Operations
Onboarding agents at sub-€8M agencies: a scoring method
Three onboarding modes for a small Dutch agency, scored on speed, DGA-modelovereenkomst defensibility, and who owns the scope-creep when an agent commits on a holiday.
It is 23:47 on December 27th. You are a Dutch agency owner, two FTE plus yourself, sub-€8M revenue. A client emailed at 23:11 to say their year-end budget came in and they need a landing page live by January 2nd. Your intake agent replied at 23:14: scope drafted, Tuesday start confirmed, link to a shared brief. You did not approve the reply. Your account manager is at her in-laws in Limburg. The agent is technically correct, and you are technically liable.
This is the situation we score for every agency we onboard, and most of them are picking the wrong mode for the wrong reason. The choice between pure agent intake, two-person account-manager rotation, and hybrid agent-first with a weekly handover is not a values question. It is a measurable trade-off across three axes that bite a Dutch agency hard.
The three modes, narrowly defined
Mode A, pure agent: a Claude-driven intake bot receives every client email, generates a scope draft, books a slot, and writes back inside fifteen minutes, including on holidays. A human reviews at the next standup.
Mode B, two-person rotation: two named account managers split the week. Every client message routes to the on-duty human. Reply SLA is one business day. No automation between you and the client.
Mode C, hybrid with Friday handover: the agent answers immediately with a holding reply, drafts the scope, and parks every commit until a 30-minute Friday call where one of the two humans signs the brief, approves the start date, and emails the client. Nothing leaves the agent's outbox without a human pressing send.
The three modes look like a continuum. They are not. They commit you to three different operating models, three different contract structures, and three different liability postures.
The rubric we score against
We give each mode a 1–5 score on three dimensions, weighted by how often the dimension shows up in a 36-project year for a small Dutch agency. The weights are not symmetric.
Dimension Weight
-----------------------------------------------
Time-to-first-deliverable per client 0.30
DGA-modelovereenkomst defensibility 0.45
Scope-creep ownership in edge cases 0.25
Defensibility carries the highest weight because it is binary at audit time and silent the rest of the year. You do not feel it until the Belastingdienst sends a letter, and then it is the only number that matters.
Time-to-first-deliverable across 36 projects
We measured this on our own monthly-retainer book and on three client agencies we rebuilt onboarding for during 2024 and 2025. The median time-to-first-deliverable (TTFD), defined as "client receives a real artefact, not a confirmation email", looks like this across 36 monthly projects:
Mode A (pure agent) median TTFD: 1.2 days, p90: 2.8 days
Mode B (two-person rotation) median TTFD: 4.1 days, p90: 7.5 days
Mode C (hybrid Friday) median TTFD: 2.6 days, p90: 4.0 days
Mode A wins on speed and stays close to its median because the agent has no calendar. The p90 spike comes from cases where the human reviewer overrules the agent's draft after the fact and the client has to be walked back. Mode B is slower and more variable: illness, school holidays, the one productive week in August. Mode C trades half a day of speed for a predictable ceiling. Nothing ships before Friday, so nothing slips past Friday either.
If you only score TTFD, you pick the agent. Don't.
Defensibility under the DGA-modelovereenkomst
This is where most agencies get it wrong because the failure mode is invisible. Under the renewed enforcement of the Wet DBA, active again from 2025 after years of moratorium, the Belastingdienst evaluates whether a freelance relationship is genuine or disguised employment. The model agreement you signed with a DGA-freelancer, or that the freelancer signed with your client via you, presumes that scope, instruction, and acceptance are decided between two contracting parties.
An intake agent that autonomously commits scope on behalf of the freelancer breaks that presumption. The agent is an instrument of your agency. If your agency is the one defining the deliverable, the date, and the acceptance criteria without the freelancer in the loop, you have inserted yourself between the freelancer and the client in a way that reads as employment, not intermediation. Rijksoverheid's own guidance on the aanpak schijnzelfstandigheid leans on exactly this question of who calls the shot.
If a freelancer DGA is invoiced via your agency on a modelovereenkomst, the intake agent must not commit scope on their behalf. The agent can draft. A named human in the loop, ideally the freelancer, has to sign before the client is notified.
Defensibility scores, then:
Mode A (pure agent): 2/5 — agent commits autonomously
Mode B (two-person rotation): 5/5 — named human signs every brief
Mode C (hybrid Friday): 5/5 — Friday handover is the commit gate
Mode A is not unusable, but only if every contributor in your book is on a payroll contract, or every project is fixed-price intake-to-delivery with you as principal. For most sub-€8M Dutch agencies, that is not the shape.
Holiday commits and the scope-creep ledger
Back to the opening scene. The agent has, in good faith, agreed to a January 2nd delivery on December 27th. The client now has an expectation. The agency now has, depending on the mode, a problem.
Mode A: ownership lands on the agency principal, you. The agent acted within its system prompt. The system prompt is your artefact. The Belastingdienst, in an audit, will not distinguish between "the agent agreed" and "you agreed". A judge in a civil dispute over the missed deadline will not either. We watched one client absorb €11.4k of unbilled weekend hours over a single year from agent-led commits that humans had to honour.
Mode B: ownership lands on the named on-duty account manager. If she replied "let's pick this up January 6th", that is the commitment. If the client pushes, the agency has a paper trail of a human decision, made by a named person, inside business hours. The commit is defensible.
Mode C: ownership lands on whoever signs the Friday handover. The agent's holding reply on December 27th, "we have received your brief, we will confirm scope and dates on our handover call on Friday January 2nd", buys you the week. The Friday call commits the work. The principal is named. The freelancer, if there is one, joins the call.
Mode A: 1/5 — ambiguous, agency absorbs by default
Mode B: 5/5 — named human, dated decision
Mode C: 4/5 — defensible, with one weekly bottleneck
The weighted score
TTFD Defens. Scope Weighted
Weight 0.30 0.45 0.25
Mode A 5 2 1 2.65
Mode B 2 5 5 3.80
Mode C 4 5 4 4.45
For a sub-€8M Dutch agency with at least one freelancer DGA in the chain and roughly 36 monthly projects, Mode C wins because it gives the agent everything it is good at and gives the human the only thing that has to stay human: the commit.
If your agency is purely employee-staffed, no freelancers, fixed-price intake, drop the defensibility weight to 0.15 and Mode A starts winning. That is the only configuration where it does.
What changes inside the agent's system prompt
The trick that makes Mode C work is that the agent has explicit hard rules about what it will not commit to. The prompt is the contract.
RULES — agent must never commit to:
- a delivery date
- a fixed scope
- a price
- a freelancer's availability
The agent MAY:
- acknowledge receipt with a timestamp
- reference the next Friday handover slot
- draft a scope into a shared doc
- ask clarifying questions
Every outbound message ends with:
"We confirm scope and dates on Friday {date} at 10:00."
That closing line is the entire defensibility argument in writing.
When we built the intake-agent for a Utrecht-based design studio with twelve monthly retainers and three freelance DGAs in the chain, the gotcha was exactly this: the agent's tone was too eager. It wrote "we'll get back to you tomorrow" because that sounded warmer than "Friday". We solved it by hard-coding the Friday timestamp into the closing line with no model freedom on that sentence. If you want the same shape in your own stack, the AI agents we ship for agencies treat the commit gate as a template and the conversation as generative.
Sit down Monday with the rubric above. Score your current onboarding on those three rows. If your defensibility column is a 2, change one thing this week: add a hard rule that the agent's last sentence is always a date you control.
Key takeaway
For a sub-€8M Dutch agency with freelancer DGAs in the chain, the agent drafts and a named human commits — the Friday call is the contract gate.
FAQ
Does the Wet DBA actually apply when an AI agent acts on a freelancer's behalf?
The Wet DBA judges the working relationship between commissioning party and worker. An agent making autonomous commitments blurs that relationship and weakens a model agreement at audit.
What if our agency has no freelancers, only employees?
Drop the defensibility weight in the rubric. With a fully employee-staffed book and you as principal, the pure-agent mode becomes viable and the speed advantage carries the decision.
How long does the Friday handover call actually take?
Around 30 minutes for twelve retainers. The pre-read is the agent's draft scope docs, so the call is a sign-off and exception list, not a discussion from scratch.
Can the agent commit to a date if the freelancer's calendar is integrated?
Technically yes, legally still no. The commitment has to be made by the named person on the modelovereenkomst, not by software acting on their behalf.