When the AI and the Manual Disagree: Franchise Manual Version Control

Franchise manual version control is the discipline of knowing not only what a policy says today but what it said on any earlier date, who approved the change, when it took effect, and who was told. Without it, a system grounded on your documents answers yesterday's truth with today's confidence.

Plenty has been written, some of it by me, about getting a manual into one clean current state. This is the part after that, which almost nobody does: keeping the states you left behind.

Two questions, and most networks can only answer one

Ask an operations leader what the refund policy is and you get an answer. Ask what it was on 10 March and watch the room change.

Those questions need different machinery. The first needs a current document. The second needs a record of every state that document has been in, with dates attached — which a shared drive cannot supply, because it holds the newest file and destroys the rest.

Four ordinary situations ask the second question:

• A location audited against a standard disputes which version was in force that quarter.
• A franchisee says nobody told them. You are fairly sure somebody did.
• Two operators in one market run a procedure differently, both correctly following a bulletin.
• A buyer's diligence asks how the brand communicates a change, and wants an example.

Geography multiplies it. Half of US franchise systems operate inside ten states or fewer, 34% run regionally across eleven to thirty-four states, and 16% are national at thirty-five or more, on FRANdata's footprint data. Once a policy carries a state-specific effective date, one manual holds several truths at once — and the network is the only party experiencing all of them.

Three dates, and most manuals record none of them

Every policy has at least three dates. Approved, when leadership signed off, useful to the author and nobody else. Published, the first moment the network could have known. Effective, when behaviour is supposed to change.

Brands collapse all three into whichever day the file was saved, and the collapse stays invisible until something starts reading the document literally.

Take an ordinary case. A packaging change is approved on 1 June, sent to the network on 8 June, and takes effect on 1 July. A franchisee asks on 15 June what the packaging standard is.

Three answers are defensible and one is right: the current standard, the fact that it changes on 1 July, and the section saying so. A system answering with the old standard alone is accurate and useless. One answering with the new standard is wrong for a fortnight, and that argument happens at a supplier's counter.

Neither is a model failure. Both are what happens when a document does not know what time it is, and retrieval quality cannot fix a corpus with no time dimension in it.

Franchise manual version control: what a change record has to carry

No document management platform is required. Six fields are, plus the discipline to fill them every time a standard moves.

1. What changed, in the words that changed. Not "Section 7 updated". The old sentence and the new one, next to each other. Operators act on a difference they can see; nobody has ever acted on a log entry that says "clarified".
2. Who approved it. A person, not a department. Somebody will need to ask them what the intent was.
3. When it was published and when it takes effect. Two fields, always, even when they are the same day.
4. What it supersedes. The version it replaces, marked retired and kept. Deleting the old text destroys the only evidence that the people following it were once right.
5. Who was told, and through which channel. A bulletin sent to a distribution list eighteen months stale is not notice, and you learn that mid-argument.
6. What came back. Which locations opened it, which acknowledged it, which never did.

Five of the six cost nothing but habit. The sixth needs a system, because acknowledgement is the one fact a document cannot record about itself.

One running cost is worth conceding. Every amendment means the material is read again, and re-reading a corpus is document-heavy work at the expensive end of what AI costs per location. Amending monthly on a fixed date beats amending continuously, for reasons unrelated to editing.

When the assistant and the document disagree, triage in this order

An operator forwards a screenshot. The answer does not match what you believe the policy is. Work through four causes, likeliest first.

A retired copy is still in the corpus. Commonest by a distance. Nothing in a lookup knows which of two approved files is the later one, so both come back and the answer is a confident average of them.

Nothing says which document wins. Manual, agreement, state addendum, last month's bulletin. If no page settles precedence the system picks, and it picks differently on different phrasings.

The effective date was never modelled. The answer is right for a date that is not today.

The document is right and the person is not. The usual outcome once the first three are fixed, and the one a version record exists for. Show the operator what changed, when it took effect and the date it reached them, and the conversation ends in a way "check the portal" never has.

Only the fourth is a people problem. The rest are defects with addresses, and the first three are what grounding forces a brand to settle in writing.

The retrospective test, which takes an hour

Pick a policy that changed in the last year. Without asking anyone: what would a franchisee in Ohio have been told about it the month before the change, and what evidence exists that they were told?

Time yourself. Under an hour and you have version control. A day and you have a shared drive with good intentions. If the answer is "ask the coach who covers that market", you have located your record — in a person.

That is a thin place to keep it. One franchise business consultant averaged 34 units in 2020, more than 21% wider than the year before — an expansion FranConnect's operations index attributes in part to the pandemic. Memory across that many buildings amounts to one phone, and the phone leaves eventually.

Two caveats. Keeping every version forever has consequences of its own, and the retention window is a question for counsel rather than a vendor's default setting. And a brand demonstrating a live assistant during governed AI franchise development has quietly promised to keep the material behind it true — a version-control commitment wearing a sales badge.

An answer given in a group chat used to be deniable and unfindable, and networks absorbed the inconsistency as weather. A cited answer is neither. The brand now has an author, a timestamp and a position — worth having, and only if you can still say what that position was last spring.

Audit trails, and the vocabulary they belong to: audit trails, defined with four others.