SableflareDescribe a process

A software company, starting with operations

We engineer the repetitive work out of your operations.

Re-keying orders from emails. Checking invoices against purchase orders. Copying the same figures between spreadsheets, portals and CRMs. We study how that work actually happens, then build software that does it reliably — and brings in a person wherever a decision needs one.

A few sentences is enough. You don't need to know what the software should be.

Example process

Supplier invoices received by email

Invoice INV-20417Complete

  1. S1Receive — Software, AP inbox: 09:14:02

    Invoice arrives as a PDF attachment.

  2. M1Read — Model: 09:14:04

    Supplier, invoice number, 6 line items and total read.

  3. S2Match — Software, ERP: 09:14:05

    Matched to purchase order PO-7781.

  4. R1Check — Rules: 09:14:05

    3 of 4 checks pass. Total is 6.2% above the PO; tolerance is 2%.

  5. P1Review — Person, AP team: 09:41:37

    Reviewer approves: freight surcharge agreed with the supplier.

  6. S3Post — Software, Accounting: 09:41:38

    Posted for the next payment run.

  7. S4File — Software, Documents: 09:41:38

    Invoice, PO, reviewer note and trace filed together.

  8. EDone — End: 09:41:38

    Complete. 1 human decision, fully traced.

  • Software
  • Rule
  • Model
  • Person
  • Exception path

Illustrative. Every system we build keeps a trace like this.

The problem

Your systems don't talk to each other, so your people do.

An order arrives by email. Someone copies it into a spreadsheet. Someone else re-keys it into a supplier portal, checks it against the CRM, and pulls it into a report at the end of the week. Each step is small. Together they take hours, rely on rules nobody wrote down, and stall when the one person who knows them is away.

That isn't a people problem. It's a design problem: the work between your systems was never designed, so people have been filling the gap.

One order, five systems

  1. Inbox
  2. ExtractCopies the orderClarifies unclear orders
  3. Spreadsheet
  4. SubmitRe-keys itApproves exceptions
  5. Supplier portal
  6. CompareChecks itResolves mismatches
  7. CRM
  8. CompileBuilds the reportActs on the report
  9. Weekly report
Today: four hand-offs, each carried by a person. It works as long as everyone is available, careful, and remembers the rules.

What we automate

Repetitive work has a recognizable shape.

If one of these sounds like a morning in your team, it's likely a good candidate. The marks under each one show what a system for it is usually built from.

  • Software
  • Rules
  • Model
  • People
  • Inbox-driven work

    Requests, orders and documents arrive by email, and someone reads, sorts and re-keys each one.

    • Customer purchase orders typed into the ERP from PDF attachments
    • A shared inbox where every message is triaged and forwarded by hand

    Usually built from software, a model, rules, people.

  • Document processing

    Invoices, contracts, statements and forms, read by a person to pull out the same fields every time.

    • Totals and line items taken from supplier invoices
    • Signed contracts checked for missing pages, dates or signatures

    Usually built from a model, rules, people.

  • Cross-system workflows

    One piece of work entered, in slightly different forms, into several tools that never talk to each other.

    • A new order recorded in the CRM, the ERP and a carrier portal
    • Status changes copied from one system to another at the end of each day

    Usually built from software, rules.

  • Reconciliation

    Two sources that should agree, compared line by line, with the differences chased down.

    • Payment provider payouts matched against bank statements
    • Carrier invoices checked against shipment records

    Usually built from software, rules, people.

  • Recurring verification

    The same checks, run every day, week or month, and recorded the same way.

    • Supplier registration numbers checked against a public register
    • Certificates and licenses tracked before they expire

    Usually built from software, rules, people.

  • Research and data gathering

    Information collected from websites, registries and portals, then written up in a standard format.

    • Company details assembled for each new lead or supplier
    • Prices and availability checked across several supplier portals

    Usually built from software, a model, rules.

  • Onboarding

    A new customer, supplier or employee, set up in every system they touch.

    • Documents collected and verified, then accounts created in four tools
    • Missing information chased until each file is complete

    Usually built from software, a model, rules, people.

  • Reporting

    The same report, rebuilt every period from several exports.

    • Morning reports downloaded, combined and summarized for the team
    • Month-end figures pulled from three systems into one pack

    Usually built from software, rules.

  • Compliance operations

    Checks that have to be performed, evidenced and recorded identically every time.

    • Names screened against sanctions lists, with each result recorded
    • Evidence gathered into folders ahead of an audit

    Usually built from software, rules, a model, people.

  • Internal administration

    Requests and approvals that follow the same path, every time, through the same people.

    • Access requests routed, approved and provisioned
    • Contract renewals flagged, approved and filed

    Usually built from software, rules, people.

Worked example

Customer orders by email, before and after.

A common process in distribution and manufacturing. The details differ from company to company; the shape rarely does.

Illustrative example, not a client engagement

As it runs today

8 steps · 5 systems · 1 coordinator connecting them

  1. Done by hand: A customer emails a purchase order: a PDF, a spreadsheet, or lines typed into the message.

    Inbox
  2. Done by hand: A coordinator finds the customer in the ERP and opens a new sales order.

    ERP
  3. Done by hand: Each line is typed in by hand: item codes, quantities, requested dates.

    ERP
  4. Done by hand: Prices are checked against the customer's price list, kept in a spreadsheet.

    Spreadsheet
  5. Done by hand: Stock is checked in the warehouse system.

    Warehouse
  6. Done by hand: If anything doesn’t match, the coordinator emails the customer and makes a note to follow up.

    Inbox
  7. Done by hand: A confirmation is written and sent.

    Inbox
  8. Done by hand: The original PO is saved to a shared drive, when someone remembers.

    Drive

Engineered

7 steps · same 5 systems · 1 human checkpoint, for exceptions

  1. Every message to the orders inbox becomes a case with its own ID.

    Inbox
  2. The purchase order is read in whatever format it arrives, and each line extracted.

  3. Stock and credit status are looked up for every line.

    ERP · Warehouse
  4. Each line is checked: known item, this customer’s price, sensible quantity, deliverable date.

    Price list
  5. A person: Only orders that fail a check reach the coordinator — with the original PO, the extracted lines and the reason. They correct, approve, or call the customer.

  6. The sales order is created and confirmed to the customer from an approved template.

    ERP · Inbox
  7. PO, order and every decision are filed together.

    Drive

What changes

Hand-offs
The coordinator stops re-typing orders and spends that time on the ones that need attention, and on customers.
Exceptions
Every mismatch lands in one queue with its reason. None of them live in someone’s inbox.
Human review
Required wherever a price, a date or a promise to a customer would change.
Traceability
Any order can be traced back to the email it came from, and every check it passed.
Reliability
When a document can’t be read confidently or a system is down, the order waits for a person. It never guesses.

How we work

We start with the work, not the technology.

Processes rarely break at the obvious steps. They break at the edges: the exception nobody mentioned, the rule that lives in one person’s head. So before we design anything, we watch the work being done.

  1. Understand

    OutputA map of the process — systems, rules, decisions, exceptions — and a straight answer on what is worth automating.

    1. Observe

      Sit with the people who do the work and follow real cases from start to finish, including the awkward ones.

    2. Map

      Write down every system, rule, decision and exception. There are almost always more exceptions than anyone remembers.

    3. Separate

      Sort each step: fully deterministic, safe to automate with checks, or genuinely needing a person. Some steps shouldn’t be automated at all, and we’ll say so.

  2. Build

    OutputA working system, integrated with your tools and tested against your own past cases.

    1. Design

      Decide what the software does, where people decide, and exactly what happens when something fails.

    2. Build

      Integrate with the systems you already use: APIs where they exist, careful browser automation where they don’t.

    3. Prove

      Run it on real past cases, then alongside your team, and compare the results before it handles live work on its own.

  3. Run

    OutputMonitoring, alerts, and a named owner for keeping it working.

    1. Operate

      Watch every run, alert on failures, and adapt the system as your process, suppliers and tools change. Systems need owners after launch.

Engineering

Each step gets the simplest technology that does it reliably.

We’re software engineers first. A process is a chain of different kinds of work, so we choose the tool step by step — not once for the whole system, and never for the pitch.

  • Software

    When the steps are known

    Plain code, APIs and integrations. Fast, cheap to run and easy to test. Most of a well-built system is this.

  • Rules

    When a policy decides

    Tolerances, thresholds, eligibility. Written as explicit rules your team can read, test and change — not buried in code or in someone’s memory.

  • Models

    When the input is unstructured

    Free-text emails, scans, inconsistent PDFs. Language models and OCR handle what rules can’t, and their output is checked before anything acts on it.

  • People

    When judgment or accountability matters

    The system prepares the case; a person decides. We design these checkpoints on purpose, not as a fallback for when the software is unsure.

Models are a component, not the premise.

We expect language models to be part of many systems we build; they are remarkably good at reading messy input. They’re used where they are the best tool for a step, and their output is checked like any other input. Where a rule or a query does the job, we use the rule or the query.

How our systems behave

Every run leaves a trace
What arrived, what was read, which rules ran, who decided what, and when.
Failures are loud
When input is unclear or a system is unavailable, work stops and the right person is told. Nothing guesses silently.
Built on what you have
Your inbox, spreadsheets, ERP and portals. You don’t replace systems to start.
Rules in one place
Business rules live somewhere your team can read and review them.
Documented for handover
Written down, so the process stops depending on one person — including us.

Start here

Tell us about the work someone does by hand.

You don’t need a specification, a budget or a list of systems. Describe the work the way you’d explain it to a new colleague. Something like this is enough to start:

“Every morning someone downloads three reports, compares them, and enters the differences into a supplier portal.”

What happens, how often, and which tools are involved. Rough notes are fine.

How often does it happen? Optional

Leave out passwords and customer data — a description of the work is enough.

Prefer email? Write to hello@sableflare.com.