Launch HN: Manaflow (YC S24) – Automate repetitive office work in tables

[austinwang115]

Hey HN! We’re Austin, Lawrence, and Wesley from Manaflow (https://manaflow.ai). Manaflow is a spreadsheet where each column represents a step instruction of a task and each row represents an AI agent executing a case of the task. With one click, you can execute thousands of tasks involving data retrieval, gluing APIs, and taking actions.

Here’s a video demo: https://www.youtube.com/watch?v=jwaaqjHGkT4

And a live demo: https://manaflow.ai/demo

The idea to build an AI automation tool for small-to-mid-sized businesses (SMBs) emerged from our extensive conversations with local managers, directors, and operators. We learned that SMBs are constantly overwhelmed with manual workflows and heavily underutilize technology, especially compared to bigger businesses with teams of engineers.

Witnessing an operations manager show us folders of spreadsheets for his business, we realized that spreadsheets have limited functionality. While spreadsheets can handle data entry and track states manually, you can’t program spreadsheets to connect with other apps, call APIs, read PDFs, search the internet, or take actions.

At heart, operation managers are expert workflow programmers as they direct humans to do tasks. We think that tomorrow’s operation managers will program AI agents in English to perform these tasks instead. The ideal way to execute workflows is not manually operating software tools and spreadsheets, but by merely clicking a button that enables AI agents to handle them end-to-end.

Manaflow’s primary interface is a spreadsheet where each column represents a step in the workflow and each row corresponds to an AI agent executing a task. The workflow powering each spreadsheet is programmed using natural language, allowing non-technical users to describe tasks and steps in plain English, eliminating the need for coding skills. Each spreadsheet has an internal dependency graph to determine the execution order for each column. AI agents assigned to each row execute tasks in parallel, handling processes such as data transformation, API calls, content retrieval, and sending messages. You can watch us build a basic workflow here: https://www.youtube.com/watch?v=TGbJN7pNb30.

One of our customers uses us to take original videos from Google Drive, watermark them with their logo, and then email the final products to their clients — all in one click of a button. This has cut their 20-hour manual workflow down to 20 minutes, and Manaflow has become a core part of their operations. Here’s a demo: https://www.youtube.com/watch?v=tTTTRrzICVg

Another use case is using Manaflow to enhance customer insight processes by tagging and classifying customer segments based on their interview transcripts. The automated tagging helps quickly sort through interview data and identify customer profiles for driving product development strategies. Here’s a demo: https://www.youtube.com/watch?v=GXEBJTh2i8Y

Manaflow has many built-in tools like oauth connections, RAG for PDFs, web crawling, Google search, and browsing LinkedIn that our customers can use. For example, researchers and consultants use Manaflow to automate the retrieval of key stats & data from fact books, while businesses use us to transform raw invoice documents into structured invoice data. Here’s a demo: https://www.youtube.com/watch?v=-hBcTjuqPFs

ASK: We’re experimenting with two ways to program workflows. 1) Use natural language instructions to build workflows where columns are sorted and executed in the order of a dependency graph, 2) Notion-inspired editor that lets you define Python tools and Manasheet columns for the AI agent to fill out. If you’ve read this far, please revisit https://manaflow.ai/demo?beta and let us know what you think! We'd also love to hear your insights and experiences in the workflow automation space!

[jakubmazanec]

I looked at the video demo and tried the live demo and I still have no idea what it does and what value it brings to the user (why adding a watermark to a video needs an AI?). Also, it's super slow and not responsive.

So many AI startups and I have yet to see one that makes sense for me. But I don't blindly trust any output from any LLM, so that's probably the reason.

[wtjangnaka → jakubmazanec]

Hi, thanks for trying out our demo! You're right - watermarking a video doesn't need AI, but we use AI to help enable nontechnical users to build that watermarking workflow, as well as any other workflow that they want to automate without having to interact directly with Python scripts.

[haswell → wtjangnaka]

Focus on things that need AI in these demos. Even a simplified non-technical UX to add a watermark does not need AI.

“As well as any other workflow they want to automate” is doing a lot of heavy lifting where concrete use cases should be highlighting why your tool does what it does the way it does.

[bpshaver]

If you're going to use a spreadsheet as your UI, I think you should be more careful how you describe the semantics of rows and columns. And make a strong case for why this representation makes sense.

People expect that rows represent observations and columns represent variables. Along those lines, would it not be more accurate to say each row represents an instance of a task and column represents a sub-task or step in that task?

"each row corresponds to an AI agent executing a task" just... doesn't make sense. The rows exist before you press the "execute" button, after all. The agent executing the (generic) task is something that happens on or with the sheet.

[gcanyon → bpshaver]

Seconding this -- I actually had copied "each column represents a step in the workflow and each row corresponds to an AI agent executing a task" to post a comment like this, but then checked through the comments to see if someone beat me to it, and here you are!

I'll add my terminology though, since it's slightly different. I think we're mostly agreed that "each column represents a step in the workflow" is fine. But for me the sheet represents an agent configuration, and each row is a specific job the agent performs.

[austinwang115 → bpshaver]

We're using a spreadsheet UI because it's familiar to millions of office workers and easy to understand the step-by-step progression of a workflow. The wording that you proposed makes a lot of sense to us.

[threeseed]

I still don't understand how multi-step LLM based AI agents work.

If the probability of an LLM making a mistake = 5% and you have 10 steps then the accuracy of the overall workflow is 60%. Which is useless. Even if we have major advancements in the performance of LLMs and it drops to 1% then still the overall workflow is 90% which is poor.

So what is the plan here ? There is a limit to how many tasks in businesses can tolerate so much inaccuracy.

[sdwr → threeseed]

Let's say the task is automatically aggregating customer support info.

First step is collecting incoming emails

Second step is summarizing each one

Third step is batching by issue/severity

Notice how there is tolerance for deviance/error. "An error" looks like coding a ticket red instead of yellow, or slightly misrepresenting what a client said. The overall workflow can still be net positive.

[alexkwood]

Re: Witnessing an operations manager show us folders of spreadsheets for his business, we realized that spreadsheets have limited functionality.

did you bother finding out where those spreadsheets come from, internal systems, external reports from vendors, are they consolidated bank statements, inventory counts and status, amazon/shopify inventory status

your ai tool should enable them to work with these spreadsheets as a starter.

I totally agree with vector_spaces comment, having a AI agent create a new workflow and train business manager on using it is a dead end. They have had the last 30years to explore VBA, Access and the other tools Microsoft comes with, and they last thing they will do is understand python the way your demo shows

[bofadeez]

Looks like https://github.com/MoritzLaurer/GPT-google-sheets

[altdataseller → bofadeez]

Now this I would use…

[bofadeez → altdataseller]

Yeah 200 lines of JavaScript, available for the last year. Super cutting edge

[altdataseller]

Does this integrate with Google Sheets or Excel? If not, i would not use this. My workflow lives in spreadsheets but not your custom spreadsheets. They live in Google Sheets/Excel so you need to meet me where I do my work if you want to make my life easier

[austinwang115 → altdataseller]

Yes, we have integrations with Google Sheets, so you can write directly there.

[gcanyon]

Based on the watermarking video, it seems that your interface is actually sort of 3-dimensional? Meaning that a single column (watermarking in the video I watched) represents multiple underlying steps? I feel like to fully leverage the spreadsheet UI you need to explode that view out so the row more clearly represents each sub-step in a column?

Also: are there conditionals? So you can skip a step/column if not needed, or repeat as many times as needed?

[macqm]

I tried to launch the demo, it's stuck in "Loading Manaflow demo...".

Console has errors:

> failed to load resource: the server responded with a status of 422 () clerk.browser.js:2 Uncaught (in promise) Error > at s._fetch (clerk.browser.js:2:48584) > at async X._baseMutate (clerk.browser.js:2:49256) ingest/static/recorder.jsv=1.139.3:1

> Failed to load resource: the server responded with a status of 404

[cyph3rpvnk]

Great job on putting stuff out there. I understand the reasoning behind using spreadsheets, but wouldn't a flow chart/node based interface make more sense for this and be more intuitive? That way people can just upload their sheets/list to the beginning of the node and the subsequent actions/nodes would be the agents.

I have more comments/feedback if you're interested (mostly UX).

[imvetri]

Great work. I didn't not go into the details, and I like the concept and motivation.

My feedback: approach this problem, solve the problem, then use a technique. Show a repetitive work that someone does, and let the viewer watch machine doing it automatically.

In other words, your aim is right, but knowledge you possess is distracting you to solve it neat and clean.

[ilrwbwrkhv]

Another one of these companies which will sell to other YC companies and "exit" in a few years.

I wish you luck but honestly it seems like you have not done the ground work behind what operation managers actually do.

[vector_spaces]

Your assumption that operations managers will be swiftly replacing their human charges with AI agents indicates that you haven't spent enough time understanding the businesses of your customers. You really need to take a job in one of these roles for several months and spend more time talking to the people whose jobs you intend to automate away with this product, otherwise you're not going to reach people with this.

Tech people seem to have this cluster of assumptions that leads them to conclude that there aren't intelligent people in non-tech roles, and that these people can't see obvious optimizations of their roles because of their lack of coding skills or something.

The reality is usually that there are layers upon layers of hidden complexity in these businesses -- ones that require a mix of domain expertise and deep awareness of the business and human context to effectively manage. Often you won't even have so much as heuristics to go on.

That isn't to say that automation and AI can't be leveraged to great effect, but it's simply not going to be the drop-in solution you claim it is. Claiming it is in this way -- esp with your smug lip-service to job annihilation -- is going to rub people the wrong way.

Instead you should re-message this around augmentation and making jobs easier so that people can focus on other concerns, and reducing costly errors introduced by manual process.

It's unclear who exactly your target is, but if you are going for e.g. parts manufacturers, local shipping & logistics companies, small CPG brands, then the examples in your videos are all wrong. Get the weird Fibonacci stuff out of the side panel, clean out any junk that says "test" and use polished examples related to reconciling purchase orders, forecasting demand for a new product line, managing production schedules, etc. You need to make the value this thing adds accessible to intelligent people who don't have a CS degree.

[threeseed → vector_spaces]

> and that these people can't see obvious optimizations of their roles because of their lack of coding skills or something.

Actually the biggest mistake is thinking technology is an optimisation at all.

Many times it's easier, cheaper and faster to do it manually especially when the magnitude or complexity of the work is low. And once a task becomes repetitive the cognitive load on the worker will be orders of magnitude higher with some app or AI agent.

[vector_spaces → threeseed]

Definitely agreed -- I was being somewhat tongue in cheek there

The hard problems at most companies whose solutions could move the needle the farthest tend to be cultural. Sometimes we make gains in that direction as a side effect of introducing the right technology. Those cases are exceptional: technology is rarely sufficient, and as you say, often not needed in the first place.