Skills and knowledge in layers
An AI agent is as good as what it knows about your work. What are your tables called? Which definition of "active customer" applies? Which conventions do you have for views and procedures? In Plainwater AI you capture that once, and every agent of every colleague works with it.
Skills and knowledge documents
- A skill is a work instruction: how you approach something. For example "this is how we write a load procedure", "this is how you check a new source", "this is what a good dashboard looks like".
- A knowledge document is background: what something is. For example a data dictionary, an architecture description or the agreements with a customer.
You can write knowledge documents in the web app or import them from Word, Excel or PDF. The platform extracts the text.
The layers
Skills and knowledge stack in layers, from broad to specific:
- Platform — what applies to everyone.
- Customer profile — what applies to a kind of customer, for example "housing association" or "wholesaler".
- Customer — what applies to this one customer.
- Task — what applies only to this piece of work.
When a task starts, the platform adds everything up. If the same skill appears on two layers, the most specific layer wins. A customer can therefore override a platform skill with its own version, and a task can do the same again.
Recorded what went along
On every start the platform records which skills and documents went along: the name, the version and the origin. That way you can see afterwards why an agent did something a certain way.
Versions
Skills and knowledge documents have versions. A new version does not overwrite the old one; both are kept. A task uses the newest published version at the moment it starts.
An agent can propose a skill
If an agent discovers something during its work that will be useful next time, for example a convention that was not written down anywhere, it can propose a skill. A human reads the proposal and approves or rejects it. Only after approval does the skill go along with later tasks.
Names in knowledge documents
Knowledge documents also contain table and column names. Those go to the model as pseudonyms, as everywhere else. See Pseudonymisation.
Who may do what
| Layer | Assign |
|---|---|
| Platform and customer profile | catalogue management |
| Customer | customer management |
| Task | editing the task |
On every layer you also see what is inherited from higher layers.
Example
A team of four consultants works for three housing associations.
- The platform has a skill "this is how you check a new source".
- The customer profile "housing association" has a knowledge document with the common terms: rental unit, complex, tenancy change.
- One customer has its own data dictionary, and a skill that adds that customer's naming to the platform skill.
- A task has a note: "careful, the contracts table was rebuilt in March".
Every consultant who starts a task for that customer gets exactly the same stack.