📢 TLDR

Series of small and big use cases that you can implement internally across operations of your company as we did. There is an illustrative demo, screenshots of actual prompts and outputs, and explanation of why to go multi-agent followed by the vision for building an autonomous company. If you want to build something useful in this area together or have feedback to share, would love to get in touch.

You can give a try to the tool we built for non-techcnial users to create multi-agent automations https://wrkdn.com/signup

Thanks to all companies and individuals who contributed their cases!

👾 Synthetic company operations data

Most of the use cases below started with internal PoCs and to make it simple we built a script that generates imaginary 4yo B2B SaaS company with 50 employees dataset – HR records, org chart, payroll, open roles, resumes received for open roles, CRM with opportunities assigned to specific AEs from HR list, emails with clients and prospects, Hubspot with leads, Facebook/Google ads attribution, some creatives, help desk with tickets, knowledge base, sales plans, sales handbook, etc. This script makes multiple OpenAI calls to generate such data united with a common thread – employees, clients, revenue. Please feel free to use it to populate your own sandbox, if any, with data.

https://github.com/LondonWizard/Synthetic-Data-Generator

🙈 How it Started

In 2022, we connected GPT to our knowledge base to enable basic self-service on the website. The next avenue was connecting it to Salesforce but despite all fine tuning we were not able to achieve meaningful consistent results even on simple cases and stopped this project. For context, we are B2B SaaS and at that time our headcount was around 200 ppl. We used AI features built in tools we already had and sometimes migrated for better AI features (for instance, in outbound) but haven’t moved on with any major in-house development of tools or buying a 3rd party solution for our internal use cases.


💥 ChatDev

Like many of you, I was very excited about a year ago seeing a simulation of R&D process in a software company with ChatDev https://www.youtube.com/watch?v=eOIXVK-i3vk – a few specialized AI agents role-play to get from idea to the working product.


🔥 And then

In 2024 we tried, with fair share of successes, to use multi-agent automation inspired by ChatDev and (later) multi-agent frameworks, and I want to share cases that I think any (at least SaaS) company can do, either on their own or with frameworks/products. There are numerous small cases that I think are great to start with (and they will over time become features of products, for sure) but the north star is in autonomous tasks/functions/departments, so I want to show a few use cases on the journey there. In the end, I want to explain why we decided to develop our own multi-agent framework reusing some of LangChain's building blocks and to turn it into a tool for non-technical users.

Quick illustrative demo for those who haven’t seen multi-agent automation - (case #8 below)

https://www.loom.com/share/4ad3b173ce0a4b50abb5b7119a4e45a4


Now lets start with the simple cases


🚀 1. Two agents: Hiring – finding candidates for open roles

Input: ATS (or where you store CVs of applicants, could be a google drive folder) API or a pile of resumes, open role description or link to it over web, guide on who are good candidates for a role (in the screens below – guide on how to hire AI engineers with schools and previous employee list to consider).

Agents: One focused on data retrieval (I call it HR Ops) and one on matching candidates (‘Recruiter’).

Output: List of candidates that are the best to invite for an interview, supported by explanation.

Note: stepping into the territory of cases that involve candidates/employee data you’d better instruct/restrict system to use certain demographic data available to you as employer (let’s say age of support agents) in analysis

Let’s start exploring cases starting with single-agent automation:

🚀 2.Scoring/analyzing support tickets

Why: To better manage company and product. In early dates, we as founders spend a lot of time on customer tickets. Then, at a certain scale, it becomes harder to review each ticket, to think of why it was created in the first place and how it was processed, and what kind of customer experiences they create. I see a lot of solutions being built to talk to CRM like Salesforce, finding insights, making forecasts, etc. But much less focus on tools like Zendesk, while there is a lot for structured and unstructured data like ticket tags, emails and transcripts of calls. This data is gold!

Inputs – Support ticket API (Zendesk), knowledge base, HR records of support employees.

Output – scoring/ranking of tickets, insights on performance of support agents, extracting customer feedback and feature requests, finding gaps in knowledge base.

🚀 3.Turning product feedback into feature requests for a roadmap.

Why: To save Product team time on analysis of customer feedback, to process feedback at scale

Inputs – Knowledge Base, article from web with a framework on how to separate potential features into rocks and pebbles, integration/sample of data from the place where you accumulate feedback. Feedback can be sourced from a tool/website via API or scraping. One case shared by YC company was about taking feedback from a dedicated Slack Channel.

Output: List of most requested features sorted into small and big, with analysis of possible implementation effort.

🚀 4.Scoring leads/deals

This is the case that Salesforce promotes with Agentforce, very simple single API-call analysis that you can extend later to multi-agent case that dives into public and your own past history data to make a better scoring.

Why: To keep better focus on good prospects

Inputs – API access to CRM/Marketing tool, information on your ICP/Buyer Personas

Outputs – notes attached to specific deals in CRM with comments and suggestions, reports for sales management. Note: the quality of output can be significantly improved going multi-agent, but even one agent can do somehow well.


🚀 5.Reviewing quality of CRM data entry.

Why: To save time of sales managers/sales ops/execs/founders

Inputs – API access to CRM, AE/SDR handbook.

Outputs – notes attached to specific deals in CRM with comments and suggestions, report for sales management.

🚀 Now, multi-agent Automation:

Why to go multi-agent? – Multi-agent automation is much harder because it requires some kind of sequential/hierarchical orchestration of how inputs and outputs are flowing around the team of agents.

With a multi-agent approach, management of context sharing, memory and integrations/restrictions becomes what you as an engineer should think hard about.

But it can tackle much more complex/interesting cases than single-agent automation and enables experiences similar to ChatDev when you plug in agents into your company operations data.

Let’s continue on with use cases.

🚀 6.Three agents: Writing release notes

Inputs: API to where you manage product (tried Jira and Linear), knowledge base, writing style and feature template guidelines.

Agents: One that knows existing product documentation, one that does data retrievals and extracts key feature description, one that creates text copy following guidelines.

Outputs: Email/slack message sent to a manager with release notes (to be reviewed, approved, enriched with screenshots and published)

🚀 7.Four agents: Performance evaluation of sales/support employees

Inputs: API to CRM or Help Desk, HR records, people consultant guidelines, knowledge base, optionally - sales plans/targets for Support metrics.

Agents: One that excels at making API calls to CRM/Help Desk (take from Case #1 if you fine-tuned it to a good performance), sales/support process analyst, business metric-focused one and people consultant who digests it all together.

Output: Preliminary report on individual employee performance is sent over email to HR/Manager for review. All conclusions in the report are supported by links to deals/tickets for the convenience of review.


🚀 8. Five agents: Finding quarterly revenue opportunities/preventing churn.

Inputs: API access to Salesforce, Zendesk, Planhat, guidelines for individual agents on what are expectations for their outputs.

Agents: Three for data retrieval and posting to the systems above, business metrics analyst and knowledge base agent that also knows what is current product offering (for upsell purposes)

Output: List of closed-lost opportunities to reopen, accounts at churn risk based on support/product usage data (from Planhat but could be MixMax), accounts with highest upsell potential.

Ok let’s stop increasing complexity and focus again on potential quick wins:

🚀 9.Improving Knowledge Base

You may notice that in the cases above, Knowledge Base becomes a basis for agents to understand context, not only in support cases but everywhere it is good to help anyone understand the company offering/how it works. So let’s invest in making it better for humans and robots.

Inputs: Support and sales conversation history over a period of time (a month/quarter) Output: Suggested new content/edits for knowledge base .

Tip: If your product documentation is heavy with screenshots you may run an OCR agent to produce text descriptions of interface layouts to better feed agents.

🚀 10.Enriching new support tickets with recommendations /guidance for frontline employees

Unlike just talking to your knowledge base, agents here can look at the previous history of this specific user/account support like transcripts of calls to extract things like technical proficiency of a support requestor. Agents can even instruct your employees on how to be more empathetic in communications with their counterparty and produce templates for these communications. Could be quite useful in a B2C context for retention and cross-sell (we developed this specific case for DirecTV).


Another 10 doable use cases:

🚀 11. Scoring accounts in CRM/Success for upsell – taking in account as much data as possible

🚀 12. Building MRR forecast (more complex – cash forecast)

🚀 13. Matching invoices to bank transactions in foreign currency (AR automation)

🚀 14. Onboarding of new hires – agents check progress in tools and guide through the process

🚀 15. Creating campaigns in Hubspot based and finding most performing creatives/messages

🚀 16. Sales/Success/Support coaching – make orchestrating agent interface with your employees to produce guidance on specific deals and answer questions as chatbots

🚀 17. Turning webinar transcripts into marketing copy

🚀 18. Monitoring IT infrastructure and notifying the team when attention is needed

🚀 19. Monitoring support tickets for commonalty for automated escalation

🚀 20. Migrate data between systems when switching tools

👾 Feedback loop challenge Our major challenge with multi-agent automation was the effort it takes to create and (even more!) to debug a case. Looking into logs of information exchange between agents, it became clear that you need two people in the room – one who understands business process and output (e.g. how a good performance evaluation report looks like in Support) and technical (who can engineer a multi-agent system). On top of that, context, memory, security, integration management became big pain but these are cool challenges to address from an engineering standpoint. So we made a decision to use our own framework that can work with inputs of high-level abstraction coming from non-technical, business users, to create teams of agents that can work with each other. So making only 1 person at a room a time to create automations – either business users to formulate tasks/goals/processes with text, and technical (Ops or IT) to set up integrations and security.

👾 Why not a workflow automation The beauty of agentic automation is that you don’t need to invest time in detailed step-by-step instructions. High-level task formulation and non-linear execution is what makes it possible.

👾 Fighting hallucinations Spending mostly a year experimenting with Agents made me witness fun, weird and dangerous situations coming from hallucinations. Fund and weird – when suddenly someone called ‘Prince Mario’ becomes your top employee of the month. These are easy to spot, and given the presence of humans in the loop, they can be easily caught. Much more worrisome is when an agent who is a product analyst suggests a feature that never existed in customer feedback.

That was another argument for using our own framework where we create multiple copies for the same agent when in doubt and ask all copies to produce a plan/output and then vote on the best course of actions. Plus internal workflow has two layers of grounding LLM output in reality, for both individual and orchestrating agents.

🎯 Why implement multi-agent automation?

1. Give your non-technical teams the power of LLMs applied to their relevant enterprise data

2. Automate simple workflows without building the workflows or writing detailed instructions - get rid of at least 20% of routine operations in business functions.

3. Make data driven decision-making the motto of your teams
   
   

🎯 Want to give it a try with early access Wrkdn tool?

https://wrkdn.com/signup


📢 Want to learn more?

Book a demo with us https://calendly.com/d/cnpv-x29-nzc/wrkdn-demo