A daily ops briefing and replacement-opportunity tracker for an HVAC and plumbing company

What the manager sees

How it works

The API calls

Every call carries an OAuth access token and the ST-App-Key header, and is scoped to the company's tenant ID.

Code step: the replacement score

A rule a manager can read and argue with. AI never decides this number.

// Score 0–100 per service location. Every point has a stated reason.
const RULES = [
  { id: 'age',       pts: e => e.ageYears >= 15 ? 35 : e.ageYears >= 12 ? 20 : 0,
    why: e => `equipment age ${e.ageYears} yrs` },
  { id: 'repairs',   pts: e => e.repairs12mo >= 3 ? 25 : e.repairs12mo === 2 ? 15 : 0,
    why: e => `${e.repairs12mo} repair visits in 12 months` },
  { id: 'r22',       pts: e => e.refrigerant === 'R-22' ? 20 : 0,
    why: () => 'uses R-22 refrigerant' },
  { id: 'costRatio', pts: e => e.repairQuote / e.replaceEstimate >= 0.4 ? 20 : 0,
    why: e => `repair quote ${Math.round(100 * e.repairQuote / e.replaceEstimate)}% of replacement` },
];

function scoreLocation(equipment, callSignal) {
  let score = 0; const reasons = [];
  for (const r of RULES) {
    const p = r.pts(equipment);
    if (p > 0) { score += p; reasons.push(r.why(equipment)); }
  }
  // A verified "strong" interest from a call adds weight, but only with evidence.
  if (callSignal?.replacement_interest === 'strong' && callSignal.evidence_verified) {
    score += 15; reasons.push('customer raised replacement on a call');
  }
  return { score: Math.min(score, 100), reasons };
}

AI step: prompt and output contract

The model gets one transcript at a time and may only answer in this shape. Numbers, prices and dates are not its job.

You review one phone call transcript for a residential HVAC and plumbing company.
Return JSON only, matching the schema. No prose.

Rules:
- Quote the customer's exact words in evidence_quote. Copy them character for character.
- If the transcript does not clearly support a field, use "unknown" or null.
  Never guess. An empty answer is better than an invented one.
- Do not state prices, dates, equipment ages or counts. Code handles those.
- replacement_interest is "strong" only if the customer raises replacing equipment
  themselves. A technician suggesting it is "possible".

{
  "call_id": "string",
  "intent": "booking | complaint | estimate_question | billing | other",
  "sentiment": "positive | neutral | unhappy",
  "replacement_interest": "none | possible | strong | unknown",
  "follow_up_needed": true,
  "follow_up_reason": "string or null",
  "evidence_quote": "exact words from the transcript, or null"
}

Validator: no quote, no claim

const ALLOWED = {
  intent: ['booking', 'complaint', 'estimate_question', 'billing', 'other'],
  sentiment: ['positive', 'neutral', 'unhappy'],
  replacement_interest: ['none', 'possible', 'strong', 'unknown'],
};
const norm = s => String(s || '').toLowerCase().replace(/[“”"]/g, '"').replace(/[‘’]/g, "'").replace(/\s+/g, ' ').trim();

function validate(raw, transcript) {
  let out;
  try { out = JSON.parse(String(raw).replace(/^```(json)?|```$/g, '')); }
  catch (e) { return { ok: false, error: 'not valid JSON' }; }

  for (const [field, values] of Object.entries(ALLOWED)) {
    if (!values.includes(out[field])) return { ok: false, error: `${field} must be one of ${values.join(', ')}` };
  }
  // The quote must really be in the transcript. Otherwise the claim is downgraded.
  out.evidence_verified = !!out.evidence_quote && norm(transcript).includes(norm(out.evidence_quote));
  if (!out.evidence_verified && out.replacement_interest === 'strong') out.replacement_interest = 'possible';
  return { ok: true, signal: out };
}
// On ok:false the Zap re-asks the model once with the exact error, then routes to the error path.

Try the validator

Pick a sample call and run it. The AI output is pre-written for the demo, but the check runs live in your browser using the same rules as the code above.

Reliability built in

Sample tracker rows