Instinct for digital marketing: what to ask
Searching Instinct for digital marketing? See which reporting, campaign and content jobs suit an AI assistant and which need a built system.
Founder, Viithiisys

What a search for an AI marketing assistant is really asking
Most people searching for Instinct for digital marketing want a job done with less manual work: reports sent, campaigns checked, copy drafted. The useful question is which of those jobs suits a chat assistant and which needs a system built around your own data.
This post sits in our AI workflow automation practice. It is written for the operations owner at an agency or in-house marketing team whose monthly reporting keeps jamming.
One thing to say plainly first. Instinct is a product we do not make, and nobody has handed us its documentation. This post makes no statement about what it does, what it costs or where it stops. If you are evaluating it, read its own product docs and FAQ, and bring the questions below to the vendor.
Provenance: Viithiisys builds AI workflow automation, so we have an interest in the "build it" answer. This post rests on public platform documentation and holds no client delivery record; the design below is labelled [Reference build].
Where client reporting jams each month
Reporting jams at the point where numbers from different platforms have to agree, and a person has to explain them. The export itself is the easy part.
Picture the first week of the month. An account manager downloads spend and conversions from Google Ads, Meta Ads Manager and GA4, then pastes them into a spreadsheet for each client.
[Illustrative example] Take a hypothetical agency with 20 clients, each on 4 platforms. That is 80 exports before anyone has written a sentence of commentary.
Then the numbers disagree. Each platform counts a conversion under its own attribution settings, and Meta's reporting lets you choose the attribution windows a figure is based on (Meta Marketing API, Insights documentation, 2026). So Meta and GA4 can both be right and still show different purchase counts for the same week.
The client asks which number is true. The account manager stops, investigates, and the other reports wait.

Why pasting exports into a chat assistant breaks down
Pasting an export into a chat assistant works well for one client and one question. It gets fragile when the same task repeats every month across many accounts, because nothing checks the figures the model writes.
Give the option its honest case. For a one-off question such as "why did cost per lead rise in week three", a pasted export and a good prompt is fast, cheap and often useful.
The trouble starts with repetition. Language models can produce statements that read well and are wrong, which is why Anthropic's own guidance suggests grounding answers in supplied material and letting the model say it does not know (Anthropic, Reduce hallucinations, 2026). A wrong figure in a chat is a nuisance. The same wrong figure in a client report is a credibility problem.
A generic reporting SaaS has a fair case too. Many handle the connections well. Ask what they do about commentary, client-specific definitions and your exceptions.
Questions to ask any AI marketing assistant before you buy
Ask four sets of questions, and ask them of every product on your list, including the one you searched for. These are requirements for you to evaluate. We are making no claim about how any named product answers them.
Multi-client reporting
Ask how the product keeps one client's data and commentary apart from another's. A mix-up between two clients is the error an agency can least afford.
Ask whether each client can have its own metric definitions. One client may count a lead at form submit and another at a booked call. Then ask how a report template changes for all clients at once, and how one client can be left out of that change.
Ad-platform data connections
Ask which platforms the product connects to directly, who maintains each connection, and what you see when a sync fails. A silent failure that leaves last week's data in this week's report is worse than an error message.
Some assistants reach outside systems through connectors, including ones built on the Model Context Protocol, an open standard for linking AI applications to external tools and data (Model Context Protocol documentation, 2026). Ask which connectors exist for your platforms today.
Platform access has its own rules. Google Ads API access runs on OAuth 2.0 credentials and a developer token (Google Ads API documentation, 2026). The GA4 Data API meters requests with quotas for each property (Google Analytics Data API documentation, 2026). Ask who holds those credentials and what happens when a quota runs out mid-report.
Brand controls
Ask where each client's voice guide and banned claims live, and who is allowed to edit them. A rule that sits in one person's saved prompt leaves with that person.
Ask how the product treats regulated wording, such as health or finance claims. Obligations vary by country and sector, so take that question to your own counsel as well.
Approval and audit trail
Ask to see the record behind one finished report. It should show which figures went in, who approved the text and when. If a client disputes a number in March, you need to rebuild how it was produced.
The reporting build we recommend
We recommend a pipeline where rules fetch and check the numbers, a model drafts the wording, and a person approves it. [Reference build] This is our client-reporting pipeline design. It is a proposed design, and no client result sits behind it.
- Scheduled pulls from each platform's official API land in one database, tagged by client.
- Rules compare each total with the platform's own figure and flag gaps, missing days and sudden spend changes.
- A model drafts commentary from the checked table only, using that client's voice guide.
- The account manager edits and approves each report before it goes out.
- Anything flagged in step 2 goes to a named person, and that client's report waits.
The exception route matters most. Our data engineering work covers the first two steps, which is where a reporting project usually succeeds or fails. Our AI workflow automation page shows how the approval and exception steps are designed.
A report is only as good as the check that runs before the model writes a word about it.
Which steps need a model and which need plain rules
Our recommendation is plain rules for anything that is arithmetic and a model for anything that is wording. Other splits are defensible, and a simple account may need no model at all.
| Step | Plain rules | A language model | Who handles the exception |
|---|---|---|---|
| Pulling spend and conversions | Scheduled API calls | Not needed here | Engineer who owns the connection |
| Checking totals against the platform | A comparison with an agreed tolerance | Not needed here | Account manager |
| Spotting an unusual week | Thresholds catch the obvious cases | Suggests likely causes for a person to confirm | Account manager |
| Writing client commentary | Templates suit simple accounts | Drafts from checked figures | Account manager approves |
| Applying voice and banned claims | A phrase list catches known cases | Drafts in the client's voice | Brand owner signs off |
Where a step needs the model to act across several systems, read our AI agent development page first. It explains when an agent is worth the extra testing and when a fixed sequence does the job.

What drives the cost of a reporting build
The cost of a build like this is driven by connections, definitions and exceptions far more than by the AI model. We do not publish a price, because each build is scoped after a discovery call.
Four drivers set the size of the work:
- Connections. Every platform is a separate integration with its own login rules, limits and version changes. Our AI integration page describes how that work is scoped.
- Definitions. Each client-specific meaning of "lead" or "conversion" is a rule someone must write down and test.
- Exceptions. Every "except when" in your current process is a branch to build, and teams usually discover more of them than they listed.
- Upkeep. Ad platforms release new API versions and retire old ones, so a connection is something you maintain for as long as you use it.
The part most often underestimated is the definitions work. It needs your account managers' time, and no vendor can do it for them.
When a chat assistant is enough and when it is the wrong tool
A chat assistant is enough when the task is occasional, one person owns the output and a mistake is cheap to catch. Brainstorming ad angles, summarising one export or tidying a draft all fit that description.
It is the wrong tool, for now, when the output goes to a client unchecked or repeats across many accounts. In that case the honest answer is that the job is a workflow first and an AI question second.
Content at volume deserves its own caution. Google's spam policies cover scaled content abuse, meaning many pages produced mainly to influence rankings with little value for readers (Google Search spam policies, 2026). An assistant that makes publishing easier does not change that policy.
So before comparing products, write down the one report or campaign task that hurts most, and who touches it.
Send us the reporting workflow that keeps getting stuck
If one report or campaign task jams every month, describe it to us step by step: who exports what, where the numbers disagree and who ends up fixing it. Our broken workflow assessment takes that description and tells you which steps suit rules, which suit a model and which should stay with a person.
You may find the answer is a reporting tool you already own, used differently. That is a fine outcome, and it is cheaper than buying an assistant for a job that was never a chat question.
Sources
- Meta, Marketing API Insights documentation: https://developers.facebook.com/documentation/ads-commerce/marketing-api/insights
- Anthropic, Reduce hallucinations: https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
- Model Context Protocol, documentation introduction: https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro
- Google, Google Ads API OAuth overview: https://developers.google.com/google-ads/api/docs/oauth/overview
- Google, Analytics Data API quotas: https://developers.google.com/analytics/devguides/reporting/data/v1/quotas
- Google Search Central, Spam policies for Google web search: https://developers.google.com/search/docs/essentials/spam-policies
FAQ
- Can an AI assistant write my monthly client marketing reports?
- An assistant can plausibly draft the commentary once the figures are in front of it. The harder parts are pulling the data on schedule, checking it against each platform and getting a person to approve it. Ask any vendor how those steps work, and what happens when a data pull fails the night before a report is due.
- What should I ask before buying an AI marketing assistant?
- Ask how it keeps one client's data apart from another's, which ad and analytics platforms it connects to, and who maintains those connections. Ask where brand rules live and who can change them. Then ask to see the record of who approved a report and which figures it was built from.
- Do I need a custom build or is a reporting tool enough?
- A reporting tool is often enough when your clients share the same platforms and the same metric definitions. A custom build earns its place when definitions differ by client, exceptions are frequent, or approved commentary has to follow brand rules. Count your exceptions for one month before you decide either way.
- Is it safe to let AI publish marketing content without review?
- Our recommended default is that a named person approves anything a client or the public will read. Language models can write confident statements that are wrong, and Google's spam policies cover content produced at scale mainly to influence rankings. Treat unreviewed publishing as a risk to assess with your own counsel and brand owner.