EiGENRA Back-Office on Three AI Agents

Departmental ops automation, demonstrated on our own books. Three AI agents draft the proposals, the Monday business pulse and the nightly lead follow-ups; code computes every figure, and nothing reaches a client without the founder's tap. The same playbook we run for SME services firms.

ⓘProject description and metrics shown here are representative — drawn from a real engagement, with the client name anonymized where confidentiality applies.

Most AI consultancies sell transformations they wouldn't bet their own business on. We flipped that: could EiGENRA's own back office — proposals, the weekly view of what needs the founder, follow-ups with warm leads — run on agents we'd be comfortable shipping to a client? Proposals waited on the founder's calendar, unpaid invoices and waiting leads sat across several screens, and warm leads went quiet once the desk's one automatic nudge had passed.

Three AI agents on shared Upstash Redis state, and code computes every fact they use. A proposal drafter (Sonnet, on demand) turns a captured lead into a scope — their problem, what we build, what is not included, the steps — while prices come from our price book at the client's regional rate, and quoted work is left at "[your fixed price]" for the founder to fill in. A Monday business pulse gathers unpaid and overdue invoices, unsigned contracts, pending approvals and the week's warm leads; when nothing needs him it sends "All clear" without calling a model, otherwise Haiku writes three to five lines of what needs him, in order, above the facts. A nightly drafter (Sonnet through the Batch API, at half price) writes each warm lead's next message in their own language. A draft carrying any figure its brief does not hold is replaced by a template, nothing reaches a client without the founder's tap, and one hard cap of $5 a month is checked before every model call.

// Every office agent: code builds the brief, the ceiling answers first,
// the model writes prose only, and a figure it invents sends the template.
const params = officeRequest({
  model: "claude-sonnet-5-5",          // proposals + follow-ups; Haiku 5.5 for the pulse
  system: PROPOSAL_SYSTEM,             // stable → cache_control: ephemeral
  user: proposalBrief(lead, offer),    // facts from code; no price in it
  maxTokens: 3000,
  effort: "medium",
  cacheSystem: true,
});

const verdict = await reserveSpend(worstCaseUsd(params)); // $5/month hard cap, fails closed
if (!verdict.ok) return templateDraft(lead, offer);        // no call at all

const reply = await anthropic.messages.create(params);
await settleSpend(verdict.reservation, actualCost(reply));

const sections = parseSections(replyText(reply));
if (!sections || unsupportedFigures(sections, brief).length) return templateDraft(lead, offer);

// Prices come from the price book at the client's rate — or "[your fixed price]".
return saveDraft({ sections, investment: investmentFor(offer, lang, priceBook, region) });

Every draft lands in one admin queue: approve and send, edit, or dismiss. A follow-up goes out through WhatsApp only inside the customer's 24-hour window; outside it, the founder copies it and sends it from his phone. The cap keeps model spend under $40 a month by construction. Every pattern here is reusable in your business — that's the offer.