Agentic AI and Automation
Examining construction, compliance, etc...
The growth of LLM’s are rapidly transforming AI Agents, enabling them to handle more diverse requests and increasing their commercial viability. We are entering an age of automation - many manual workflows will be replaced, efficiency will increase, and business models will drastically shift in the onset of agents. This article will provide a high-level overview of the mechanisms behind an AI Agent, current areas of interest being disrupted, and a few interesting startups within the space.
Mechanisms behind AI Agents:
In my previous article, I defined the workings of LLM’s and described “single shot” models, where a model is given a prompt with one example, and outputs an answer to that. Agents, on the other hand, use an iterative approach that involves reasoning behind actions and reflecting on decisions. Below is a visual describing an example process using Chain of Thoughts.
This form of reasoning follows an Agentic Framework defined by Andrew Ng as:
“Reflection: The LLM examines its own work to come up with ways to improve it
Tool use: The LLM is given tools such as web search, code execution, or any other function to help it gather information, take action, or process data
Planning: The LLM comes up with, and executes, a multistep plan to achieve a goal (ie: writing an outline for an essay, doing online research, then writing a draft, etc…)
Multi-agent collaboration: More than one AI agent work together, splitting up tasks and discussing ideas to come up with better solutions than a single agent would.”
This framework advocates adding intermediate reasoning steps including: Planning, Reasoning, Criticism, and Actions, which ultimately help reduce LLM hallucination and improve performance. What’s particularly interesting is the concept of multi-agent collaboration. In a complex environment, multiple agents with specialized, fine-tuned training, can be connected to a common knowledge base through simple APIs. Say we have an agent with an LLM trained on pharmaceutical data and access to medical documents on prescriptions. Let’s also assume the common knowledge base is a generalist, conversational bot on GPT4. If the conversational bot is asked a question about a specific prescription, it can delegate the expertise to the pharmaceutical agent and collaborate with it to produce more accurate outputs.
Multi-agent collaboration has already seen growth, with the rise of many multi-agent orchestration (MAO) platforms. Improvements in agent-collaboration may bring about new business models: Agentic-based pricing, a version of seat-based pricing where each ‘seat’ is an agent, or Agent-as-a-Service, where agents are provided as a cloud-based service and businesses only pay for the services they use.
Here is an example case study I found particularly interesting, which highlights autonomous agents put to the test:
Case Study: Autonomous cloud platform
High Level Goal: Allocate computing resources including CPU, memory, and storage to users dynamically in ways to achieve maximum performance, security and lowest user cost.
Sensory signals: Real-time computing usage of the users, total users cost, performance metrics
Expertise AI Agents: Resource Allocation Agent, User Cost Optimization Agent, Security Agent, Monitoring Agent, Cloud Manager as orchestrator
Environment: The hardware resources pool, APIs, and tools
The individual agents collaborate together to improve the user experience. For example, instead of users looking for ways to save cost, the User Cost Optimization Agent and Resources Allocation Agent will explore ways to reduce costs when users don’t need certain features. This autonomous cloud relies on the quality and accuracy of the agents, as well as ensuring that they act in the best interest of people - this can be negotiated through a “mediator agent” that has mechanisms in place to check the actions of other agents.
Now let’s examine some impactful use cases of AI automation.
Key Use-Cases of Agentic AI and Automation:
AI in AE&C (Architecture, Engineering and Construction)
The insights and startups I gathered below are based off of this article. Generative AI has displayed a variety of powerful use cases in the construction industry, which the graphic below helps visualize:
Some of the startups below were sourced from this spreadsheet compiled by Riccardo Cosentino. I came across unique solutions within AI-powered concrete solutions, payment platforms for construction, compliance automation, and construction documentation enhancement:
Document Crunch (9M Series A): Contract compliance and risk review solution designed specifically for construction. From pre-signature to project closeout, their purpose-built AI platform helps identify risks, understand complex terms, and ensure project teams stay compliant.
Contract Co-Pilot helps redline documentation faster and proactively identify risks —> they won the Best Overall LLM AI Breakthrough Award, which OpenAI won last year. (Document Crunch’s LLM is developed on a custom-built, construction-specific knowledge graph, claiming to have high precision)
UpCodes (3.5M Pre-series A): UpCodes is a comprehensive compliance and product research platform that accelerates design to construction. They offer a centralized, searchable library of construction regulations.
They recently launched UpCodes Co-Pilot, an AI research assistant that makes building codes easier and accelerates compliance research
Swapp (11.5M Series A - Israel based ): AI-powered construction documents solution. “SWAPP harvests value from your firm's previous projects and takes it forward to boost your architecture teams' performance.”
Converge.io (18M Series A): Concrete operating system: Predictive AI for concrete planning, and precast concrete logistics optimizations.
Kwant (3.9M Seed): Operating system for the modern construction workforce. An IoT solution that leverages smart wearables and sensors to connect with its platform insights tech, getting real-time data on productivity and safety on the job site.
Adaptive (19M Series A): The AI-powered financial management platform for construction. Adaptive combines expert customer support with automated bills, receipt capture, budgets, draws, reporting, and more - all fully integrated with QuickBooks.
Supply Chain Automation:
The supply chain process is often very large, with siloed teams and large amounts of unstructured data. Managing buyers and suppliers becomes increasingly complex with scale. However, for mid-market companies, supply chain management is a challenge because they may lack the necessary IT resources or can’t afford ERP software like SAP. Didero AI is looking to solve this problem:
Didero.ai ($7m seed round - July 17th): Didero is the “AI Agent powering superhuman procurement teams.” Didero integrates with existing systems to automate common workflows - from checking orders to running an RFQ process. Didero has some specialized models that do things like extract data from tables, purchase orders, and price lists
The key here is that AI brings an advantageous benefit for mid-market companies: these companies do not have strong negotiating power, so Didero can help them manage the grunt work and be proactive with negotiating contracts and making payments.
Kojo: ($39M Series C 2022): Construction material management software
Kojo tech is a company demonstrating a strong use-case of automation in the construction industry. They are a purpose-built construction materials procurement platform that supports the process of researching, selecting, ordering, and paying for the raw materials required for a construction project.
Kojo’s Workflow Automation: In 2023, Kojo launched AI-powered “Kojo-intelligence layer.” - Finds optimal price for a given material, automatically generates price quotes, generates real-time inventory projections for available materials, and allows contractors to see previous purchase analytics.
Why I’m interested: In May 2024, they launched Kojo AP: fintech offering to modernize payment process for contractors. Kojo AP eliminates many of the manual data entry processes for contractors' accounting teams, automatically catching mistakes on invoices and enabling vendors to get paid directly in one system. Kojo AP offers direct invoice integration with all major accounting software, invoice matching, and supports direct payments with Quickbooks and Vista. Additional integrations are planned for late 2024.
With the launch of Kojo AP, Kojo is a truly end-to-end software: this will significantly drive their revenues, upsell existing customers, and increase retention as they roll out the new feature to all customers soon.
Although Procore is a large competitor, they actually cross-sell with Kojo's construction materials procurement platform because it is "the best" in that space.
Automating workflows for compliance:
There has been a growth in the number of startups targeting compliance and documentation review, especially with the development of AI Agents. Here’s two example startups in this space:
Norm AI (27M Series A): Norm Ai pairs compliance and business teams with AI agents to proactively identify and mitigate risks. This platform of regulation-professional agents helps teams across tech, healthcare, law, environment, and more, stay compliant.
Legalfly AI (16.3M Series A): LegalFly offers a full suite of Legal AI agents designed to automate legal services, draft contracts, and review compliance risks. The company uses on-prem anonymization so data is secure and anonymized before it leaves a company’s workplace.
Infrastructure & Platforms for Building AI Agents
SuperAGI (10M Series A): A developer-first open source platform that provides infrastructure for building autonomous AI agents. They hope to build a full-stack artificial general intelligence platform based on large agentic models.
Platforms to build monetize, and integrate AI-Agents to automate workflows for your business:
ELNA: Decentralized AI Agent creation and monetization platform on blockchain
Fetch AI: Open-source platform to build AI agents
Nexus: Automate AI workflow with AI agents for different use cases
Dify: Build LLM apps using AI agents, workflow and RAG engine.
Beam: Agentic AI Process Automation platform for verticals including healthcare.
Ultimately, AI Agents are still in the early stages of development - many decisions are made at design time rather than run-time, agents currently require human intervention and assistance, agents must be highly fine-tuned to combat the increasing complexity of diverse tasks, and performance is reliant on the capabilities of our current LLMs. It will be exciting to see how AI Agents continue to evolve and integrate with many enterprises, leading to improved productivity and newer business models. That’s all for today, Stay Curious.
Great read!