The most honest systems documentation in your firm is the spreadsheet nobody sanctioned. Your AI roadmap is already written in your workarounds, if you know how to read them.
Ask a controller how the monthly numbers actually come together. Somewhere in the answer lives a spreadsheet nobody sanctioned, maintained by one person, named something like FINAL_v7_USE_THIS_ONE.xlsx.
Nobody put that file in an architecture diagram. Yet it is the most honest piece of systems documentation your firm owns, because every workaround is testimony. A system refused a requirement, and a person routed around it. The export that runs every Monday, the numbers re-keyed between two platforms, the colleague who "just fixes" the categories before the report goes out: each one marks the exact spot where your official systems stopped meeting the business.
Which is why the standard way firms build an AI roadmap, a whiteboard, a brainstorm, and a list of use cases nobody has validated, starts in the wrong room. The demand signal already exists. It is sitting in the file share, holding the month-end close together.
1. Run a workaround census, not a use-case workshop
Workarounds beat brainstormed use cases on every dimension that matters. They are pre-validated: someone cared enough to build the ugly version and has maintained it, unpaid and unthanked, for years. They come with a named owner, a real frequency, and a measurable cost. A use-case workshop produces wishes. A census produces evidence.
Where to look:
Recurring exports. Anything pulled from a system into Excel on a schedule means the system's own reporting cannot answer the real question.
Re-keying. A human typing values from one screen into another is a missing integration with a salary.
The human API. "Send it to Maria, she fixes it." Maria is running an unencoded business rule from memory. When Maria takes a vacation, so does your data quality.
Version-numbered attachments. Files with dates and initials in the name are an approval workflow that never got built.
The shadow master. The shared spreadsheet everyone treats as the real customer list, the real rate table, the real headcount. That one is the most dangerous, and the most informative.
2. Translate each workaround into its requirement
The census only pays off if you read each artifact for what it encodes, then pick the fix that matches the requirement, and only that.
| The workaround | The requirement it encodes | The likely right fix |
| Monthly export to Excel for reporting | The system cannot answer the question people actually ask | A reporting layer. No AI required. |
| Re-keying between two systems | A missing integration | Plain automation |
| Someone hand-classifies transactions or tickets | A judgment rule nobody ever encoded | A real AI candidate, with human review |
| Shadow spreadsheet as the "real" master list | Nobody owns that data domain | Ownership and master data governance, before anything else |
Notice how the table splits. Some of it needs AI. More of it needs plain automation or a decision about ownership. Knowing the difference is the discipline that keeps budgets intact and reputations un-singed, because the fastest way to discredit an AI program is to point it at a problem an integration would have solved for a tenth of the cost.
3. The governance stakes just went up
Shadow spreadsheets were tolerable when only humans read them. They are ungoverned data stores, with no owner, no validation, and no access control, and they already feed real decisions. Now firms are wiring AI assistants to the file share and grounding chatbots on "our documents."
Ground an assistant on a drive full of workarounds and the workarounds get a louder voice and better grammar. The stale rate table stops being one analyst's private shortcut and becomes the confident answer the whole firm receives. Retrieval does not launder data quality. It amplifies whatever it finds.
4. The 90-day version
Weeks 1 to 4: inventory by amnesty. Ask people what they maintain to get their job done, and make it explicitly safe to answer. A census framed as an audit finds nothing, because workarounds hide from auditors.
Weeks 5 to 8: rank. Score each item on frequency, blast radius, and how much undocumented judgment it contains. The shadow masters and the human APIs float to the top.
Weeks 9 to 12: retire, promote, or rebuild. Retire what a report or an integration kills outright. Promote the judgment-heavy survivors into governed requirements with named owners. Pick your AI pilots only from that promoted list, each with an owner, a review step, and a finish line.
Ninety days later you have an AI roadmap made of proven demand, priced by real frequency, staffed by people who already care, and a data estate that is one census cleaner than it was.
Final thought
Firms keep asking where to start with AI, then commission research to find out. The answer has been on the shared drive the whole time, hiding in a filename with a version number. Read your workarounds. They are the requirements document your users have been writing for years, one unsanctioned spreadsheet at a time.
AI, Properly is the Ayvede field report: AI news read plainly, with a takeaway you can use. Ayvede turns ungoverned workarounds into governed AI roadmaps for high-trust firms. Advise. Vet. Deliver. Ask a question in writing.