From a presentation I gave to the Heraklion Makers Club yesterday…
To use Coda AI to automate BUSINESS PROCESSES, we need to be able to build AI AGENTS.
Each AI AGENT is an expert at performing a single task which involves a certain amount of LOGIC (branches and loops) as well as AI prompts to the Coda AI. So we use Coda Formula Language (CFL) to specify the logic and calculations, and use Coda AI table-columns to do the AI parts of the task. We use Coda Tables to keep track of the data needed for the process.
Here is an example of an AI AGENT that translates emails from Brokers that contain Instructions to be processed. It has a mixture of CFL Formulas and Coda AI prompts that complete the task.
-
We ensure the word-count is smaller than the ‘Coda AI Context Limit’ using a formula
-
We use a simple Coda AI prompt to determine the language of the email
-
We then translate the email to English using a slightly more complex Coda AI prompt
-
Now we must determine the QUALITY of the translation to protect against Hallucinations
As you can see, nothing here that is beyond the ordinary Coda Maker - its not rocket science.
The behavior of the AI AGENT is controlled with a single Start! button which contains all the LOGIC needed by the agent to complete the task in a reliable way.
The formula in this button can be summarized as follows
-
if word-count is too high, generate an ERROR message
-
do the AI prompt to detect the language of the email
-
if the language is English, skip the translation
-
if the language is UNKNOWN, generate an ERROR message
-
do the AI Translation prompt and the Translation-Quality prompt
-
if the quality-score is too low, generate an ERROR message
-
otherwise the result is a SUCCESS
The next AI AGENT in the process will then parse the translated email to detect all the Broker’s instructions, and send EACH of them to the appropriate AI AGENT for that transaction type.
So the entire Business Process is broken down into simple Tasks, with an AI AGENT built to do that task in a reliable way.
To “trigger” an AI AGENT we use the CFL formula below
-
create a new AI AGENT instance with the
addrow()& set the input values -
push the AI AGENT’s
Start!button to execute its procedure -
the
Start!button becomes disabled when the AI AGENT has defined its result
Again, none of this is beyond the abilities of most Coda Makers who are experienced with making Coda automations. Its not rocket science, does not need Python programming skills, and does not require a degree in computer science.
We originally tried to achieve this using very big complicated prompts. But LLMs are very bad at following complex logic and are useless at doing ANY calculations. Both of which are needed to automate even the simplest Business Process. Add to that the tendency for hallucinations to occur when the prompts and results are very long (small prompts and results almost never cause hallucinations).
So, this approach works extremely well, because the whole process is broken down into smaller tasks, and the use of CFL formulas for logic and calculations is more reliable.
And all this is possible BEFORE Coda launches it Coda-Brain technology later this year.
Coda-Brain will only enhance the power of this approach, when it arrives.
My goal was to illustrate to Coda Makers how they can exploit AI to vastly improve their Automated Business Processes without needing programming skills - just good-old Coda know-how.
We have encapsulated this technique in our AGENTIX FRAMEWORK pack which will be launched next week.
Respect,
Max