ADA Tips & Tricks: The Right Model for Each Task

The problem

One model is no longer enough

For years, adding AI to marketing analysis meant plugging in an assistant tied to a single model. It was fine to get started. But the pace at which new models arrive left that approach behind.

Today ADA's model picker lists fourteen models from four providers (Anthropic, OpenAI, Google and xAI) side by side on the same screen. And the list keeps moving: while ADA Flash runs on Gemini 3.7 Flash, the catalog already offers 3.8. Every release brings something new: a larger context window, better image reading, more speed or lower cost.

The trouble with a single model is lock-in: when a better one shows up for your task, you are stuck until someone rebuilds the integration. You are always one version behind.

The best model of the month stops being the best the following month. Marrying one is betting the market will stop improving.

The architecture

ADA is not a model: it is what picks the model

ADA is multi-model, or model-agnostic. Put simply: several frontier models can run underneath the assistant, and ADA decides which one to use based on what you ask. It is not a chatbot wired to one LLM; it is the layer that orchestrates all of them.

Two terms worth separating, because they get confused constantly. Multi-model is this: several models available, one per task. Multimodal is something else: a single model's ability to understand text, image and video at once. ADA is multi-model and, on top of that, leans on multimodal models when the task calls for it, such as reading a creative.

That orchestration works in two ways:

• ADA on its own. It operates unassisted. It decides which model to use to solve each task with the lowest possible credit consumption: you ask, ADA handles the rest.
• ADA in conversation. You pick the model. When a specific task needs a particular profile, you select it by hand from the assistant itself.

In conversation mode, ADA presents its engines as named tiers, each backed by a different model:

• ADA Flash · Gemini 3.7 Flash: high volume, low complexity
• ADA Pro · GPT-5.6 Terra: the default engine
• ADA Max · Claude Sonnet 5: deep reasoning
• ADA X · Grok 4.6: real-time signal from X

…plus an open catalog of individual models for cases where you need maximum depth or the lowest possible cost.

The architecture changed with ADA 2.0 and brought three things you notice in daily use. Transparency: you can see what actions the assistant runs, what information it pulls and what context it builds the result from. Conversational memory: it remembers the previous exchange, so you do not have to repeat your instructions in every message. And smart widget updates: when you ask to adjust a chart, ADA keeps the procedure and the logic it originally used instead of starting over.

Behind every tier there is a criterion, not a whim. And when ADA works on its own, that criterion is applied automatically: choosing well is the part you never see, and it is exactly where the value sits.

The comparison

Every model is good at something different

An important clarification: it is not that each model is useful for one thing only. They can all do almost everything. The difference is in how much they consume to get there and which task they handle best. Choosing well is, above all, a question of efficiency.

Relative consumption is a ratio, not an exact credit value: it shows how much each model weighs against the others. Multipliers from ADA's model picker, September 2026.

The pattern is clear. To read what an ad shows (which people, objects, colors or logos appear), a model with strong multimodal capability like Gemini detects more and detects it better. For a simple, repeated query, an inexpensive model clears it for a fraction of the cost. And for analysis that leaves no room for error, Claude's ecosystem holds long reasoning better.

There is one case where efficiency wins by name: when analyzing X, it does not matter which model you are chatting with. ADA turns to Grok to gather the information, because that is the route that optimizes credits on that source.

The rule

The rule that saves you credits

If you take one thing away from this article, make it this: always start with ADA Pro. It is the default engine and it is built to cover most of a marketing team's daily work with the best quality-to-cost ratio. Move up a tier only when a specific task proves it needs to.

It sounds obvious. It is not: the natural reflex is to leave the most powerful model selected «just in case», and that is where credits disappear without anyone noticing.

The three moves, then:

• Drop to Flash when volume rules and the task is simple. Classifying hundreds of comments, tagging, translating, summarizing meetings, a first moderation pass. Paying for deep reasoning to tag things is throwing credits away.
• Stay on Pro for 90% of the day: campaign analysis across Google Ads, Meta, TikTok or GA4; copy and content; reading tables and KPIs; benchmarks and competitors; executive presentations.
• Move up to Max when the cost of an error is high: churn, Go-to-Market, attribution or pricing models, annual planning, multi-source research synthesis. The test is simple: if redoing the work costs more than the analysis, use the good engine.

And the open catalog comes after Max, not before. It is surgical use for specific cases, not a replacement for the normal path.

The deliverables

Ask for the deliverable, not just the answer

Most teams use ADA as a search box with better manners: ask, read the answer, copy and paste. It is the poorest possible use. ADA 2.0 delivers in five different formats, and you name the format in the prompt.

• An executive answer in chat. Direct and actionable, without unnecessary jargon. For settling a specific question.
• An interactive chart. It drops into the conversation, ready to consult or pin to a panel.
• A structured table. A tabular view ordered by metric relevance, for when you need to compare rows rather than read prose.
• A full visual report in HTML. A self-contained document with editorial design, KPI cards, embedded charts and a business narrative. It saves inside Bunker, with no external tools.
• A data export. CSV or XLSX when the volume is large and the destination is another tool.

The HTML report is the deliverable that changed the most, and not only in what it says: also in how it looks. That deserves its own section.

The design

Your report no longer looks like the tool that made it

There is a cost almost no team accounts for. The platform delivers a report that is correct but generic, and someone spends a couple of hours rebuilding it in Slides with the brand's identity before presenting it to the client. That work adds not a single conclusion to the analysis: it merely translates it into another format.

ADA now generates the report with the brand identity already applied. Design stopped being an afterthought and became one more parameter of the deliverable.

What you set before generating it:

• The logo. You upload it as PNG, JPG or SVG and ADA extracts the palette from it. You do not need to know your brand's hex values.
• The colors. Primary, secondary and tertiary, plus titles and buttons, either by hand or by picking from a set of ready-made palettes. Background, cards and text derive themselves.
• The background. Flat, gradient or mesh.
• The cards. Solid, glass, gradient, outline only or no box at all.
• The typography. Poppins, Playfair Display, the system typeface or any Google Font.
• The rhythm. Density and corner radius, on two sliders.
• The navigation. Horizontal, one section at a time with previous and next, built for presenting live. Or scroll, the whole report on one page, where tabs become anchors.

If you would rather not decide every detail, ADA suggests. Below the canvas you get design shortcuts ready to apply one at a time: dark mode, gradient background, glass cards, parallax, bigger titles, rounder corners, softer colors. You try them, see the result and move on.

And when no option in the panel does exactly what you have in mind, you ask for it in words. Customization is conversational too: you describe the change on any element of the report and ADA applies it. It is the same assistant, working on the aesthetics instead of the data.

In practice

Seven ways to get more out of ADA

None of this is advanced. These are the things that separate a team that uses ADA from one that gets real leverage out of it.

1. Start with Pro and escalate for a reason. Pro is the default engine and it handles almost everything day to day. Drop to Flash if the task is simple and repetitive; move up to Max if an error would cost you. It is the habit that saves the most credits over a month.
2. Stop repeating the context. ADA 2.0 has conversational memory: it remembers the previous exchange. Instead of rewriting the full brief, chain your asks: «now the same thing but for Meta only», «compare it against the previous quarter». Every unnecessary repetition burns credits.
3. Review the process before you share. ADA shows which actions it ran and what information it pulled to build the result. When the analysis feeds a sensitive decision, open it up and check where the data came from. It takes thirty seconds and it is what separates an insight from a well-formatted mistake.
4. Ask for the fix, not the rebuild. If a chart did not come out the way you wanted, ask for the correction on the same widget: ADA keeps the original procedure and logic. Starting from scratch pays for the work twice and usually gives a less consistent result.
5. Name the format in the prompt. «Give me this as a table», «build an HTML report using the brand palette», «export it to CSV». Asking for the right deliverable up front saves you the round trip.
6. Schedule what you repeat. If you ask for the same weekly report every Monday, turn it into a scheduled task: it runs on its own daily, weekly or monthly over relative periods (the last full week, the closed month) and goes out by email to the recipient list you define. The report is waiting for you.
7. Save and share the prompts that work. When you land on the one that gives you exactly the analysis you need, save it and switch it on for the rest of your brand. One analyst's know-how stops being theirs and becomes the team's.

Two operational notes to close. Data stays in memory for up to an hour, so if you come back to a conversation much later it is worth pulling it again. And consumption counts both what you ask and what ADA answers: a vague prompt that produces three pages of text costs more than a precise one that produces half a page.

The criterion

More power is not always the answer

If every model can do almost everything, why not always use the most powerful one? For two reasons: cost and fit.

Asking an inexpensive model for a complex analysis can get expensive in a different way. A lightweight model, pushed past what it handles well, tends to hallucinate: it returns an answer that sounds convincing but is wrong. That is why, when the task is demanding, it is worth checking both where the data came from and how it was used.

At the other end, a top-capability model sharply reduces that risk on complex tasks and gives high-confidence answers for sensitive decisions. But using it for a trivial question is overpaying for nothing.

It is not about using the most powerful AI. It is about using the right one.