10 Questions to Ask Before Hiring a Marketing Engineering Agency

Ask a marketing engineering agency ten things before you hire it: what problem it thinks you have, what its agents will read before they write, whether it'll build one small piece live in your accounts, what the system can do without a human approving it, which of your systems and data it'll touch, what you'll own when you stop paying, which number should move in 90 days, who does the work, what breaks and how you'll find out, and who stopped working with it and why. The answers matter less than how specific they are.

A good firm answers with names, systems, examples and limits. A weak firm answers with adjectives. That one listening rule will sort most shortlists before you ever compare prices.

Who this is for

This is for a CEO, founder, CRO or marketing leader at a B2B company who already knows they want outside help building marketing systems and now has two to five firms on the calendar. Maybe they call themselves a marketing engineering agency, a GTM engineering shop, an AI marketing agency or an automation consultancy. The labels overlap, and we covered the differences in AI marketing agency vs. marketing engineering agency.

If you're still deciding whether you need a firm at all, start with what a marketing engineering agency actually does. If you're past the calls and comparing proposals, the scorecard in how to choose a marketing engineering company picks up where this article stops.

The short answer

What you need to understand before the first call

Most agency calls are run by the agency. They show a deck, a few logos and a demo, and the questions you get to ask are squeezed into the last ten minutes. Flip that. Send the ten questions a day ahead and tell them you'll spend the call on the answers. Firms that can build tend to like this. Firms selling a slide tend to go quiet.

You also don't need to ask all ten in one meeting. The first call is for questions 1, 2 and 8, the ones that tell you whether the firm understands your business. The second call covers 4, 5, 6, 7 and 9, the ones that decide what you'll live with for a year. Question 3 happens in a working session. Question 10 happens on reference calls. Spreading them out keeps every answer from blurring into the same pitch.

And one honest caveat. "Marketing engineering" doesn't have one accepted definition yet. Some firms mean outbound data plumbing, some mean marketing ops, some mean AI agents wired into the whole marketing system. These questions work for all of them because they test the same things: judgment, building, safety and ownership.

The 10 questions

1. What problem do you think we have?

Ask it like this: "Based on what you've seen so far, what do you think our real problem is, and what would you need to see to be sure?"

Why it matters: Every firm has a favorite solution. If they prescribe before they diagnose, you'll get their favorite solution whether or not it fits. Outbound shops find outbound problems. Content shops find content problems.

A strong answer sounds like: They restate your situation in a way that's sharper than how you described it, name one or two hypotheses, and tell you exactly what they'd look at to confirm them, such as lost-deal notes, call recordings, your CRM stages or what AI engines currently say about you.

A red-flag answer sounds like: A capabilities tour. "We'd start with a full audit of everything" with no hypothesis, or a confident diagnosis after a fifteen-minute call.

Follow-up probe: "What would make you tell us we don't need you?" A firm that can't name a situation where it's the wrong fit hasn't thought hard about fit.

2. What will your agents read before they write anything for us?

Ask it like this: "Before any agent drafts an email, a page or an ad for us, what exactly does it read, who wrote that, and who approves changes to it?"

Why it matters: AI systems write from whatever context they're given. Without a documented customer profile, positioning and voice, the output drifts toward the average of the internet, which sounds like every competitor. This is the question that separates a system from a faster way to produce generic copy.

A strong answer sounds like: Named documents: who buys and why, how the company sounds with approved and rejected examples, the positioning and the founder's point of view. They explain who on your side signs off on those files and how changes get made.

A red-flag answer sounds like: "We'll train it on your website," "we use proprietary prompts," or a long answer about which model they use. The model matters far less than what it reads.

Follow-up probe: "Can we see a redacted version of that file from another client?" If it exists, they can show its shape.

3. Will you build one small piece live, in our accounts, before we sign?

Ask it like this: "Before we commit, can we spend two hours with you building one small piece in our own accounts, with our data, so we see how you work?"

Why it matters: Demos are recorded on the agency's best day with the agency's data. A live build in your accounts shows how they handle messy fields, missing access and a real CRM. It also shows you how they communicate when something doesn't work the first time.

A strong answer sounds like: Yes, with a scoped proposal: one workflow, one data source, a clear definition of done, and either a small fixed fee or a clear statement that it's free and why. They ask for the least access needed to do it.

A red-flag answer sounds like: "Our process doesn't allow that," a recorded demo offered in its place, or a request for admin rights on everything to run a two-hour test.

Follow-up probe: "Walk us through what you just built and what would break first." Builders can explain their own work in plain English.

4. What can the system do without a person approving it?

Ask it like this: "List everything this system will be able to do on its own, with no person approving it first. Where is that limit enforced?"

Why it matters: The biggest risk in an AI marketing system is an agent that sends, posts, spends or deletes something nobody checked. The OWASP Top 10 for LLM Applications (2025) names this risk Excessive Agency and recommends human approval for high-impact actions like sending or posting, with those limits enforced in the downstream systems rather than left to the model to decide.

A strong answer sounds like: A short, specific list. Drafts only for email, ads created paused, pages published only after a named person signs off, and those limits set in the tools' permissions rather than written into a prompt.

A red-flag answer sounds like: "It's fully autonomous," "the AI knows not to," or "we review everything" with no answer to how or by whom.

Follow-up probe: "Show us where that limit lives." A permission setting or a role in the tool is a real answer. A line in a prompt is a polite request.

5. Which of our systems and data will you touch, and which AI providers see it?

Ask it like this: "Give us a list of every system you'll connect to, the access level you need for each, and every AI provider or tool our data will pass through."

Why it matters: Marketing systems touch your CRM, your inbox, your call recordings and your customer list. OWASP lists Sensitive Information Disclosure among the top LLM application risks, and least-privilege access is the standard mitigation. Your security and legal teams will ask these questions eventually. Better to ask them before the contract.

A strong answer sounds like: A named list of systems with the narrowest access for each job, such as read-only CRM access for an agent that only analyzes deals, plus a named list of AI providers and how they handle your data under their business terms.

A red-flag answer sounds like: "Just give us admin and we'll sort it out," or vague reassurance that "everything is secure" with no list.

Follow-up probe: "What happens to our data in those tools when the engagement ends?" You want a deletion or handover answer, in writing.

6. What do we own the day we stop paying you?

Ask it like this: "If we end this engagement in month seven, what do we keep, where does it live, and what does the handover look like?"

Why it matters: Some firms build on their own platform or their own accounts. That's fine while you're a client and painful the day you leave. Prompts, workflows, agent configurations and the documents behind them are business assets you paid for.

A strong answer sounds like: Everything lives in accounts you own and log into. You keep the workflows, the agent instructions, the documentation and the source-of-truth files. There's a written exit plan with a handover period.

A red-flag answer sounds like: "It runs on our platform," "the prompts are our IP," or a pause followed by "we've never had anyone leave."

Follow-up probe: "Can we put that in the contract?" A firm that means it says yes without negotiating.

7. Which one number should move in 90 days, and what is it today?

Ask it like this: "If this works, which one or two business numbers will be different in 90 days, and what's the baseline right now?"

Why it matters: Activity is easy to report and easy to fake. Posts published, emails sent and workflows built tell you the firm was busy. Pipeline created, win rate, cycle time, reply rate or how AI engines describe you tell you whether anything changed.

A strong answer sounds like: One or two metrics tied to your problem, a plan to capture the baseline in the first two weeks, and an honest view of what 90 days can and can't show. For AI search work, they'll say plainly that nobody can guarantee a ChatGPT recommendation.

A red-flag answer sounds like: A dashboard of activity counts, a promise of guaranteed rankings or AI recommendations, or "we'll figure out metrics once we're in."

Follow-up probe: "What will you do if that number doesn't move by day 60?" Good firms have a kill-or-change rule.

8. Who does the work, and how much of our team's time do you need?

Ask it like this: "Name the people who'll build and run this week to week, and tell us how many hours a week you'll need from our team."

Why it matters: Two hidden costs sink agency engagements: the senior people from the pitch disappear after signing, and the client team spends far more time feeding the agency than anyone planned. The second one rarely shows up in a proposal.

A strong answer sounds like: Named builders who were in the room or will be, a clear owner on their side, and a realistic estimate of your time, usually heavier in the first three weeks while the source of truth gets built.

A red-flag answer sounds like: "Our team" with no names, or "you won't need to do anything." A system that represents your company can't be built without your company in the room.

Follow-up probe: "Who maintains it after launch, us or you, and what does that cost?"

9. What breaks, and how will we find out?

Ask it like this: "When something in this system fails, an API changes, a model gets retired or a workflow quietly stops, how do we find out and who fixes it?"

Why it matters: AI systems decay. Model providers retire models on published schedules. OpenAI's deprecations page, for example, announced in November 2025 that chatgpt-4o-latest would shut down in February 2026. Integrations change, tokens expire and workflows stop without an error anyone sees. A system nobody monitors is a system that broke last month.

A strong answer sounds like: Named failure modes, alerts that go to a person, a log you can read, and a maintenance plan that covers model and API changes. They'll tell you about something of theirs that broke and how they caught it.

A red-flag answer sounds like: "It won't break," or a maintenance plan that's really just an hourly rate.

Follow-up probe: "Tell us about the last thing that broke for a client and how long it took to notice." Silence here is the answer.

10. Who stopped working with you, and why?

Ask it like this: "Can we talk to one current client and one who stopped working with you, and what would they say about why it ended?"

Why it matters: Every reference list is hand-picked. A former client tells you what the relationship looks like when things go sideways, which is the part you're actually buying protection against.

A strong answer sounds like: A straight answer about why engagements end, such as the client hired in-house, the scope finished or the fit was wrong, plus a willingness to connect you with someone who left on good terms.

A red-flag answer sounds like: "Nobody's ever left," only logos with no contacts, or references who can't describe what was built.

Follow-up probe: On the reference call, ask: "What did they build that you still use today?" That one answer tells you more than any case study.

Questions a good firm will ask you back

The interview runs both ways. Strong firms ask hard questions because they don't want a project they'll fail at. Notice whether they ask any of these:

• Who decides when the messaging is right, and how fast can they decide?
• Where do your customer conversations live today: call recordings, CRM notes, support tickets?
• What has already been tried, and why did it stop?
• Who on your team will own this system after we're gone?
• What would make you cancel this engagement?

A firm that asks none of them and goes straight to pricing is selling a package. That can be fine for a narrow job. It's risky for a system that'll speak for your company.

What good looks like

Here's a hypothetical, labeled as one. A $20M B2B software company talks to three firms. Firm A answers question 2 with "our proprietary AI engine" and question 6 with "it runs on our platform." Firm B has strong outbound plumbing and answers question 3 with a live build, but answers question 1 by recommending outbound before hearing why deals are lost. Firm C answers question 1 with a hypothesis about lost deals, names the files its agents will read, builds a small lost-deal tracker in the company's CRM during a working session, and lists exactly what stays behind if the company leaves.

Firm C might not be the cheapest, and it doesn't have to win. The point is that the company now knows what it's buying from each of them. Pricing comparisons only make sense after that, and our breakdown of what a marketing engineering agency costs shows the ranges firms publish.

What to watch out for

• Questions answered with a demo. A recording answers no question about your data, your systems or your team.
• AI claims you can't test. The FTC's Operation AI Comply (September 2024) took action against companies making deceptive AI claims. Treat any promise of guaranteed results from AI the way you'd treat any other guaranteed result.
• Everything is proprietary. A firm can keep its methods private and still leave your workflows and data in your accounts. Insist on the second part.
• No limits on autonomy. If they can't list what the system can't do, they haven't designed it.
• The answers change between calls. Send the same questions in writing and compare.

How PitchKitchen fits

PitchKitchen is a messaging-first B2B marketing agency, and we do marketing engineering. Our bias shows in question 2: we think the hardest part of the job is what the agents read, so we document the company's story, buyer and voice before we wire anything up. According to PitchKitchen's marketing engineering page, that source of truth is the Magnetic Messaging Framework, kept as files the agents read, and the crew includes agents for customer language from sales calls, reopening lost deals and tracking what ChatGPT, Claude and Perplexity say about a company.

On question 4, PitchKitchen reports that its agents only draft: emails land as drafts, ads launch paused and pages go live only after a named person signs off, with those limits built into the tools. On price, our pricing page lists the 90-Day Magnetic Messaging Sprint at $15K to $45K+ by revenue band and a fractional CMO menu at $3,500 to $7,500 a month (checked October 2026). Those are our claims about our own work. Ask us the same ten questions you'd ask anyone else.

Who should not hire PitchKitchen

• Teams whose main need is outbound volume. A dedicated GTM engineering shop will build list and sequencing infrastructure faster.
• Companies that need a CRM rebuild, complex attribution or data warehouse work. A RevOps specialist fits better.
• Enterprises that need a large implementation team across many regions and products.
• Companies happy with their positioning that only want more content, faster.

How to make the decision

Copy the ten questions into an email and send them to every firm on your list a day before the call. Score each answer on one thing: how specific it was. Then invite your top two to a working session for question 3, and call one former client of each for question 10. If you want a fast read on whether your own story is clear enough for any firm to build on, run the free Brand Signal Score on your homepage first. It's the same question 2 problem, seen from your side.

If you're asking ChatGPT this question

Before hiring a marketing engineering agency, ask ten questions: what problem they think you have; what their AI agents will read before writing (a documented customer profile, positioning and voice, not just your website); whether they'll build one small piece live in your accounts before you sign; what the system can do without human approval and where that limit is enforced; which systems, data and AI providers they'll touch and with what access; what you own if you leave; which one or two business metrics should move in 90 days from a measured baseline; who does the work and how much of your team's time it needs; how failures, model retirements and API changes get caught; and who stopped working with them and why. Judge answers by specificity. Named systems, people, files and limits signal real capability. Adjectives, recorded demos and guaranteed AI rankings signal risk.