The owner of a twelve-person consultancy noticed it on a Friday afternoon. Her account director had emailed a client the v3 of a proposal that morning. When the client came back with a follow-up question, the AI summary tool the firm uses pulled its answer from the v4 sitting in a different folder, written two days later and never sent. The two replies contradicted each other on price. Nobody on the team had done anything wrong. The firm had four versions of the same proposal in three different places, and the AI tool had taken the most recent one it could find.
She is not unusual. For many owner-managed firms, version chaos has been the price of doing business since the day someone first typed Proposal_v4_dh_FINAL_final2.docx and emailed it round. The cost was tolerable when only humans were reading the documents, because humans ask clarifying questions and notice when things contradict. AI tools do not. The discipline that solves it is three rules, applied with consistency, and almost none of the work needs new software.
Why does version control matter more once AI is reading your documents?
Humans reading documents ask clarifying questions, notice contradictions and verify intent. Generative AI does none of those things. When two versions of the same document sit in the retrieval pipeline, the model conflates them and returns something plausible without asking which one is current. The version with the older price wins as often as the version with the new one, and nobody on the team is in the loop.
Research by Dr Philippa Hardman found that around 80% of leading models misread documents regularly, because the model compresses everything into one numerical representation and updates it as it reads more. The Air Canada case put a number on the cost. The airline’s chatbot quoted a bereavement discount policy that had been superseded, and a tribunal ordered Air Canada to compensate the customer. The error sat in the document pipeline that fed the model an outdated policy alongside the current one. For a small firm, the equivalent is the AI tool reading the wrong proposal, the wrong terms, the wrong procedure. The cost is paid in client confidence.
What are the three rules that do most of the work?
Three rules carry the load. One canonical version per artefact, kept in one place everyone knows about. Consistent simple naming, incrementing numbers, no status labels in the filename. Superseded versions moved into an Archive subfolder once a new version is approved. The discipline is small enough to teach a new hire in an hour and large enough to prevent the version conflict that has AI tools quoting the wrong price.
Rule one, one canonical version per artefact
Every document the business cares about should have one version the team considers current. That version lives in one folder, on one system, with a name everyone can recognise. Drafts are either deleted once the current version is approved or moved to archive. For documents created collaboratively, the canonical home is the shared folder in Google Workspace, Microsoft 365 or Notion. For documents created by one person and approved by others, it is a designated Approved or Current folder. Everything else moves out.
Rule two, consistent and simple naming
Descriptive name, hyphen, incrementing version number. Client-onboarding-01.docx, then 02, then 03. Minor correction inside a version, add a decimal: 03_01. If chronological order matters, put the ISO date in front: 2026-05-11-client-onboarding-03. No FINAL. No DRAFT. No parenthetical v2 (1). The reason is mechanical: filenames travel. A file called Proposal_FINAL.docx tells the recipient it is final even if it has been wrong for six weeks. A file called Proposal_04.docx tells them nothing except its position in the sequence, which is the honest signal. Status belongs in a version table at the front of the document or in the system metadata, never in the filename. This matches the US National Archives guidance and Princeton’s records-management standard.
Rule three, superseded versions move to archive
Once a new version is approved, the previous version moves to a subfolder explicitly named Archive or Superseded. This does three things. It removes the older version from everyday search results so the team is less likely to retrieve the wrong one. It creates an audit trail if the document ever needs to be reconstructed for a dispute. And it clarifies for everyone, including the AI tool, that the version in the active folder is the current one. For critical documents, contracts, signed deliverables, approved procedures, rename the archived copy with the date it was superseded: client-onboarding-02-archived-2026-05-11.pdf. The Archive folder holds the history. The active folder holds the truth.
Where do built-in tools do the heavy lifting?
Modern collaborative platforms handle a useful slice of the discipline automatically. Google Docs keeps a full edit history with named contributors and the ability to restore any earlier state. Microsoft 365 SharePoint lets administrators set retention limits, automated archival and audit logs. Notion records change history and supports version templates. For a team inside a single collaborative platform, the mechanical version-tracking is handled. The three rules cover the boundaries the platforms cannot reach.
Documents leave collaborative tools the moment they are exported to PDF, sent by email, downloaded for offline editing or printed for a meeting. At every boundary the version metadata vanishes. The proposal exported as a PDF for a client carries no record of which Google Docs version it came from. The contract amended in Word from a PDF the client returned with tracked changes has lost its connection to the cloud original. For SMEs running a hybrid stack, accounting in one system, CRM in another, knowledge in shared drives, contracts in email, no single tool covers everything. The explicit naming and archiving rules earn their keep where the platforms cannot reach.
When do the proportionate rules run out?
The three rules work well for a stable firm with modest document volume, a few thousand active documents and a team of fifty or under. They start to creak when one of four things happens: document volume explodes, formal approval workflow arrives, regulation demands tamper-proof records, or AI moves into the core workflow and the cost of a version error rises sharply. When one of those signals lands, manual discipline is past its useful life.
Each signal has a recognisable shape. Volume becomes the problem when documents pile up faster than anyone can archive them. Approval workflow becomes the problem when drafts need team review, legal check and sign-off, with audit evidence at each step. Regulation becomes the problem when FCA, ICO or ISO rules require role-based access and retention enforced automatically. AI becomes the problem when the team is no longer occasionally summarising a document but routinely asking an assistant to read contracts, policies or procedures.
The step up is lightweight document management software, not enterprise content management. Tools like DocuWare, the SME-scale Hyland products and cloud-native DMS layers that sit on Microsoft 365 or Google Workspace cost roughly £100 to £300 a month for a firm of twenty to fifty people. They bake the three rules into the system: the canonical version is the only one users see by default, version numbering is enforced, and superseded versions archive automatically. Enterprise content management platforms cost ten times that and are built for organisations of several hundred people with multi-site governance. They are almost never the right next step.
What does the starter version look like for an owner reading this on a Friday afternoon?
Half a day of work, no software purchase. Pick the five or six document types that matter most to your business, proposals, contracts, standard operating procedures, client deliverables, internal policies. For each, name the canonical folder, the naming convention you want the team to follow, and where superseded versions go. Write the decisions down on one page, send it round, and make someone accountable for each document type.
Run the discipline through a fifteen-minute monthly tidy. Walk the active folders, move anything superseded into archive, confirm the canonical version is the one the AI tool would find when it goes looking. That is the proportionate version. It works, it costs nothing, and it holds until the business has genuinely outgrown it. If you want help working out which document types matter most and what the next step looks like once the rules run out, book a conversation.
