BAZ26's approach to AI
An integrated practice. Three cases drawn from our daily work, and the method that holds them together. A mindset without which AI produces confusion instead of value.
2026 · state of the art
Every team has its own tool. Every tool has its own subscription. Nobody talks to anyone else. Companies keep telling us the same story: «we have a thousand AIs, we don't know who uses them, how, or why.»
The average company in 2026 reports · 12-18 active AI subscriptions Actual usage · below 30%
Two ways to adopt AI
On the left: AI as most people use it, with tools that don't talk to each other, each with its own password, each with its own cost. On the right: AI as we practice it, with an orchestrator that knows our context and tools that execute specific tasks under its direction.
Common model · 2024-25
Ten tools, ten interfaces, ten subscriptions, zero shared memory. When context changes, every tool has to be re-briefed from scratch. The cost is the cognitive load of holding it all together in the head of whoever is doing the work.
BAZ26 model · 2026
One orchestrator at the center, which knows our world. Below it, specialized tools it runs itself. Context is one, shared. Memory is permanent. Cognitive load goes back where it belongs: on decisions, not on management.
What everything rests on
Every process of ours is built on these three. If even one is missing, the system doesn't hold.
01
The AI has to know who we are. Not through a 200-word prompt, but through a living knowledge base: brand book, site, social, calls, emails, database. Every answer cites the sources it used. Without sources, it's not a valid answer.
02
The AI steps into our flows, it doesn't bypass them. It talks with our tools, doesn't replace them. It works next to the team, not in place of it. The goal isn't to save hours, it's to make them better.
03
The AI proposes, flags where it isn't sure, asks for confirmation on conflicts. The decision stays with our team. Nothing ships without being truly understood: every step is inspectable, editable, rejectable.
A precondition that's on us
The three principles above only work if we also shift two underlying habits in the way we work. Simple habits, but they need to be kept alive every day.
Habit 1
Before every task we ask ourselves: can AI handle at least 30% of this?
The question isn't "do I use AI or not?", it's "how far does it go?"
Habit 2
AI always answers. It's on us to decide whether it makes sense. We discuss, we don't accept passively.
Never accept an output without asking why
The other half of the system comes from us
The cost of the transition
Adopting integrated AI produces an initial drop in productivity of around 20%, before the curve climbs back up. Almost everyone quits during the dip. Those who stay recover 30-50% of their time, and keep it.
Redesigning flows takes time. That initial period where "we're slower than before" is the investment. Without the patience to cross it, the gains that come after never arrive. That's what separates those who truly learn to leverage AI from those who stay on the surface.
How principles become architecture
From the three principles comes a concrete architecture: three levels we build one on top of the other, always in this order. Context first, then connections, then capabilities. Skipping one always produces the same result: agents that stall, vague answers, context lost.
01
what the AI knows about us
02
which tools it's wired into
03
what it can produce
Architecture · three layers of an integrated AI
Under the hood · technical choices
Underneath the system there are two architectural decisions that change a lot. The first is about how the AI's brain is built. The second is about the horsepower we use for each job.
Choice 1 · how the brain is built
The trend is to build teams of AIs that chat with each other. We tried it. It doesn't work, because each one sees only a slice of the problem and errors multiply down the chain. We do the opposite: one AI that changes role between phases, with real tools that execute.
Common approach · to avoid
BAZ26 approach · in use
Choice 2 · the right model for each job
Inside the same AI run models of different power. For simple, repetitive jobs we pick the fast and cheap models. For strategic decisions, the more powerful ones. Same system, costs under control, quality where it counts.
Fast
Sorting email, classifying content, summarizing, handling repetitive tasks. Around 3x cheaper, with 90% of the capability.
Balanced
Research, writing, reasoning, tool use. The model that does the bulk of the work across all our systems.
Deep
Campaign strategy, creative direction, architectural decisions. Slower and more expensive, used where the quality of thought matters more than the cost.
Beyond Claude, every process of ours also integrates specialized AIs for specific tasks: image generation, video analysis, data collection. They're called in when needed, inside the same flow and with the same context. Claude orchestrates, the other AIs run their part.
In production · AIgency Reborn · Social Analyzer · Smart database Cost percentages change · the principle doesn't
· Part IV ·
We don't sell an off-the-shelf product. We're sharing how we tackled three problems from our own work: three different answers, one philosophy. Each case attacks a different problem, but they all use the same method.
Case · 01
The creative agency that respects the brand. Starts from zero on new clients, works against the knowledge base, has binding controls.
The case · AI creativity
Case · 02
From a URL to a high-value strategic report, with metrics, SWOT and recommendations the client keeps in hand.
The case · competitive analysis
Case · 03
Our internal archive. People, projects, calls and emails connected in a graph you query in plain language.
The case · company archive
Different surfaces · same architecture
Case · 01 · brand-aware creativity
AIgency Reborn
We start from zero on a new client and walk all the way to the final pitch, always holding the correct tone of voice and brand identity.
7 step pipeline · 3 models · 18 KB sources
Gate · 5/5 green
One AI orchestrates: absorbs the brand, writes, generates images, keeps everything aligned Brand always binding · traceable pipeline · every output cites its sources
What happens inside
From a three-line brief to a publishable post. The AI changes role along the pipeline, every step cites the sources it used, every output passes a chain of automatic checks before delivery. If a check fails, the AI flags the conflict and asks the team.
Project Father's Day Gift · Concept 02 of 03
Iron in ten minutes, no iron, no board. The gift that changes the morning of someone you love.
Model per phase · Haiku for debrief · Sonnet for research, writing and review · Opus for strategy and creative direction When there's a conflict, the AI declares it instead of improvising
What actually changes
In a cold-start pitch for a new brand, AI handles the execution side: the debrief, the preliminary research, the first draft of the concepts. What's left is time for the work that really matters: the strategy, the conversation with the client, the critical review.
"Fragmented AI" model
1-2 days
Coordinating separate tools for a cold-start pitch: NotebookLM for debrief, Perplexity for research, ChatGPT for strategy, Nano Banana for concepts, Canva for the deck. Each tool has its own chat, its own password, its own context to rebuild every time.
Fragmented use · many separate tools
AIgency Reborn
Optimized approach
A single interface, one AI that handles debrief, research, first draft, citing sources at every step. The brand always stays binding. The remaining time goes to strategy, client, critical review.
Integrated use · validated on Bellairon
Numbers are indicative · they depend on brief complexity It changes how we work, not just how long it takes
Case · 02 · competitive analysis
Social Analyzer
We pick an Instagram, LinkedIn, Facebook or TikTok profile. We add the brand's site and its competitors. The system produces a reliable report with metrics, SWOT, brand audit and recommendations.
Profile URL to analyze
Brand site · optional
Competitor handles · optional
Fully local · data stays on the client's machine · no cloud Format · interactive HTML + printable PDF
What happens behind the scenes
The pipeline works in stages: each stage saves its result before moving on. If the connection drops mid-analysis, it resumes at the right stage, not from scratch. Designed to run on unstable connections and to respect privacy: everything stays local, nothing leaves the machine.
Stage 01
apify · firecrawl
Target profile, last 50-100 posts, brand site, competitor profiles. All collected and saved as raw JSON.
Stage 02
gemini · multimodal
Engagement metrics, content mix, recurring themes, top-performing formats, descriptive audience, site brand audit.
Stage 03
gemini · reasoning
SWOT, competitive edges, brand gaps, operational recommendations ranked by impact and feasibility. Every claim backed by evidence.
Stage 04
jinja2 · playwright
Structured HTML report + printable PDF. Every section cites the posts and data that produced it. Zero pages without a source.
Staged pipeline · any stage can be re-run on its own Data stays local · never leaves the client's machine
What the client receives
The report isn't a list of metrics. It's structured like a strategic document: executive summary, performance benchmarks, content audit, audience description, recommendations ranked by priority. Every number has a link to the post that produced it, so no claim sits disconnected from the source that backs it.
Page 04 · Performance
Post distribution by format, last 90 days. Reels dominate engagement, carousel beats single photo.
Page 12 · Strategy
Synthesis grounded on the data from pages 04-09. Every cell is linked to the evidence.
Page 24 · Action
Complete sections · executive summary, benchmarks, content audit, audience, recommendations, brand audit Precise citations · posts, screenshots, data
Case · 03 · company archive
Smart database
People, projects, calls, emails, documents: everything in one space. Connected. Queryable in plain language. The AI stops treating our world like a stranger.
Fully local · markdown organized into a graph The AI reads the archive like a library · never like a wall of text
One call · one routing
A call recording lasts an hour. Its transcript is a block of raw text. A BAZ26 routine routes it in seconds: identifies the people mentioned, the projects, the companies; links everything to the right file in the archive; adds precise links and citations. What was raw material becomes live context for every future conversation with the AI.
Before · raw markdown
00:03:17 Client We'd like to review the Q3 launch plan together. We're thinking about a new product for the premium segment. 00:04:13 Client On the ongoing project we've already aligned on the brand book. For the roadmap we're waiting on the partner's reply by Monday.
After · routed markdown
--- frontmatter --- participants: [[client]], [[daniele-faccio]] mentions: [[partner]] related: [[q3-launch-plan]], [[brand-book]] tags: [meeting, area/strategy, kickoff] --- 00:03:17 · [[client]] review of [[q3-launch-plan]], new premium product. 00:04:13 · [[client]] [[brand-book]] · aligned. roadmap · waiting on [[partner]] by Monday.
Auto-routing · every new call processed with no manual intervention One graph · many agents reading it
An AI without memory
is temporary consulting.
With memory,
it's your team.
Method summary · Smart database · BAZ26
How to get in touch with us
We don't sell licenses, we don't sell a platform, we don't have a price list. We work in three different ways, depending on the starting point. We decide the shape of the engagement after we've listened.
A
We listen. We map the AI already in the company. We identify the gaps.
For those · who want to understand before moving
B
We design the solution around your problem. No off-the-shelf product.
For those · who already know the what
C
We work alongside your team all the way to autonomy. The goal is to stop being needed.
For those · who want to keep the capability in-house
Often combined · A opens, B builds, C consolidates
From here on
We propose a sixty-minute call. We listen to where you stand with AI today. We tell you what's realistic in the next 90 days. From there we decide if it makes sense to continue. If it doesn't, we'll say so.
write to us · info@baz26.it