Everyone’s looking for the best ways to add AI to their company’s workflows.
ChatGPT to generate cold email templates. Midjourney to generate assets for a slide deck. Github Copilot for an AI pair-programmer.
All standard stuff in 2024. But most organizations are looking for something more.
Instead of automating generating a template, let’s do the whole flow. Let an AI agent copy the leads from Salesforce and generate the email templates, then queue up drafts in Gmail. Imagine how much time your salespeople could save.
A new class of tools called AI agents lets us do exactly that. Their potential is groundbreaking.
Problem is, they’re a bit of a gimmick in their current form. Thankfully, we have the technology to fix that.
First, let’s take a look at where we can improve.
The disconnect with most AI agents
You don’t actually need to automate everything
Demos float around of AI agents organizing an entire vacation based on a prompt.
They’re cool, sure.
Take a look under the hood, and this isn’t actually a great fit for an AI agent.
You’ll spend tons of time giving context and background, and there’s no guarantee the entire workflow (researching, vetting, booking each thing) will work. You’ll spend more time investigating, debugging, and retrying than you would have if you’d done it yourself.
All that for a task you might do once or twice a year. How valuable does that end up being for your team?
There’s a whole other category of tasks that are perfect for automation, but they don’t usually make for sexy demos.
The web is designed for humans, not robots
The perfect webpage for a robot probably looks like Salesforce.
One giant mishmash of buttons, forms, and information. A robot’s dream; predictable, with no hidden behaviors. Maybe Salesforce knew what was coming.
Of course, most web pages don’t look like this. Salesforce isn’t winning any (human) design awards. Modern web design is simple, because humans don’t like being overwhelmed. Gestures (like swiping) are implied and discovered instead of explained. We’ve spent 30 years building web pages around humans looking at them, not autonomous agents.
It’s not any better under the hood. The main components of any webpage (like the DOM and HTML) are brittle. They change frequently, and they’re complex even for AI.
Let’s imagine a website like LinkedIn. At its core, there’s a navigation bar at the top, a feed of content, a sidebar with connections…you get the idea.
Here’s what a vastly simplified version of LinkedIn looks like to an AI.
Trust us, this is as annoying to navigate for an AI agent as it is for you. We get all this muck abstracted away when we visit LinkedIn or Apple. AI doesn’t get that privilege!
In some places, the internet is actively designed against robots. How many times have you had to use a CAPTCHA when logging in, filling out a form, or doing anything useful?
Asking a generative AI to stumble around a webpage designed for humans is an inefficient (and unreliable) way to design automations.
Generative AI doesn’t translate well to automation (on its own)
LLMs are excellent at understanding natural language and intent. Throw them sixteen pages of research papers, and they’ll send back the key takeaways.
The problem with natural language is losing valuable context. Transferring knowledge takes time, and things get lost in that translation. Critical things that you know, and the AI has no way of knowing.
Even if you spend time to get it right, generative AI is not deterministic.
The same prompt in ChatGPT gives different results each time you run it.
That’s fine, because ChatGPT is mimicking human language. You want variance and that’s how humans use language.
If you’re building a workflow to do the same thing every day, this variance is a pain in the a**.
Even if you do account for variance, errors compound when you’re chaining events.
Build an automated single task (like filling out the first name in a form) with a 95% success rate, and that’s certifiably reliable.
But you’re never automating just one thing. It’s a step-by-step process - pull leads from HubSpot, dump them into Excel, sort by revenue, take top five, and generate cold emails.
Ten “95% success rate” tasks in a row is now a 40% chance of failure. Keep in mind, a good AI agent today succeeds 40% of the time on ONE task.
We’re trying to build a robust, deterministic outcome for automation. Not roll the dice.
How AI agents can transform organizations
AI agents aren’t perfect today. They’re a square peg for a round hole.
But they’re still high-leverage, groundbreaking technology that will change how we interact with computers.
We just need to narrow the focus and change the approach.
Spend time on repeatable, high-leverage tasks
Investment in AI automations should lead to time saved. Sure, there’s hours spent upfront, but it pays off quickly, not over a period of months or years.
Every job at a company has tasks like these. Ask your sales team what they hate doing every week, and you’ll probably get a good answer. Manually importing leads from Hubspot into Excel and then from Excel into Gmail is not what you want your salespeople spending their time on.
How many leads do you need to manually do this for, every week, to make this automation a good investment?
Probably fewer than you’d think.
These tasks are high-leverage and repeatable. They’re the best candidates for automation because they’re saving time AND automation is high-impact.
Automate to transform an organization, not just one person
One salesperson’s automations shouldn’t just help them do their job better. An automation that works for one will probably work for many. Distributing agent workflows changes their impact.
This is transforming within and outside of the organization. Within the team, everyone benefits from the best AI agent solutions, regardless of who creates them. The best employees at qualifying a lead can automate their expertise, share it, and everyone gets better at their job.
Outside, they can reduce friction. Communication between teams can be slow; frequent questions and asks can turn into automations.
A great salesperson can help customer service with an automation that brings up the right account info when a customer logs a ticket. Both people’s jobs get easier because of this saved time. This is context and expertise the salesperson can easily share without coaching anyone else.
Everyone works better, but just as importantly, you get data you didn’t have before. Look at where the automations are focusing on, and you learn where the time waste is.
Augment agents with APIs
The internet isn’t going to change; AI agents will still not excel at navigating it. Let’s throw out the notion of AI agents handling every task themselves.
At the same time, if your automation involves filling 100 cells in Excel, there’s no point in asking a human to step in. There goes all that time saved!
Thankfully, we’ve built an amazing way for robots to navigate the internet called APIs.
APIs allow us to interact with websites via code instead of a mouse and keyboard. They let us do repetitive tasks quickly, and you don’t have to worry about the first API call being different from the 500th.
An AI agent navigating through an 100x800 Excel sheet would take a long time. Remember earlier how we showed that errors compound over a ten-step process? Imagine 80,000 steps.
APIs are deterministic and fast. They’re great for the repeatable tasks that interact with the internet. AI agents understand natural language, and can work with the information that APIs send them.
Remember how a web page looked for AI earlier? Confusing at best.
def get_values(spreadsheet_id, range_name):
creds, _ = google.auth.default()
try:
service = build("sheets", "v4", credentials=creds)
result = (
service.spreadsheets()
.values()
.get(spreadsheetId=spreadsheet_id, range=range_name)
.execute()
)
rows = result.get("values", [])
print(f"{len(rows)} rows retrieved")
return result
except HttpError as error:
print(f"An error occurred: {error}")
return error
This entire API code block reads an Excel sheet cell in the same way every time. Reliable, predictable outcomes.
I know what you’re thinking, and no. A salesperson shouldn’t be expected to learn what an API is, or how it works. The right AI agent will know when to use APIs in a workflow. No human understanding needed.
Don’t abstract humans away, place them in the driver’s seat
There’s a notion that today’s agents just need to be told what to do once. That’d be nice, but reality is a bit different. The more autonomy you give AI to make decisions, the more time you’ll spend debugging, validating, and rerunning the automation.
The solution is staring us in the mirror. Since AI agents aren’t doing well in the driver’s seat, let’s keep humans navigating the process.
The perfect flow always starts with humans explaining the intent in natural language. The critical change is a feedback loop; the agent should return at each step in the process, allowing the human to verify correctness. Course correction happens here and prevents error.
Generative AI understands the intent and the mission, and creates the orchestration. The agent can hand the task off to an API (for repeatable tasks and robot-first automations) or to a human (for validation and status checks).
If we add features like memory and backtracking, we’re approaching a reliable system.
Discover automations you didn’t know you needed
The toughest part is deciding what to build. A great AI agent will discover tasks you’re doing that are repeated, and suggest them as automations.
This can work outside of workflows, and within them. AI agents can detect a manual step and recommend automation.
This proactive discovery can prevent errors later on. If you’re starting a new workflow in Salesforce, for example, the AI agent can pre-train with common Salesforce web pages and data to improve the chance of success.
The ideal AI agent shouldn’t just sit idly by waiting to improve.
What this looks like today
Let’s see this in action. An agent by itself versus the workflow we’ve described - agents + humans + APIs.
We’ll take the same workflow from before; take a company from Google Sheets, scrape the company’s webpage, then update that company’s information in HubSpot.
It’s important to note that an agent, by itself, could not do this. Google Sheets renders on the web page in a way that it can’t see.
Let’s assume, somehow, we got around that - just updating HubSpot alone with the UI is five tasks. At a (generous) 80% task success rate, our “Update HubSpot” step becomes a coin flip. And that’s one of the FOUR steps in the complete flow.
Let’s try that workflow using Bardeen, an AI agent that uses APIs and humans to improve performance.
Ready to get started? We can help.
AI agents alone can’t do everything. A capable AI agent, armed with APIs, is the perfect companion for any team member driven to work smarter.
We let humans explain the need to turn prospects into leads. LLMs to scan prospects and flag good leads. APIs to migrate their data from Hubspot to Gmail.
Thankfully, that’s exactly what we’re working on at Bardeen. If you’re stuck running repeated, automatable tasks all day, well, you shouldn’t be. Give me a shout at Artem at bardeen.ai. Better yet, try it out yourself for free here and let me know what you think.
.jpeg)
.jpg)
