Promote your first marketing engineer instead of hiring one. Give it to whoever already decides what your data means, usually whoever runs marketing ops or RevOps, with hours off their week, one job and a build partner. A first agent breaks on meaning long before it breaks on code, and that person is the one who notices.
That holds when the first agent takes on work your team already does by hand and the operator wants the role. If nobody owns your definitions yet, the question is which of the three titles to hire. If nothing can come off the operator’s week, hire or wait.
Empact Partners builds these agents. We have worked inside the marketing teams of B2B software partners as their consultancy since 2020, and an agent that routes leads, reports or prospects is our RevOps work. We sell the build. The verdict costs us the rest, because it tells you to keep the agent with your own person instead of paying us to run it.
A Builder Who Does Not Know Your Data Asks What It Cannot Answer
We know this from the wrong side of it. In June we wrote a data request for a B2B software partner: questions their own agent would answer from their database, so that our content work could stand on their numbers. We knew the tools well. We did not know their data, and the first version had 363 questions.
Their marketing lead decides what each term means on their side. He ran our request past the agent their data team already uses and sent back the questions their data could not answer. About 140 survived.
Much of what we cut asked about what happens after a record leaves their product, which their system never reads back. A whole group of questions assumed their product had something it does not have, and we kept using a word that means something else in their data. He had to correct us more than once.
What their team ran on its own
In mid-June their data team handed the request to that agent, and it answered from a read-only replica of their data. We never had access to it, and never needed it. In August, when a new data source reached their system, their team reran the whole set without us: 142 questions, 141 answered, one marked unavailable.
Their lead left seventeen comments on the August results, all in one day. One flagged an answer their own agent had got wrong. Another set the format every later request has to follow: name the exact measure, say how it is broken down, and list every option of the breakdown. What stayed with us was the request itself and every edit to it.
.png)
Look at what their lead did not do. He did not write the request, build the agent or run the queries. He owned the meaning, which was the one part no builder could have supplied. On a first agent that is the marketing engineer’s job, and the code is the part you can rent.
The Person Who Explains The Number Catches What The Formula Hides
The same thing happens after launch, and we have watched it in our own house. The Empact Panel is our own build: it measures how often ChatGPT, Google AI Mode and Claude name more than a dozen of our partners each week. Its lift compares how often an engine names a partner when it cites the partner’s own pages with how often it names them when it does not.
In October Sharné McDonald, a senior GTM consultant on our team, filed a proposal from one partner’s weekly report, where the lift had fallen sharply and was printed as a decline. Nothing had got worse. The partner was being named more often everywhere, which raised the baseline the lift is measured against.
The formula did exactly what it was built to do, and the report told the partner the wrong thing. Sharné spotted the misreading in a number she explains to partners every week, and she could trace it herself because every consultant here is an admin on every panel. Twelve days after her proposal, the report stopped calling that pattern a decline.
“My Ops Lead Has No Time For This”
The objection is right, and usually it is true. The person who knows what your data means is busy precisely because everybody already asks them, and a promotion that adds an agent to their week and removes nothing fails, whatever the new title says.
We have watched access without a job do nothing. In October a partner’s marketing lead told us on a call that he had not logged into the panel we built for his team. He read our consultant’s report instead, which was the sensible use of his week.
The hours an agent will give back cannot pay for the hours before it works, either. Promising them is how a promotion like this gets approved and then fails.
.png)
With a builder beside them, the owner does not build or repair anything. The marketing lead’s whole part in the August rerun was seventeen comments in one day, and those comments now bind every request we send. Sharné’s catch came out of a report she was already reading.
So the time is real, and it has to be found first. Nobody logged the minutes on either side. What the record shows is the shape of the work: one day of rulings on a rerun his team did, not a build he ran.
Give Them Hours, One Job, And A Builder
Write these into the promotion before anything is built:
The split between owner and builder is the part teams leave vague, and it is where the first month goes wrong. On the partner’s data request it divided like this:
| The work | Their marketing lead | Us, the builder |
|---|---|---|
| Which questions the data can answer | Said so, before the first run | Wrote 363 and cut to about 140 |
| What each term means | Ruled on it in writing, binding on every later request | Corrected our questions to match |
| Running it | Their data team and their own agent, twice | No access, none needed |
| The format of the next request | Set it | Follows it |
| Editing the request | Never | Every edit |
A RevOps engagement with us has the same shape. Empact Partners starts with an audit of the work your team repeats, scopes one agent to one job, puts a consultant in charge of what it produces while it is built, and agrees the measure first.
In the first week we also put two things in writing: your operator as the owner of the definitions, and the hours that come off their week before the build starts.
Hire The Builder For The Second Agent
Once the first agent has an owner, hire for tool fluency, because by then you know what the builder will build and whose answers they will build it on. Hired before that, the builder is a stranger learning your definitions on your data, and your ops lead is the one teaching.
If your first agent is waiting on a hire and the person who knows your data already works for you, book a time with me, and we can work out together whether Empact should build it beside them.
