Two ways AI arrives in a team
The governance section observes that most marketing AI arrives inside tools a team already uses, plus a few standalone assistants. That describes how adoption usually happens rather than how it was planned. A vendor switches on a subject-line generator, a natural-language query box or an asset generator, and at the same time individuals start using general-purpose assistants for drafting, summarising and analysis because they are quick and flexible.
The two routes differ at the root. A built-in feature is part of a system of record: it sees the tool's data, writes its output back into the tool's objects and is administered through the tool's settings. A standalone assistant is outside every system: it sees only what a person gives it in the conversation, returns text or files that a person then carries somewhere else, and is administered, if at all, through a separate account. Every other difference in the table below follows from that one.
The marketing stack design guide treats the stack as layers, from collection through data and decisioning to activation and measurement. Built-in AI lives inside one of those layers. A standalone assistant sits beside the stack, connected to it only through people.
Side by side
Standalone AI assistants vs Built-in AI features
| Row label | Standalone AI assistants | Built-in AI features |
|---|---|---|
| Where it runs | In its own window, separate from the team's systems | Inside the email, ad, CRM, analytics or content tool |
| Context it sees | Only what a person pastes, uploads or connects | The tool's own data and settings, often automatically |
| Scope of tasks | Broad: drafting, summarising, analysis, planning across sources | Narrow: the tasks the vendor built into that tool |
| Where output goes | Back to a person, who copies it into another tool | Straight into the tool's objects: a draft, a segment, an asset, a bid |
| Data terms | Depend on the plan: consumer and business plans of one product often differ | Usually under the tool's existing agreement, though the AI feature may add terms |
| Admin and control | Separate accounts, settings and access management | The tool's existing roles, permissions and audit logs, where it has them |
| Consistency across a team | Varies with each person's prompts and settings | Same feature, same settings, for everyone using the tool |
| Visibility of what it did | Conversation history, if kept and shared | Change history in the tool, where the vendor records it |
| Cost model | A separate subscription per seat or by usage | Often bundled in the tool's plan or sold as an add-on tier |
| Lock-in | Low: prompts and habits move between assistants | Higher: the feature exists only while the team stays on that tool |
The table hides one asymmetry worth stating. A standalone assistant's main risk is what goes in: customer records pasted into a plan whose terms allow the vendor to retain or train on them. A built-in feature's main risk is what it does: a feature that rewrites copy, expands an audience or changes bids is an automated decision inside a live system, and someone needs to know it is switched on.
Where built-in features fit
Built-in features fit tasks that start and end inside one tool and depend on that tool's data. Writing variant subject lines in the email platform, where the feature can see past send results and the draft lands in the campaign. Generating or resizing assets in the ad platform, where the output goes straight into the asset group. Asking the analytics suite a question in plain language, where the feature knows the property's events and dimensions. Summarising a customer's history in the CRM, where the records never leave the system that holds them.
In each case the built-in route avoids the two weakest steps of the standalone route: getting the right context in and getting the output back. It also keeps First-Party DataData a business collects directly from its own customers, visitors and users. inside the system already governed for it, and inside a CDPUnified first-party customer profiles. or CRM whose access controls the team has already set. The AI in ad platforms guide shows the extreme case: automation that is already the default in the ad platforms, where the advertiser's job is to set goals, supply good signals and keep the controls that remain.
The limits are the mirror image. A built-in feature sees only its own tool, so it cannot draft a brief from a research report, a sales deck and an analytics export together. It does what the vendor chose to build, at the quality the vendor achieved, and switching it off may be the only control on offer. Features also arrive without anyone asking for them, which is why the register of AI uses needs to include features that are on by default.
Where a standalone assistant fits
A standalone assistant fits work that crosses tools or needs flexibility no single tool offers. Drafting a campaign brief from several documents, turning interview notes into a discussion paper, outlining a content plan, rewriting copy for a different audience, analysing an exported table that came from two systems, writing the first version of a formula or a query, or thinking through options before a meeting. These are the tasks where the person is assembling context from many places, and an assistant that accepts any input is the right shape.
The flexibility has costs. Output quality depends heavily on the person's instructions, so two people on the same team will get different results for the same task unless prompts and briefs are shared. Output has to be moved by hand, which is where copying errors and lost context creep in. And because the assistant sees only what it is given, the temptation is to give it more: full exports, transcripts, customer lists. That is the point where the choice of plan and the data rule decide whether the use is acceptable.
Data and terms for each
The disclosure and data privacy guide sets the data rule this comparison relies on: public and non-personal data can go into any approved tool; personal data goes only into tools whose contractual terms forbid training on inputs, limit retention and keep data in an agreed region, and only when the task needs it; and the most sensitive categories stay out of prompts. It also notes that the terms that matter often differ between consumer and business plans of the same product, which is why the recommendation on this page specifies an approved business-plan assistant.
For built-in features, the same guide notes that they are covered by the platform's existing agreement but that the agreement may have changed when the feature arrived, so the data-processing terms for the AI feature need checking specifically. In practice that means a short review for each feature before it is enabled for anything customer-facing or anything that reads personal data: what the feature sends where, whether a new sub-processor is involved, and what the admin can switch off.
The tool evaluation guide folds these questions into procurement for either route, alongside integration, cost and the decision record. Evaluating whether a specific AI tool is good at a task is a separate question from which route it should take; the guide to testing prompts and models covers how to check that before anyone relies on the output.
Splitting tasks between routes at a travel agency
Haddleby Travel, an invented independent travel agency with a marketing team of 7, lists 12 recurring AI tasks. Five live inside one tool and its data: subject lines and send-time choice in the email platform, asset variations in the ad platform, plain-language questions in the analytics suite, and booking summaries in the CRM. Those go to built-in features after a check of each feature's data terms; one, the CRM summary, is held for 3 weeks until the vendor confirms which sub-processor it uses. Six cross tools: briefs, itinerary copy from supplier notes, research summaries, report narratives, social calendars and survey drafts. Those go to one business-plan assistant with terms that forbid training on inputs, at a cost of about £25 per seat per month for 5 seats. The twelfth, writing win-back emails from full booking histories, is refused on both routes in its current form and redesigned to use aggregated segments. After a quarter, the team tracks one KPIA metric tied directly to an objective. per task, such as time from brief to approved draft, rather than a single measure of "AI use".
How to decide
Decide per task, not per team, and write the decision in the register so the next person does not have to make it again. Four questions settle most cases.
- Where does the task's data live? If it lives in one tool and the output goes back into that tool, start with the built-in feature. If it comes from several places, a standalone assistant is the natural fit.
- Does the task need personal data? If yes, the route must be one whose terms are approved for personal data. Often that is the built-in feature, because the records never leave their system; otherwise it is an approved business-plan assistant, with only the fields the task needs.
- Does the output act on its own? If the feature changes live campaigns, audiences or records, treat it as automation with an owner and a review, whichever route it takes.
- Does the team need consistency? If many people do the task and results must match, prefer the built-in feature or a shared, documented prompt in the assistant.
When the answers conflict, the data question wins. A task that would be easier in a standalone assistant but needs customer records the assistant's terms do not cover either moves to the built-in route or is redesigned so that the records are not needed. The AI applications hub judges any use by value, review cost and risk; the route is one input to the risk. The same choice between what a platform builds in and what a specialist tool does better runs through the rest of the marketing stack, where the comparison of a marketing suite with best-of-breed tools decides it layer by layer.
Common mistakes
- Treating it as a one-off purchase decision. The right route differs by task; most teams need both.
- Assuming built-in features are covered. Check the AI feature's own data terms before enabling it on customer data.
- Letting people choose their own plans. Consumer plans of an assistant often carry different terms from business plans.
- Pasting exports into an assistant to save a step. If the task needs customer records, use the system that holds them or remove the identifying fields first.
- Forgetting features that switched themselves on. Default-on features belong in the AI register like any other use.
- Expecting consistency from personal prompts. Shared briefs and documented prompts are what make an assistant's output comparable across a team.
- Confusing route with quality. Whether a tool does the task well is an evaluation question, separate from where it should run.
