How to build an AI agent that actually runs inside your dashboard.

Give it a self, and make it one of "us"

The single highest-leverage line in the whole system is the identity statement. Our agents don't say "the data shows" — they say "we". They're cast as a specific executive with a specific temperament: the CEO is calm, assured and direct; the CFO is a cash-obsessed realist; the CMO has taste and cultural fluency. Naming the role and the temperament up front colours every sentence that follows.

// One builder. Every prompt in the system flows through it,
// so the voice is identical in a briefing, an email or a chat.
function buildSystemPrompt(mode, staffMap) {
  const base = `You are the AI CEO of Hoovu, a B2B puja-flower
    supply chain in 9 cities. Voice: calm, assured,
    data-driven, direct. Use "we" — you are part of
    this team. Open with what is on track. Attach a
    concrete action to every gap. End with direction,
    not alarm.`;
  return base + toneFor(mode) + rulesBlock() + staffBlock(staffMap);
}

Define tone as switchable modes, not one voice

The same fact needs to be said differently to different people. To a founder, "cash is tight this week" is useful candour. To the warehouse team, it's needless anxiety. So tone is a mode the builder switches on — same brain, different register. We ban specific words in team mode and require honesty in founder mode.

Rank what it cares about — explicitly

An agent with no priorities gives you a flat list of everything. An agent with a ranked value system gives you a point of view. The CMO, for instance, is told its priorities in order — and when two pull against each other, it knows which wins. The CEO leads with what's on track; the CFO thinks in cash before anything else. Rank order is how you get judgment instead of a data dump.

Bake in the rules so it never sounds naïve

The fastest way to lose trust is for the agent to flag something that isn't actually a problem. So we encode the business's lived rules directly into the persona: revenue is final by 6 AM, so never say "tracking well"; fill rate reads 0% before 2 PM because invoices aren't entered yet, so don't panic; partial invoices land the next day, so don't flag a 48-hour gap. These rules are applied silently and never narrated.

// Apply silently — never name these rules to the reader.
- Revenue is FINAL by 6 AM IST. Never say "at this pace".
- Fill rate = 0% before 2 PM is normal. Flag only after 4 PM.
- Partial invoices fill NEXT day. Don't flag a <48h gap.
- Procurement 36–45% is fine; flag only if > 45%.
- Benchmarks: fill ≥97% · labour ≤8% · AOV ≥₹35.

Turn raw rows into narrated knowledge

Raw data answers "what happened". Memory answers "what does it mean". We walk the whole operation and write a memory layer organised into eleven categories — clients, cities, flowers, cashflow, festivals, labour, wastage, farmers, pricing, tasks and the company's North Stars. Crucially, each category gets both structured metrics and a short narrative in plain language — because the narrative is what the agent reasons on most fluently.

AIMemory/
  clients/narrative: "Zepto is our largest account but
     receivables have crept to 40 days. Swiggy steady…"
  clients/zepto: { daysSinceLastOrder: 0,
     revenueTrendPct: -8, fillRate: 0.97 }
  flowers/narrative: "Rose running hot pre-Navaratri…"
  cashflow/summary: { peakGapDay: "Thu", gap: 420000 }
  cities, festivals, labour, wastage, farmers,
  pricing, tasks, northstars // 11 categories total

Compress aggressively; the context window is precious

Once the memory exists, every surface reads it — not the database. The morning briefing doesn't re-aggregate a month of orders; it reads the pre-written P&L narrative and the day's numbers. This keeps responses fast, cheap and — most importantly — consistent: every surface tells the same story because they all read the same memory.

• One expensive computation at night → many cheap reads all day
• The same memory feeds the briefing card, the emails, and the chat console
• Compression is a feature: a tight narrative beats a giant JSON blob in a prompt

Re-seed on a schedule, while everyone sleeps

The memory is rebuilt every night at 11:30 so the morning briefing reads a fresh picture. Scheduled jobs do the heavy lifting off-peak: the nightly memory build, the morning briefing, the end-of-day summary, personal daily plans, and a weekly recompile of platform best-practices for the CMO. Each job is idempotent and logs its own failures, so a bad night never corrupts the memory.

23:30  → rebuild AIMemory (all 11 categories)
06:00  → morning briefing  (reads last night's memory)
10:30  → personal daily plans  (per manager)
18:30  → end-of-day summary
Mon 04:00 → recompile platform skills for the CMO
// every 30 min → anomaly scan → proactive nudges

Give it tools — with strict contracts

The nightly memory can't hold everything, so the chat console gets tools: a set of functions it can call to query live data mid-conversation. The model plans, calls a tool, reads the result, and either answers or calls another — a tool-use loop. The discipline is in the contract: every tool caps its date range, caps its result size, and returns a predictable { summary, data, notes } shape so the model never drowns in rows.

getCityPnL({ city, fromDate, toDate }) {
  // 1. cap the range — never let it ask for 2 years
  range = clampDays(fromDate, toDate, 90);
  // 2. compute, then cap the payload (~30KB)
  return {
    summary: "Bangalore: ₹2.1L revenue, 41% procurement…",
    data:    rows.slice(0, 200),
    notes:   "Fill rate excludes today (invoices pending)."
  };
}

Cache like the calls are expensive — they are

AI calls are slow and cost real money, so nothing recomputes that doesn't have to. The briefing card uses a three-tier cache: it checks the browser first, then a shared server-side cache (so the first teammate to log in pays for everyone), then the API only if both miss. Everything is timestamped — "generated 2h ago" — with a manual refresh button for when you truly want a fresh take.

• Tier 1 — localStorage: instant, this user, this device
• Tier 2 — shared server cache: one generation per day, reused by the whole team
• Tier 3 — the API: only when both miss, or on explicit refresh

Choose the surface to match the moment

The same agent shows up in four different ways, each suited to a different need. Get this mapping right and the agent meets people where they already are — instead of asking them to come find it.