Teach Certant what your people know
Tell Certant once what a term or column means. It keeps the correction in a ledger and replays it whenever new documents arrive.
"In the salary schedules, the Year column means years of experience within the classification. It isn't a fiscal year."
rebind_propertyYear becomes years_of_experience. Evidence: values 1 to 8, none of them castable to a date.unpivotFour rate columns, one per effective date, become one row per date.extend_vocabularyAdds a type for a step within a classification.
Keep what your experts know
Every organisation has words and columns that mean something specific to it: a Year column that counts experience, a code only the payroll team can read, a policy that replaced another without anyone renaming the file.
Correct a general AI tool and the fix usually lives in one chat. In Certant it goes into the knowledgebase, where every chatbot, agent, dashboard and monitor picks it up.
After a year, the ledger is a record of what your people decided about your own records, and every one of those decisions still applies.
From page 146 of a published disability services enterprise agreement, captured 12 June 2026. The two figures are one 3% pay rise apart, and nothing else flagged the difference. Read how it was found.
How a correction works
You describe the problem in a sentence. Certant does the modelling, shows you the result, and changes nothing until you apply it.
- Describe it Open the dialog from the data model, any record, or under any answer, and say what's wrong in plain English.
- Check the proposal Certant turns the sentence into typed changes, each citing the statistics behind it. If the data disagrees with you, it says so.
- Preview the rows See the change list and real rows from the new version. Nothing is saved yet.
- Apply it Applying makes a new version and keeps the old one, so you can roll back.
From then on, code applies the correction on every rebuild. The model is told about it too, but the result doesn't depend on the model doing as it's told.
Find the gaps in your records
Certant also keeps a list of the questions it couldn't answer. Read it once a month and you'll see which documents to add, which policy has gone out of date, and what your staff or members keep asking.
Your team corrects what's wrong and adds what the list says is missing, so the knowledgebase gets more complete the longer you use it.
- Does the travel policy cover contractors?No document covers contractors
- What's the notice period for casual staff under the 2026 agreement?The 2026 agreement hasn't been added
- Is there a template for a hardship request?No matching document
Each one points at a document to write, update or add.
Chatbots and agents use it too
The same knowledgebase sits behind chatbots, the Agent Builder, Analytics, Strata and the AI FDE. When you correct the knowledgebase, each of them uses the correction.
Corrections are added and never edited. Undo is a new entry.
Each correction keeps the evidence behind it, including any disagreement with the data.
Each applied correction makes a new version of the model and keeps the previous one.
Changes that touch live data wait for a person to approve them.
Before your team starts correcting
What if someone makes a wrong correction?
Undo it. The undo goes into the ledger as a new entry, and you can also roll the model back to the version before the correction.
Does the AI have to follow the correction?
Code applies the correction on every rebuild, so the result doesn't depend on a model following an instruction. The model is told about the correction as well, so it builds around it.
What can we correct?
What a column or term means, which property a value belongs to, how a table should be laid out (for example, turning dated rate columns into rows), and new kinds of value your data needs.
Where do we make a correction?
From the Data Model page, from any record's page, or under any answer.
What if the data disagrees with the person correcting it?
Certant says so, and shows the evidence. The person can still go ahead, because they know the organisation, and the disagreement is kept with the correction.
Show it what your team knows
Book a demo on your own documents, and bring the one term everyone outside your team gets wrong.
Year in the salary schedules means years of experience within the classification.