Prompt chaining
Break a task into small model calls that pass results to each other: summarise the account, then classify the risk, then draft the message.
Platform · Logic and workflows
Chain AI steps, decisions and hand offs into one flow you can read and change.
Logic and workflows is the engine behind every app and agent: the order of steps, the decisions between them and who gets the work next. You build it from reusable blocks, keep model costs in check and see exactly why each case went where it did.
AksharAIFlow: collections remindersAK
Flow: collections reminders
- Trigger
- Daily at 8:30 am, accounts due within 3 days or overdue
- Steps
- Read history, check promise to pay, classify, draft message, route
- Languages
- Gujarati, Hindi, English
- Token limit
- 4,000 per account, 2 retries
- Always to a person
- Over 60 days past due, disputed accounts
- Last edited
- Rina Shah, 22 Sep 2026
- Flow drafted
- Rules reviewed by collections head
- Rerun on August accounts
- Compliance sign off on message wording
- Publish to live queue
The problem
One prompt cannot run a real process
An NBFC's collections team handles thousands of overdue accounts each week. Each one needs several steps: read the account history, check the last promise to pay, choose the right reminder in the right language, and decide whether a call from an agent is needed. A single chatbot prompt gets some of this right some of the time. The team needs each step written down, in order, with clear rules for when a person takes over.
What it gives you
Inside logic and workflows
Calling agents within a flow
Hand part of the work to an agent built in Agent studio, such as a document reader or a reminder agent, and carry its answer back into the flow.
Token management
Set limits per step and per run, trim long inputs before they reach the model, and see the cost of each flow so bills do not surprise you.
Reusable logic blocks
Save a tested sequence, such as "verify PAN and match name", once and use it in every flow that needs it. Fix it in one place and every flow picks up the fix.
Decision trees
Write branching rules on amount, days past due, language or customer segment, and see the whole tree on one screen.
Task routing
Send each case to the right queue or person by region, skill or workload, with the reason attached so nobody has to guess why it arrived.
How it works
The order you work in
- Step 1Map the processList the steps your team follows today, where they decide something and where they hand work over.
- Step 2Build the chainTurn each step into a block: a model call, a rule, a lookup or a call to an agent.
- Step 3Add decisions and routesDraw the branches and say who gets each outcome, including the cases that must always go to a person.
- Step 4Set limitsCap tokens, retries and run time for each step so a bad input cannot run up cost or loop forever.
- Step 5Run and watchEvery run records which path it took and why. Change a rule, rerun past cases and compare before you publish.
AksharAIFlow runs, week of 5 Oct 2026AK
Flow runs, week of 5 Oct 2026
Accounts processed
3,412
Sent to officers
486
Avg tokens per run
2,180
Failed runs
11
- Mon82
- Tue74
- Wed69
- Thu77
- Fri91
An example
Daily reminders for an NBFC collections desk
Each morning the flow picks up accounts due in the next three days and those already overdue. A chain reads the repayment history and any promise to pay noted by the field team. The decision tree sorts accounts: a gentle WhatsApp reminder in Gujarati or Hindi for first time late payers, a call task for anyone who broke a promise, and straight to a collections officer for accounts past 60 days. The flow never sends a message that mentions penalties or legal action; those stay with the officer.
An illustrative example with sample data.
Questions
About logic and workflows
What happens when a step fails or the model gives a poor answer?
You decide per step. A failed step can retry, fall back to a simpler rule, or stop and send the case to a person with the error shown. Output checks can reject answers that miss required fields or break your wording rules.
Can we see why a case went down a particular path?
Yes. Every run keeps a trace of the inputs, each step's output and the rule that sent it onward. That trace is what a reviewer or auditor looks at when a decision is questioned.
Which AI models does a flow use?
The one chosen for your project. AksharAI works with leading model providers, for example OpenAI, Anthropic Claude, Google Gemini and open models, and different steps in one flow can use different models where that makes sense for cost or language.
Keep going
Related parts and reading
Platform
Agent studio
Configure an agent's knowledge, actions and limits, then test it like a new hire.
Learn morePlatform
App builder
Put together a working AI application on a canvas, then test it before anyone uses it.
Learn moreGuide
Writing a prompt flow your team can maintain
A prompt that one person wrote in a weekend becomes a problem the day that person goes on leave. Small steps, instructions in plain language and tests that run before each change keep the flow maintainable by the whole team.
Learn moreSee it on your work
Bring one slow, repetitive process. We will show you an agent doing it.
A working session with someone who builds these every week.
Or get a ballpark proposal in two minutesWhat happens next
- Step 1A short call about one process and the systems involved
- Step 2A demo of an agent working on a version of your task
- Step 3A clear view of fit, approvals and what it would take