Most "AI marketing" tools optimise for output — more posts, more ads, more creative. Mira optimises for decisions. Her core object is not a post or a campaign. It is a single, repeating loop: Hypothesis → Experiment → Evidence → Decision → Learning. Everything below hangs off that. This page is the honest version — what runs today, what's shipping next, and exactly where the hard problems are.
A human CMO is not IF ctr>2% THEN scale — that's how you buy junk traffic. Mira runs a judgement loop, not a rule sheet.
The loop matters because the same number means different things in different contexts. A 0.7% click-through rate can be a winner if it brings serious buyers; a 3% rate can be a loser if it brings tyre-kickers. Mira reads the metric through the hypothesis, not in isolation.
No black box. A real CMO can tell you why. So can Mira, and you can read the receipt.
When Mira pauses an ad, shifts budget, or ships creative, she writes a decision record: the diagnosis, the confidence, the evidence and the counter-evidence, the rule that triggered it, and whether it needed your approval. This is what makes her auditable — and it's the raw material the teaching layer uses to explain herself to you in plain language.
| Field | Example |
|---|---|
| Diagnosis | High spend, low qualified conversion on Google Search |
| Confidence | 0.74 — medium-high |
| Evidence | CAC 2.3× target · 7-day activation 28% below benchmark · CRM lead quality low |
| Counter-evidence | Click-through rate above benchmark (so the creative isn't the problem) |
| Rule triggered | CAC > 2× target after minimum spend |
| Action | Pause Google campaign · reallocate to the LinkedIn test that's converting |
| Approval | Not required — pausing spend is inside her autonomy envelope |
A good CMO says "this is my best read, with medium confidence." Mira never reports certainty she doesn't have. shipping
This is the difference between a clever dashboard and a CMO you can trust with a budget.
Attribution is rarely a "fact" — it's a model with caveats. Mira knows that, and refuses to launder uncertainty into confidence. Before she acts on a number, it passes a data-quality layer:
If she can't trust the source of a signup, she will not scale spend on it. Full stop.
Too little traffic to be meaningful → the result is marked inconclusive, not "winner" or "loser."
She reports lead volume, and is explicit she can't yet vouch for lead quality.
Channel data disagrees with revenue data → she escalates an attribution mismatch instead of guessing.
"Inconclusive" is a first-class outcome here, alongside winner, loser, needs-revision, attention-without-intent, and wrong-audience-signal. Without it, an agent over-decides — it acts on noise. shipping (Phase 1)
With a human CMO you can't open their head and correct a wrong assumption. With Mira you can.
Everything Mira believes about your business lives in one place you can read — and partly edit. No guesswork about what she's operating on. It's two panes:
Positioning, banned words, approved claims, your ICP, pricing, constraints, hard goals. Fully editable, versioned, attributed. This is principal truth — what you say, goes.
Evidence-based rules, each carrying a confidence and a source ("Founder ICP responds worse to 'content generator' framing — confidence: medium, from Campaign X"). You can challenge or flag any of them (which forces a re-evaluation) — but you can't silently delete her evidence, because that's the memory she compounds.
Two rules keep this honest rather than dangerous:
Conflicts surface, they don't resolve silently. If you edit a belief that her data contradicts, she says so: "You've set us to target enterprise, but the last three campaigns show SMBs convert 4× better — keep your override, or update?"
Overrides are logged and tracked separately. When you overrule her, she records it, notes the risk, and tracks that outcome on its own — so over time she learns when your instinct beat her evidence, and when it didn't. Some founder instincts are right. Some aren't. She learns from both. shipping (Phase 4)
Autonomy without boundaries is either dangerous or useless. Mira's is bounded on purpose.
Every action sits in one of four tiers, and the boundary is coded, not vibes:
| Tier | What it means · examples |
|---|---|
| 🟢 Autonomous (within rules) | Draft & schedule approved content · generate ad variants · pause a broken campaign within thresholds · move test budget inside the envelope. Pausing money is conservative — it's allowed. |
| 🟡 Needs approval | Scaling spend above the ladder · launching a new campaign type · new positioning or offer · any public performance claim. Scaling money is aggressive — you sign off. |
| 🔵 Advisory only | Repositioning proposals · pricing ideas · partnership suggestions. She recommends; you decide. |
| ⛔ Refuses / escalates | Guaranteed-return or risk-free claims · scaling on broken attribution · spam · fabricated proof. And she escalates immediately on spend spikes, ad-account restrictions, or public backlash. |
Approvals reach you as a tap on Slack — yes / no / pick-one — batched into a daily queue, never a stream of pings. Decisions in buttons, deliberation in conversation. live
Self-grading is worthless. So a different model grades her — for you, not for us.
A separate, rival model reviews Mira's decisions adversarially and reports to you — its only job is to find where she's wrong. The philosophy is simple: the only party that matters is the tenant, and visible self-critique is how trust is earned. live
If you want to understand what's happening and why, there's a path. If you don't, she respects that.
Every decision ships with a plain-language "why" on tap — translated into your numbers, not textbook jargon. Not "CAC ₹3k > target, pausing per kill rule" but: "We're paying ₹3,000 to win one paying customer; we agreed ₹2,000 was the ceiling, so I've paused this before it burns more. Tap to see what CAC means and why we drew the line at ₹2k."
You set the depth — just tell me what you did / explain the reasoning / teach me the concept. Opt out and she stays terse, but the explanation is always one tap away, never gone. The teaching lives heaviest at the start, where a misunderstanding is most expensive: when you say "dominate the US market," she walks you through what that means — signups, revenue, or brand? — before she builds anything. shipping (Phase 5)
We build in the open. Here's the order, and why it's that order.
Real metrics as facts — CAC by channel, cohort trial-to-paid, funnel rates — plus the data-quality layer and bottleneck diagnosis.
The experiment object, the decision object, auto kill/scale, and learning that changes the next plan.
Decision & creative scoring with confidence, and the conversational goal interpreter.
The open, editable two-pane memory, with override logging and confidence decay.
The on-tap "why," the grounded glossary, tiered depth.
Weekly memo, monthly reset, and broader paid-ads execution as platform approvals land.
Phases 1–2 are most of the felt intelligence. Phases 4–5 are what let you trust and understand it — and they're the part no "AI that makes content" can copy.
Talk to Mira and ask her to walk you through how she'd market your company.