Custom AI Marketing Solutions: Wrappers vs Engines

Why this distinction matters

"Custom AI marketing solution" can describe almost anything. At one end is a web page with a text box that sends your input, plus a hidden prompt, to a general model and shows the result. At the other is a system that knows your business, reads your data, follows defined processes, checks its own output and reports on results. Both may be sold with the same words and similar demos.

We call the first a wrapper and the second an engine. Neither term is an industry standard; they are useful labels for a real difference. A wrapper is not automatically bad. It is bad when you pay engine prices for it, or rely on it for work that needs an engine.

Four options side by side

General assistant, wrapper, engine and managed service compared

Criterion	General assistant used well	Wrapper	Engine	Managed service
What it is	ChatGPT, Claude or similar with your own briefs and instructions	An interface and fixed prompts over a general model	A system with retained context, data connections, processes, checks and measurement	A provider runs the work using its own systems and people
Knows your business	Only what you give it each time or store in a project	Only what the form asks for	Yes: stored and maintained context	Yes: the provider maintains it
Uses your real data	If you paste or upload it	Rarely	Yes, through connections	Yes, with your permission
Checks its own output	No; you check	Rarely	Yes: rules for claims, format and brand	Depends on the provider: ask what checks and human review run before work reaches you
Measures outcomes	No	No	Yes, if built to	Should be part of the service: ask to see a sample report
Main risk	Output quality depends on your briefing time	Paying engine prices for a prompt	Maintenance falls on whoever built it	Weak fit or vague reporting leaves you unable to judge the work
Who maintains it	You	The vendor, if they still exist	You or the builder	The provider
Right for	Individuals and small teams with time to brief well	Narrow, low risk, occasional jobs	Repeated work where quality and consistency drive results	Businesses that want outcomes without operating a system

What makes something an engine

Five parts, each of which a wrapper usually lacks.

1. • Retained context. Your offers, audience, proof, voice and exclusions stored and maintained, so every output starts from the same truth.
2. • Data connections. Access to analytics, Search Console, ad platforms or the CRM, so work is based on what is happening rather than assumptions.
3. • Defined processes. The steps for each job written down and followed, so the same job is done the same way.
4. • Checks. Rules applied before output reaches a person: claims against an approved list, character limits, banned words, required sources.
5. • Measurement. A link from output to outcome, so the system (and you) can see what worked.

An engine can be simple. A well organised folder of context files, a few routines and a weekly report can meet all five. Complexity is not the test; the five parts are.

Ten questions that reveal which you are being sold

1. • What model or models does it use, and can that change without notice?
2. • Where is our business context stored, and how do we update it?
3. • Which of our systems does it connect to, and is access read only?
4. • What checks run before output reaches us?
5. • How does it stop invented facts or claims?
6. • What does a log of its actions look like?
7. • What does it measure, and how is that reported?
8. • Where does our data go, is it used for training, and who can see it?
9. • What happens to our context and outputs if we stop?
10. • What could we not do with a general assistant and a good brief?

Question ten is the most revealing. If the honest answer is "not much", you are looking at a wrapper. That may still be worth buying for convenience, at a wrapper's price.

When a wrapper is the right choice

Wrappers are fine for narrow jobs with low risk: generating first draft alt text, reformatting copy into a template, suggesting subject line variations. Anthropic's guidance for builders is to use the simplest solution that works, and for these jobs a wrapper, or a general assistant with a saved instruction, is that solution. Do not pay for an engine to do them.

When you need an engine

You need an engine where the same work repeats often, quality differences show up in results, and mistakes cost money or trust: search content that must rank and be accurate, ad copy that must follow policy and brand, lead handling that must reach the CRM correctly. Here the missing context, checks and measurement are exactly what makes the difference.

Build, buy or have it run

If you need an engine, there are three routes. Build one yourself if you have the skills and time to maintain it. Buy one if a vendor's engine matches your processes closely; check its answers to the ten questions. Or use a managed service, where the provider's engine and people do the work. OWASP's security guidance on language model applications is worth reading whichever route you take: an engine with broad permissions and weak checks is riskier than a wrapper.

AfM is a managed service. We run marketing using our own internal AI operating system, which retains context and supports production, and the founder supplies judgement, prioritisation and review. Clients get managed work and reporting on outcomes; they do not buy or operate the system. If you are comparing this with building or buying, the platform comparison and AI marketing vs agency cover the trade offs.

FAQ

Is every AI tool built on ChatGPT or Claude just a wrapper?

No. Most AI products use a model from a provider such as OpenAI or Anthropic. What matters is what is built around the model: context, data connections, processes, checks and measurement. Those decide whether it is a wrapper or an engine.

Is custom always better than off the shelf?

No. Custom work makes sense when your process is genuinely different or the value of fit is high. For common jobs, a well chosen off the shelf tool is usually cheaper and better maintained.

What happens to a wrapper when the model underneath changes?

Its output can change without warning, for better or worse, because its fixed prompts were tuned to the old model. Engines with test sets can detect this; most wrappers cannot.

Who owns the prompts and context in a custom solution?

Whatever the contract says, so check it. Your business context is your asset; make sure you can export it in a usable form if you stop using the solution.