Skip to content

AI consultancy

Useful AI, not AI theatre

Workflows · Tooling · Quality standards · Advisory

Every marketing team has been told to use AI. Very few have been told what for. I help teams work out which parts of their process genuinely get better with a model in the loop, which get measurably worse, and how to tell the difference without betting the brand on finding out.

01

Why most AI rollouts in marketing do not stick

The pattern is consistent enough to be predictable: a burst of enthusiasm, a spike in output, a quiet decline in quality, and six months later nobody is using it for anything that matters.

Tools without a job
Licences for everyone and no decision about what they are for. Usage spikes, plateaus, and dies, and the only measurable outcome is the invoice.
Volume as the only win
More content, faster, is the easiest thing to measure and usually the least valuable. Ten times the output at half the quality is not a productivity gain.
No quality floor
Nobody has defined what “good enough to publish” means, so the standard drifts downward one acceptable-seeming piece at a time until the brand sounds like everyone else.
Nobody owns the risk
Particularly in regulated categories, where a fluent, confident, wrong sentence is not an embarrassment but a compliance problem.
02

Who this is for

Teams that want the actual gains and can do without the pilot-project theatre.

  • Teams told to “use AI”

    A mandate from above, no brief, and no idea which parts of the job it should touch.

  • Small teams, wide remit

    Two or three people covering what ought to be six roles. This is where the real leverage is.

  • Regulated businesses

    Where the upside is obvious and the failure mode is a compliance incident, so the guardrails have to come first.

  • Founders doing their own marketing

    You are the strategist, the writer and the analyst. Getting the workflow right buys back the most time.

  • Content teams under volume pressure

    Being asked to publish more without more people, and wanting to do it without the quality collapsing.

  • Teams that already tried

    Adoption stalled, output got blander, and people quietly went back to the old way. That is a workflow problem, not a tooling one.

03

What you get

Each workstream is designed to move a specific metric, not to pad a report.

01

Workflow audit

I map how the work actually gets done now — not the documented process, the real one — and mark where a model in the loop would help, where it would hurt, and where it would just add a review step. Most teams are automating the wrong half.

02

Tooling and setup

Which tools, configured how, with what context. Most of the value is in the setup rather than the subscription: the prompts, the source material, and the connections into the systems you already use.

03

Quality standards

An explicit floor for what gets published, who reviews what, and which categories of work never go out without a human decision. This is what stops the slow drift toward generic.

04

Enablement

Training your team to do this themselves, with worked examples from your own work rather than a generic deck. The goal is that you do not need me afterwards.

04

How it runs

Four phases with real durations. Nothing here is a proprietary framework — it is the order the work has to happen in.

01

Diagnose

Week 1

How the team works now, what it spends time on, and where the genuine constraints are. Often the answer is that AI is not the highest-value fix available, and I will say so.

02

Pilot

Week 2–4

Pick two or three workflows and rebuild them properly, with a quality bar agreed up front and a way to tell whether it worked. Narrow and real beats broad and theoretical.

03

Standardise

Month 2

Turn what worked into something repeatable: documented prompts, review steps, and the standards that keep quality from sliding once the novelty wears off.

04

Hand over

Ongoing

Train the team, document the decisions, and leave. Check in periodically as the tools change, because they will.

05

What this has produced

I have built and run the AI operating system for a regulated financial firm’s marketing function — the skills, the eval suites, the governance layer and the quality gate that decides what is actually good enough to publish.

Task-specific skills covering research, drafting, SEO, design, and compliance review
53Task-specific skills covering research, drafting, SEO, design, and compliance review
Eval cases across the six skills where a bad prompt change would do the most damage
50Eval cases across the six skills where a bad prompt change would do the most damage
Checks a piece passes before it can publish — 28 diagnostic, 8 that block it outright
36Checks a piece passes before it can publish — 28 diagnostic, 8 that block it outright

What makes it work is not the model, it is the scaffolding. Content requests are filed and triaged automatically; the build and the approval stay with a person, because that is where judgement actually earns its keep. A governance layer records the constraints and the decisions behind them. Nothing publishes without passing an anti-AI quality gate I wrote to catch the templated phrasing, negative parallelism and synonym-cycling that makes generated copy recognisable — and in a regulated category that gate doubles as a compliance control. The eval suites cover six skills rather than all fifty-three, which is the point: they sit where a bad prompt change would do real damage, not everywhere for the sake of a round number. I have deliberately not attached a traffic figure to this page — the growth that firm saw came from a replatform and a content programme, and crediting it to the tooling would be the kind of unearned causal claim the rest of this site avoids. Client unnamed; the system is mine.

06

Questions people actually ask

Including the ones about money.

The other four

Most engagements use two or three together. A strategy sets the direction; the rest do the work.