Spark Tech Advisory
Case studies

Two engagements, described by the problem rather than the client.

Client names are withheld by default. What matters is the shape of the problem, the decisions we made, and what the team was left holding.

Manufacturing ~40 staff Multi-state sales Discovery → build

A finance portal that replaced a spreadsheet from 2017

Five manual processes — payroll, vendor invoices, commission holds, exemption certificates, sales-tax thresholds — all running through one co-owner, consolidated into one application his team operates.

Before
  • Payroll calculated by hand every Thursday in a spreadsheet built in 2017
  • Payments entered one at a time in a bank portal, notifications written by hand
  • Manufacturer invoices tracked as saved emails in an Outlook folder
  • Commission holds and clawbacks calculated manually, with no way to audit them
  • If the owner was unavailable on a Thursday, payroll did not happen
After
  • The Thursday payroll hour returned to the business every week
  • Two trained backup operators, so payroll no longer depends on one person
  • An auditable commission history for every rep, recalculated when profit changes
  • An alert when a vendor payment comes due before the customer has paid
  • Roughly 18 hours of December compliance work reduced to review and signing
The problem

Discovery started by watching the work, not asking about it. One co-owner spent an hour every Thursday, under constant interruption, computing commissions from CRM data in a spreadsheet nobody else fully understood. The logic was genuinely hard: individual rates, a payout cap, deposit-coverage holds, split deals, cancellation clawbacks that cascade into bonus clawbacks, and a delivery commission pool.

He said plainly that he could not be certain every weekly calculation had been correct. There was no systematic way to check.

Four adjacent problems shared the root cause: manufacturer invoices managed in an inbox, high-value open balances tracked by memory, exemption certificates re-typed by hand each December, and sales-tax thresholds reviewed once a year — which meant a single large sale could cross a threshold nine months before anyone noticed.

The decision

Three architectures were evaluated in writing before a line of code. Building inside the CRM meant fighting its automation limits with every commission rule, on a payment integration unproven at twenty-plus payees. A standalone app meant bypassing the accounting system already used for P&L and 1099s.

We chose the middle path: the CRM stays a read-only source of truth, a purpose-built portal holds calculation and approval, the accounting platform executes payment and keeps the ledger. Before committing we confirmed its API could create bills and trigger transfers at that volume, and documented a contingency path in case it couldn't.

The architecture document was signed before the build began. It named what we chose, what we rejected, and what we would do if the chosen path failed.

What we built

Payroll automation

Pulls the pay period's closed deals, applies every commission rule, and presents a report for review. Nothing moves until an owner approves it — then the portal creates the bills and triggers the transfers.

Vendor payment queue

Invoices arrive at a dedicated address and are parsed and matched to their deal. The queue is color-coded by urgency and flags any building where the vendor is owed before the customer has paid.

Holds and clawback ledger

A running record of every held amount by rep and deal. When a deal cancels, both the commission and the proportional bonus clawback are calculated and surfaced for review before any deduction.

Certificate preparation

An AI-assisted workflow that finds the current official form for each state and vendor, flags where it differs from the saved copy, pre-fills it, and produces a PDF ready to review and sign.

Threshold monitoring

Net sales and transaction counts tracked by state against each state's registration thresholds, with an alert at 70% and a status report each December. It surfaces the data; the registration decision stays with the business and its advisors.

Delivered with owner training for three operators, a payroll runbook for the backups, and an admin interface so rates, bonus schedules, and thresholds can be changed without a developer.

Handmade goods Solo studio Marketplace + own site AI writing system

One maker's voice, turned into a writing system

A one-person craft studio needed listings, captions, and emails written constantly — and could not afford for any of them to sound like they came from a machine.

The problem

The studio's whole advantage is that a real person makes each piece by hand, and the writing has to carry that. Generic AI drafts broke it immediately — hype adjectives, title case, invented delivery promises, emoji where the brand never uses them.

The owner was rewriting every draft from scratch, which cost more time than writing from a blank page.

What we did

We interviewed the owner, read everything she had written, then wrote the voice down as rules a system can follow: sentence case, US spelling, verb-first calls to action, concrete claims only, and a fixed list of usable proof points — no inventing a shorter turnaround to make a sale.

Each channel got its own format rules — listings, captions, email, blog, and bios differ in structure, length, and convention, including where a single emoji is welcome and where it never is.

The last piece is a brand check the system runs on its own output, fixing failures rather than shipping them and flagging judgment calls for the owner.

The outcome

The owner now asks for a caption or a listing and gets back copy in her own voice, formatted for where it is going, with a short audit of what was checked. The rules live in one place, so when the brand shifts, the system shifts with it.

The useful part was not the AI. It was making somebody write down what "sounds like us" actually means.

Start here

Recognize your own Thursday afternoon in this?

Tell us what the manual process is and who it depends on. Thirty minutes is usually enough to know whether there's something here.

Choose a day to see open times.