Generative AI has been used in Agilebase for some time, with various features to help developers of systems, like

• generating SQL calculations, or debugging problems with existing calculations
• generating help text or documentation for tables, fields and views
• filling newly created tables with sample data to allow quick testing
• suggesting and creating tables and fields – i.e. given a descriptive prompt, build a basic application

We’re now adding the ability for customers to use AI directly and have it operate on the data in their systems.

We are opening a two month limited beta testing programme for individual customers who can use up to 2000 calls per month at no cost. If you are interested in joining, please contact us.

AI Workflows

The new feature we’re developing is called AI workflows.

A standard workflow in Agilebase typically uses calculations to work out what the content of fields should be, then creates new records with those field values, or updates existing records.

That provides great flexibility, the only limit is what calculations you can do in SQL.

AI workflows work in a similar way, but rather than the calculation generating the value to save in the database, it will generate a prompt, to be sent to an external AI. The AI’s reply will be what gets saved to the database. Note that you can configure things so the AI either replies with plain text, or with a data format suitable for further processing, such as a comma separated list, or JSON data. The AI can return one ‘value’ or a list of items.

The AI can also use the internet.

Some examples might help you picture the possibilities:

• Given a company number, populate our contacts database with a list of primary contacts
• We have a list of customers with whom we haven’t interacted with in 6 months. For each, generate a polite message of enquiry asking how they are doing with using the particular products they’ve purchased and whether they would like any support or training (taking into account the content of previous communications)
• The Amazon page for our product is X. Read all the reviews and return a brief summary of the best and the worst customer experiences

Taking the first example, the prompt you create could be

‘The Companies House page for {company name} is {url}. Please return a list in JSON format of all people, including their roles and dates of birth’

The response might be something like

[{

    "name": "Clifford John Calcutt",

    "role": "Director",

    "date_of_birth": "November 1957"

  },

  {

    "name": "Mark Thomas Garrett",

    "role": "Director",

    "date_of_birth": "April 1969"

  },

  etc.

which Agilebase would use to create one record for each contact.

Generic vs. customised AI

Out of the box, Agilebase will use a generic AI such as OpenAI’s ChatGPT, which has been trained on a massive corpus of public data. However, it is relatively easy to add your own context. For example, you can train an AI with details of your own products and services. You can even allow it access via APIs, to public or private databases of information, so it can look up specific things.

To try that out quickly and simply (without coding), take a look at Bristol company Helix.ai (https://tryhelix.ai). Agilebase can connect to a custom Helix model as easily as a generic one.

Data Privacy and AI

If you stick to the defaults, data will be sent to a public AI such as OpenAI, via their API. You are then subject to their privacy policy and data security.

If you set up your own custom AI, you can optionally host it yourselves, or have it hosted for you, in which case you can completely control the data and it needn’t leak beyond the boundaries you set. For example, use open source models or Azure OpenAI with a private endpoint.

We can put you in touch with one of our consultancy and delivery partners to help you explore that option.

Here in Bristol, technology providers such as Helix (as above) or Mystic AI (https://www.mystic.ai) can form part of the solution. 

Pricing

AI workflows are free to use up to 100 calls per month per client, i.e. 100 record updates. Beyond that prices start at 8 pence per call, reducing down 3p for higher volumes. Usage is capped to avoid incurring unexpected costs. Please request our standard pricing spreadsheet for full details.

Note: final pricing is subject to change following cost monitoring during the beta test.

Timing

This isn’t ready for release yet, but if you’d like to take part in the beta programme, please get in touch and we’ll contact you soon.