There's a hard truth buried in most stalled AI initiatives: the workflow was never ready. The process lived in three inboxes and one veteran's memory. The data was scattered across spreadsheets with column names only their author loved. No model fixes that — a model amplifies it.
You cannot automate a process no one can describe.
01The pre-AI checklist
- ▢The workflow has a system of record — one place where its state lives, even if that place is simple.
- ▢The handoffs are explicit: who does what, when, and what "done" means at each step.
- ▢The data path exists: inputs arrive digitally, or the first project is making that happen.
- ▢The exceptions are known: the veteran's edge cases are written down before they retire into legend.
02Modernization is the wedge
Here's the reframe: modernization isn't the boring prerequisite to the AI project. Done right, it is the first AI project. Wiring a workflow into a system of record, structuring its data, making its state visible — that build creates the substrate every intelligent feature needs, and it pays for itself before a single model is invoked.
Companies that skip this step buy AI twice: once when the tool fails on top of chaos, and again when they rebuild the foundation and try again.
03What "enough" modernization looks like
The trap on the other side is over-modernizing: eighteen months of platform work before anything intelligent ships. The standard isn't perfection — it's sufficiency for the one workflow you're about to automate. Modernize the lane, not the highway system.
- ▸One workflow, end to end. Its inputs arrive somewhere queryable, its state is visible, its outputs land somewhere structured. Everything else in the company can stay messy for now.
- ▸Interfaces over migrations. You rarely need to replace the legacy system — you need an API, an export, or a sync job that lets new software read and write alongside it. A read-only mirror of the data is often enough to start.
- ▸Weeks, not quarters. If the modernization slice for a single workflow is scoped past six weeks, the slice is too wide. Cut scope, not corners: fewer fields, fewer integrations, one direction of sync.
04The ROI nobody counts
Modernization gets undersold because its returns land under other line items. Count them anyway:
- ▸Visible state. When the workflow has a system of record, status meetings shrink and "where is this?" emails disappear.
- ▸Faster onboarding. New hires learn a documented process in days instead of shadowing a veteran for months.
- ▸Cheaper everything after. Every future feature — AI or not — builds on structured data instead of excavating it first.
- ▸Optionality. A modernized workflow can adopt whatever the next wave of tooling is. A chaotic one can't adopt anything.
05A representative pattern: the quoting lane
Consider the pattern we see most often — details vary, the shape doesn't. A services company quotes custom work. Requests arrive by email, a senior estimator builds each quote in a spreadsheet copied from the last similar job, pricing rules live in their head, and finished quotes go out as PDFs with no structured record of what was quoted or won.
The naive AI project starts at the glamorous end: train something to draft quotes. It fails immediately on substrate — there is no structured history of quotes to draft from, no record of win rates to calibrate against, and no system the draft could land in.
The sequenced version runs the other way. Phase one gives quoting a system of record: requests land in a queue, each quote gets structured line items, sent quotes and outcomes get logged. No AI yet — and already the pipeline is visible, handoffs stop dying in inboxes, and the estimator stops rebuilding spreadsheets. Phase two adds the intelligent layer on top of state that now exists: draft quotes assembled from similar past jobs, priced from the now-recorded rules, routed through the estimator for judgment. The drafting assistant that was impossible in month one is a natural extension by month three — because the modernization was the first project, not the prerequisite it kept getting filed under.
06Sequencing the two projects
In practice the modernization build and the AI build aren't separate initiatives — they're phases of one scope. Phase one: give the workflow a system of record and a data path (this alone usually removes a chunk of the manual work). Phase two: add the intelligent layer — extraction, drafting, triage, retrieval — on top of state that now exists.
Sequenced this way, the project shows value at the end of phase one instead of holding its breath until the model works. And if priorities shift after phase one, you've still bought a permanently better workflow instead of a stalled science experiment.
OPERATOR NOTE — When a team asks for "AI for our operations" and can't name the system of record, the deployment path always starts one layer down.
Put this thinking to work.
A 30-minute strategy call with an operator — we'll map your first deployment path, not send a deck.