TIER 4 / STANDARDS-FIRST AI

Standards-First AI

Agents trained on your standards, not on nothing.

Most AI agent projects fail because they are trained on nothing. No written voice, no claims rules, no quality bar; just a model, a vendor and a prayer. Gartner predicts over 40 per cent of agentic AI projects will be cancelled by the end of 2027. I do it the other way round: standards first, agents second. Your marketing standards written down and made explicit, then agents built on those documents, then governance and handover so your team runs it without me.

Before you commit to anything: five minutes tells you which phase you actually need.

Take the readiness check

The paragraph you can repeat to your board

We are paying a senior marketing operator to write down our marketing standards, the voice, the quality bar, the claims rules, who we sell to, and then to build AI agents on those documents so the team can draft, review and repurpose marketing to that bar without adding headcount. The documentation is ours and stays useful even if we never switch an agent on. Nothing the agents touch publishes without a named human signing it off, quality is measured against the written standards rather than anyone’s opinion, and the AI Act transparency marking is handled as a workflow step. It costs less than the content hire it replaces, and unlike the hire, the standards do not resign.

The method

Three phases, strictly ordered. Stop after any one and keep everything.

Phase 1 stands alone and is the prerequisite for everything else. Phases 2 and 3 are only sold on top of it. A client can stop after any phase and keeps everything delivered to that point.

3 TO 4 WEEKS / EUR 9,500 FIXED

Phase 1: Standards capture

The documentation most marketing teams keep in one senior person’s head, written down and made explicit. This is the moat. The agents are just the printer.

What ships

Voice and register rules: sentence style, register by channel, banned vocabulary, worked good and bad examples per rule
Claims discipline: which claims need evidence, how a claim carries its receipt, what never ships
Quality bars per content type, each with an annotated example that meets the bar and one that fails it
ICP definitions and messaging hierarchy: who you sell to, in what order, with what argument
Templates and structural rules for each content type
The review checklist: the single document any draft, human or agent, is checked against

What I need from you

4 to 6 hours of founder or senior marketer interview time; your best and worst existing content, with your own verdict on which is which; win/loss notes or sales call access; and a named internal owner identified by the end of the phase. Without the owner, Phase 2 does not get sold.

Accepted when

Every rule has a worked example. The blind test passes: a writer or model given only the pack and a brief produces a draft you recognise as yours. You sign the pack as the bar you will hold your own team to. The pack is in version control with an owner named on the cover.

4 TO 6 WEEKS / EUR 18,000 TO 26,000 FIXED BAND

Phase 2: Agent build on the standards

Agents built on the Phase 1 documents, running on your real backlog during the build, not on demo briefs. Tool-agnostic by design: the standards are plain documents and survive any model or vendor change. The band is scoped on content types, agent count, portal state, brands and languages, and fixed in your Sprint Brief or proposal before work starts.

The agents

A drafting agent: first drafts of your defined content types, inside your templates, to your documented bar. Drafts, never publishes.
A review agent: checks any draft against the review checklist. Flags claims without receipts, voice drift, banned vocabulary, structural misses. Built on different instructions from the drafting agent, deliberately, so they do not share blind spots.
A repurposing agent: cuts approved long-form into the derivative formats your templates define, preserving claims and receipts verbatim.
A CRM hygiene agent: HubSpot property normalisation, lifecycle consistency, duplicate flagging. Propose-and-approve only; it never writes to your CRM unattended.

What ships

The agents, configured and running in your accounts, not mine. An agent contract sheet per agent: what it receives, what it produces, what it is not responsible for, what it can touch, what happens when it fails. An evaluation set of 20 or more real tasks with scored baseline outputs, kept for regression testing. A scoped-access map, signed off by you.

What I need from you

The named internal owner, 2 to 4 hours a week through the build. Sandboxed, least-privilege access: read-only CRM views to start, a drafts-only content space, no production credentials. A real backlog to run the agents on.

Accepted when

Agent drafts score to your documented standards on the evaluation set, in side-by-side review with your owner, verdicts recorded. The review agent catches a set of deliberately seeded violations, a fabricated claim, a voice break, banned vocabulary, before any human flags them. Every agent has a defined failure behaviour. No agent can publish externally or write to production without a human approval step, and we verify that by attempting it.

2 TO 3 WEEKS PLUS A 60-DAY WINDOW / EUR 6,500 FIXED

Phase 3: Handover and governance

The phase that makes this a system you operate, not a dependency on me.

What ships

The operating manual: who runs what, on what cadence, with what escalation path. The human sign-off map: which outputs need approval, by whom, with what turnaround. The quality routine: a monthly sample of agent output scored against the standards by your owner, with a drift log; when output is wrong, the standard gets fixed, not just the output. The AI Act transparency workflow: provenance marking as a template field and a publishing step, with the audit trail it produces as a by-product. Training for your owner, with a recorded walkthrough they keep. A 60-day support window and one standards revision.

What I need from you

Your owner’s calendar honestly committed to the monthly quality routine, and a decision on the publishing sign-off chain, made by the founder, in writing.

Accepted when

Your owner runs one full cycle, brief to draft to review to approved output, with no involvement from me, and the trail is logged. A deliberately bad input, a fabricated statistic in a brief, is caught by the gates before publication. The governance log has real entries, not a template. You can state, in one sentence each, what the system does when an agent fails, when quality drifts, and when the standards need to change.

The prices

Fixed prices, published

Item	Price
Phase 1: Standards pack, standalone, yours outright	EUR 9,500 fixed
Phase 2: Agent build, scoped band fixed before work starts	EUR 18,000 to 26,000
Phase 3: Handover and governance, incl. 60-day support window	EUR 6,500 fixed
Full programme (1 + 2 + 3)	EUR 34,000 to 42,000
Governance retainer, optional, sold only after Phase 3	EUR 750 a month, 3-month minimum

Compare it to the alternative. A mid-level content hire in the Netherlands costs roughly EUR 56,000 to 72,000 a year fully loaded, and that recurs every year. The full programme is bought once, the standards are yours outright, and the standards do not resign. For active retainer clients, governance upkeep sits inside the retainer’s build allowance at no extra fee.

50/25/25 milestones per phase, deposit on signing. Ex VAT. Dutch law. Phases do not credit into each other; each is priced for its own work and survives standalone by design.

Governance

Your last vendor sold you agents. Nobody sold you the controls.

Human sign-off, fixed, not optional

Nothing external-facing publishes without a named human approval. That is contractual, not configurable. CRM changes are propose-and-approve until the hygiene agent has a clean multi-week record, and you decide if that ever changes. The review agent’s “claim without receipt” flag always escalates to a human; it is never auto-resolved. And when an agent is uncertain, it stops and asks. A stalled draft is cheaper than a published error.

Quality measured against the written bar, not anyone’s opinion

The standards exist in writing precisely so quality is checkable rather than a matter of taste. The review agent scores every draft against the checklist. Your owner scores a monthly sample against the same checklist. Every miss traces to either a standard that was unclear, so we fix the document, or an agent that ignored it, so we fix the agent. The fix is logged. The feedback loop is the system.

EU AI Act, Article 50

Transparency obligations apply from 2 August 2026. For a systematised content operation, provenance marking is a workflow step and a field in the template, not a compliance project. AI-generated and AI-assisted content is marked at the point of creation, the publishing step carries the disclosure where your context requires it, and the audit trail, which agent produced what, against which standards version, approved by whom, when, falls out of the governance log as a by-product. One boundary, stated plainly: I provide the operational workflow and the audit trail. I do not provide legal advice, and your counsel owns the interpretation for your context.

What never enters the agents

Customer personal data beyond what a task strictly requires. Credentials and API keys. Anything under NDA. Unreleased financials or fundraising material. HR and personnel matters. Legal disputes. Agents run on least-privilege access: scoped CRM views, not exports; a drafts space, not your production CMS. Vendor terms are checked before build: no training on your data, EU data residency where the stack offers it. And your own accounts hold everything, so revoking my access revokes everything.

Read this first

When this does not work

Read this before you pay me anything.

No standards-worthy volume. If you publish twice a quarter, buy Phase 1 and stop. The agents will not pay back below a real content cadence.
No internal owner. Without a named owner the system decays in months. I do not sell Phase 2 until the owner is identified. That is a stated rule, not a preference.
Content-mill expectations. If the goal is fifty posts a week, this is the wrong service and I will say so. The system raises the bar. It does not abolish it.
Standards that do not exist yet. If you are pre-product-market-fit and still discovering your voice and ICP, come back when there is something true to document.
Outsourcing judgement. The agents enforce the bar. Setting the bar, and changing it, stays human. If you want the thinking automated, you are buying the wrong thing from the wrong person.

The scope fence

Not a chatbot project. No customer-facing support bot, no website chat widget. Different problem, different risk profile.
Not martech rescue. A broken portal or a CRM in chaos is a systems build, quoted separately. The hygiene agent assumes a workable portal; it does not rebuild one.
Not me as your retained writer. After handover, your team plus the system produce the work. If you want my hands on content monthly, that is the fractional CMO retainer, priced as such.
Not AI strategy slideware. No vendor selection decks, no transformation roadmaps. Everything delivered runs, or it is a document you keep and use.

Two ways to start

If you already know AI is on your desk, start with the Scoping Sprint. EUR 1,500 fixed, credited in full against any engagement within 90 days, and the brief will tell you which phase you need, or that you need none of them yet. If you are not sure, the readiness check will tell you in five minutes, free, and one of its possible answers is “do not buy agents yet, including from me”.

Start with a Scoping Sprint

Take the readiness check