routine work so people can spend their time on judgement, taste and influence. See what happens to the people.
01 · What we do
What we do on every deployment.
This is not new for us. It has been designed in from the beginning and is written into the runbook our operators work from.
Two named humans on every deployment
One accountable for the machinery, one for the outcomes. Both named in the engagement record, not implied by a job title. If nobody on your side can own the outcome, we say so rather than work around it.
Every playbook begins by proposing, never by acting
It puts forward what it would do and a person decides. We measure how often it was right across real work before it is trusted to act on its own, and that threshold is agreed with you rather than set by us.
Gates are reviewed, and amended only on a body of evidence
An approval point is never loosened because a playbook has had a good week. It is reviewed against an accumulated record of real decisions, amended only when that record supports it, and every amendment is a formal change carrying your signature.
Nothing that affects you is a black box
Every decision a playbook takes about your leads, your campaigns or your spend is explainable: the weights are documented, the reasoning is recorded, and figures trace back to the query that produced them. Uncertain data is labelled uncertain rather than smoothed. Our own method, the frameworks and rubrics behind an assessment, remains ours. What we do with your data never is.
The limits that matter are built in, not switched off
A permission that is turned off can be turned on again by whoever holds the password. So our most important limits are not permissions at all. No playbook can answer a question about price, contract or a complaint: it escalates to a person, and there is no setting that changes that. No playbook scores, ranks or reports on your staff. Neither was built, so neither can be switched on. And in a standard deployment no playbook composes prose to one of your customers: a reply to an individual is drafted and waits for a person to send it. The rest, including spending caps, how many records a playbook may touch at once, and the switch that returns everything to manual, are set in the software and tested before go-live, not left as a line on a settings page.
Only what a step needs leaves your systems
Each client has an isolated space, hosted in the UK or the EU, and the decision log lives there. When you leave we switch it off, hand you the full log of everything that was done, and delete what remains. The register of sub-processors comes with the data processing agreement, which is signed before any access is granted.
We are model-agnostic on purpose
Which AI provider sits underneath is a configuration choice, and we keep it that way. A lock-in to one vendor is a conflict of interest, so we do not take one.
Every playbook is assessed for ongoing value
Monthly, against the baseline measured before it was switched on: what it did, what it cost, and what it was worth. Reported to you and to whoever owns the number it moves. A playbook that stops earning its place is one we will tell you to switch off.
02 · What we refuse
A position is only worth reading for the things it rules out.
No playbook improvises to your customer
Three of our fifteen can send a message: event follow-ups, campaign variants, and meeting times on an inbound enquiry. Every word they send was written or approved by a named person before it could go out: a template, a variant, a time from your own calendar. The machinery selects and sends, and it never composes. A reply written for one individual is drafted and waits for a person to send it. And price, contract and complaint are never answered by a machine at all: that escalates, permanently, by design.
We do not recommend AI because we sell it
We supply all four answers: hire it, borrow it, automate it, or stop doing it. Because we can deliver any of them, the evidence decides rather than our revenue model. Anyone who only sells one answer will never tell you to stop.
We take no commission from the platforms we assess
We do not join platform partner programmes that pay us a share of what you spend with a vendor. A trailing commission on a recommendation is an interest you would never see on an invoice, and we assess the martech portfolio those vendors sit in. The one fee we earn from a recommendation is retained search, when the evidence says hire: that is our own service, disclosed before the work starts, and the role design is delivered whatever the evidence concludes.
We do not claim results we cannot evidence
We measure exposure and risk, not a forecast. Where a number is indicative, it says so. We would rather give you a smaller claim that survives a board’s questions than a larger one that does not.
We do not build systems that watch your people
Our playbooks monitor campaigns, pipelines and accounts. They do not score employees, rank individuals or report on anybody’s working day.
03 · What happens to the people
Our answer starts with the capability framework, not with the playbooks.
We mapped what a marketing function has to be good at now, 36 capabilities across seven domains, and we did not conclude that machines should do more of it. The capabilities which decide whether marketing matters to a business are human, and they are becoming scarcer rather than less valuable.
The capabilities that make marketing matter are human ones
14 of the 36 are frontier capabilities, the ones the shift to AI created or changed. Research judgement in a synthetic age. Creative direction and taste. Applied AI fluency. Agent supervision. Capacity design. These are what decide whether marketing is treated as a cost line or as the function that grows the company, and not one of them can be automated.
The playbooks exist to make room for that judgement
Routing, deduplication, pre-flight checks, chasing, assembling Monday’s numbers. Work people do because nobody built anything better, not because it needs a person. Removing it matters less than what the freed time is spent on.
Automation is one of four answers, and often not the right one
Hire it, borrow it, automate it, or stop doing it. A Stocktake recommends building permanent capability as readily as it recommends a playbook, because the evidence decides. Where the gap is judgement, taste or influence, the answer is a person, and we say so.
We have a commercial interest in saying this
We place people as well as deploying playbooks. If the answer were always “automate”, our own business would be smaller. Treat that as a reason to believe us rather than as a disclaimer.
A deployment is a training ground, not a handover
Your team supervises the playbooks on real work from the first week, approving each decision while accuracy is measured. The capability you are left holding is the ability to run them without us.
Check us
Hold us to it.
Take the seven questions into your next conversation about AI, with us or with anyone else. Our answers are above.
Statistics from Responsible AI in Practice: 2025 global insights from the AI Company Data Initiative, Thomson Reuters Foundation and UNESCO, 2026, covering 2,972 companies across 11 sectors.