AI & DATA STRATEGY
Road between mountains and the sea at sunset
Synaptiq helps you develop your AI and data strategy as well as accelerate your roadmap to achieve successful business outcomes. Assess your AI and data readiness so you can prioritize the gaps you need to fill.
Read More ⇢ 

 

    DATA LAKE
    Boat on a lake - AI Datalake
    Synaptiq helps you unify structured and unstructured data into a secure, compliant data lake that powers AI, advanced analytics and real-time decision-making across your business.
    Read More ⇢ 
      AI AGENTS & CHATBOTS
      Person on a construction crane - AI Agents and Chatbots
      Synaptiq helps you create AI agents and chatbots that leverage your proprietary data to automate tasks, improve efficiency, and deliver reliable answers within your workflows.
      Read More ⇢ 
        LEGAL SERVICES
        Wooden gavel on dark background
        Learn how Synaptiq helped a law firm cut down on administrative hours during a document migration project.
        Read the Case Study ⇢ 

         

          GOVERNMENT/LEGAL SERVICES
          Large white stone building with large columns
          Learn how Synaptiq helped a government law firm build an AI product to streamline client experiences.
          Read the Case Study ⇢ 

           

            strvnge-films-P_SSMIgqjY0-unsplash-2-1-1

            Mushrooms, Goats, and Machine Learning: What do they all have in common? You may never know unless you get started exploring the fundamentals of Machine Learning with Dr. Tim Oates, Synaptiq's Chief Data Scientist. You can read and visualize his new book in Python, tinker with inputs, and practice machine learning techniques for free. 

            Start Chapter 1 Now ⇢ 

             

              ⇲ Artificial Intelligence Quotient

              How Should My Company Prioritize AIQ™ Capabilities?

               

                 

                 

                 

                Start With Your AIQ Score

                  11 min read

                  From Answers to Action with Enterprise AI Agents: Webinar Recap

                  Featured Image

                  Chatbots can tell you your churn rate. In contrast, enterprise AI agents can find the accounts at risk, work out why, and execute the next step.

                  In a recent Synaptiq webinar, Dr. Tim Oates, Co-founder and Chief Data Scientist, explained how enterprise AI agents operate, how companies can integrate them, and why data governance and guardrails are what make them trustworthy. For most organizations, the difficulty isn't making AI smarter, but connecting enterprise data to enterprise work, safely.

                  From AI Answers to AI Action

                  Dr. Oates framed the shift as a progression, using an example of a company concerned about losing their customers:

                  • Chat: "What is our churn rate this quarter?"
                  • Answer: "Which accounts are most at risk, and why?"
                  • Assist: "Draft outreach for the top ten and recommend an offer."
                  • Act: "Send the offer where policy allows, escalate the rest to a human."

                  AI agents are moving fast. Citing Gartner, Dr. Oates noted that internal deployment is increasing and that 42% of enterprise tech leaders already have agents in production. Gartner also projects the share of companies using task-specific agents will grow from under 5% to 40%.

                  The gap between experimenting and delivering remains wide. Nearly two-thirds of enterprises have experimented with agents, yet fewer than 10% have taken them to tangible value.

                  Why AI Agents are Practical Now

                  The good news is that AI agents are now practical to build and run. Dr. Oates pointed to four developments that have come together to make this happen:

                  • Better models. Reasoning, tool use, and structured output have all improved.
                  • Accessible enterprise data. Lakehouses, semantic layers, vector search, and APIs put that data within reach.
                  • Models that use tools. Models can now work directly with SQL, search, applications, APIs, and code.
                  • Evaluation and governance. Tracing, permissions, monitoring, and human review make it possible to evaluate and govern agents.

                  Together, these developments let AI move beyond analysis and into the work itself. A problem, however, still stands in the way.

                  The Last Mile Problem

                  Dr. Oates went on to describe what he calls the last mile problem.

                  Today, it’s common for companies to extract information from data sources, build pipelines, centralize this information in a lakehouse, and build analytics and machine learning. Then a human is required to turn insights into action.

                  AI agents provide action. They identify the affected customers, investigate the underlying causes, review recent support interactions, recommend interventions, draft outreach, create CRM tasks, and escalate unusual cases when necessary. So, how do we ensure the AI agents are turning insights into the correct actions?

                  How Do We Ensure Enterprise AI Agents Act Appropriately?

                  An AI agent is more than a model. It is the system built around the model, which serves as its reasoning engine. An enterprise AI agent combines four core components:

                  • Model. The reasoning engine that understands requests and generates responses.
                  • Context. Data, documents, history, and business rules give the agent what it needs to know.
                  • Tools. Search, SQL, APIs, applications, and code let the agent take action.
                  • Control loop. The agent reasons, chooses a tool, acts, and observes the result.

                  AI agents don't just answer; they iterate. The result of one action becomes the input to the next decision. For example, a churn agent might find at-risk accounts, check their history, and retrieve the relevant policy before making a recommendation.

                  Agent Doesn't Have to Mean Autonomous

                  Although AI agents can act, that doesn’t mean that they have to act without oversight. Dr. Oates was clear that autonomy is a dial, not a switch. There are several levels of autonomy that can be granted depending on the guardrails in place:

                  1. Inform: "Here is the account's usage and support history."
                  2. Recommend: "I recommend a retention offer for this account."
                  3. Prepare: "I've drafted the outreach and the offer."
                  4. Execute with approval: "Approve the offer, and I'll send it."
                  5. Execute: "The offer is within policy, so I'll send it."

                  High-stakes tasks that demand a lot of reasoning should keep a human involved. Human-in-the-loop (HITL) is not all or nothing; it's a workflow design decision.

                  The Enterprise Version Adds Constraints

                  While it’s true that enterprise AI agents add value by taking on more autonomy, it’s important to acknowledge the constraints that keep that autonomy safe, accountable, and within defined boundaries. Dr. Oates highlighted five questions every organization should be able to answer:

                  • What data does the agent have access to?
                  • What actions will we allow it to take?
                  • How do we know whether it’s doing the right or wrong thing?
                  • What happens when it’s uncertain?
                  • Can we go back and look at what it did?

                  Building the agent is the easy part. Building one you trust inside the enterprise is much harder.

                  Finishing the Last Mile

                  Building an AI demo is easy. A demo has one user, one data source, one tool, a happy-path prompt, and manual oversight. An enterprise deployment, by contrast, has many users, many data sources, permissions, multiple tools, failures and edge cases, auditing, cost, and compliance, all of which create challenges.

                  When Enterprise AI Agents Break 

                  Dr. Oates described this as a stack of problems rather than a single one, so teams need to evaluate every level for weaknesses. Common failure points include:

                  • Data. Incomplete or incorrect data yields a confident, fluent, wrong answer.
                  • Access. The agent shouldn't see all the data. A churn agent doesn't need HR records.
                  • Reasoning. The agent can pick the wrong tool, misread a result, or stop early.
                  • Action. A wrong action or answer, such as emailing a customer that they're getting a refund when they aren't.
                  • Evaluation. Variable outputs and paths make testing much harder.
                  Data Governance and Guardrails

                  Dr. Oates also traced how the stakes rise as systems get more capable, from traditional analytics to agentic AI. More autonomy means a greater need for trusted data.

                  • Traditional Analytics. A person asks a known question and receives a known report.
                  • GenAI and RAG. A model retrieves and summarizes enterprise information.
                  • AI Agents. A model decides what information to retrieve and what to do with it.

                  That's why a firm data foundation has to come first. It's also why guardrails belong at every layer, with evaluation, tracing, and monitoring running across all of them:

                  Layer

                  Guardrail

                  User Request

                  Identity & Permissions

                  Agent Reasoning

                  Policy checks, logging, tracing

                  Data Access

                  Approved sources, row & column permissions

                  Tool Use

                  Tool allowlists and approval gates

                  Business Action

                  Threshold rules & human approval

                   

                  The principle running through all of it: never rely on the model itself to enforce the rules that constrain the model.

                  How Databricks Enables Enterprise Agents

                  Dr. Oates's approach is to start with governed data and build the AI platform on top of it, rather than moving data over to an AI stack. Databricks supplies the infrastructure around the AI agent, though not how the agent itself reasons. The key pieces he walked through are outlined below.

                  Structured and Unstructured Knowledge 

                  AI agents need both structured and unstructured knowledge. Databricks Genie agents handle structured data such as customers, transactions, usage, and financials. AI Search handles unstructured knowledge such as policies, contracts, support tickets, and documents.

                  Genie isn't just a chatbot; it's a tool that becomes part of a larger enterprise agent.

                  Model Context Protocol (MCP)

                  MCP is a standard way for an AI agent to discover and invoke tools. It lets a system say, "Here's what's out there, here's the input I need, and here's the output you'll get back," so different platforms can communicate. Databricks provides managed MCP servers for:

                  • Genie agents
                  • AI Search
                  • Databricks SQL
                  • Unity Catalog functions, which are governed calculations and a way to specify what an agent can do

                  AI agents can then reach enterprise systems such as CRM, Slack, Drive, APIs, and custom apps. Knowledge makes an agent useful; tools let it take action.

                  The churn agent, for example, pulls metrics from a Genie agent, supports records from AI Search, then pricing from a governed function, and writes to the CRM.

                  Unity AI Gateway: Control at Runtime

                  Every request from an AI agent passes through the Unity AI Gateway on its way to models, tools, and MCP services, carrying identity and permissions from the Unity Catalog. The gateway covers four areas:

                  • Access. Controls which models a team may use and which tools and MCP services it may invoke.
                  • Traffic and cost. Rate limits, quotas, budgets, and usage attribution.
                  • Guardrails. Policy checks are applied to requests and responses.
                  • Audit. Records which model was called, by whom, and with what, so you can reconstruct what an agent did.

                  The enforcement point sits outside the model, which is the whole idea. You can tell an AI agent it may only answer HR questions, preventing the agent from overreaching and answering questions outside of its governed scope.

                  MLflow 3: You Can't Manage What You Can't Measure

                  MLflow 3 for GenAI captures prompts, model calls, intermediate steps, tool calls, outputs, latency, evaluation scores, and human feedback. That supports a continuous loop: evaluate, improve, deploy, monitor. If a prompt isn't working well, you can improve it, redeploy, and keep watching. As Dr. Oates put it, observability is part of correctness.

                  From Prototype to Application 

                  The full stack runs from enterprise data in the lakehouse and Delta tables, up through the catalog (what data you can access), models, agents and multi-agent systems, the gateway (what an agent can do), and MLflow 3 (measuring everything). To put agents in front of people, Databricks offers an AI Playground for prototyping, an agent SDK that works with tools like OpenAI and LangGraph, and Databricks Apps for hosting experiences people can log in to. Synaptiq has built several of these.

                  The through line: data → knowledge → models → tools → agents → applications, with governance and evaluation carried throughout. 

                  Putting it Together: A Churn Walkthrough 

                  Dr. Oates closed the demonstration with one example question that carries a lot of enterprise context: "Which customers are most at risk of churn this quarter, and what should we do about them?"

                  Answering it takes revenue and account history, product usage, support tickets, contracts, renewal dates, and company policies. No single database contains the answer.

                  From Question to Investigation 
                  1. Identify high-risk accounts with a Genie agent and SQL
                  2. Investigate why, using a Genie agent and AI Search
                  3. Gather relevant context from contracts and support history
                  4. Recommend an action through agent reasoning
                  5. Prepare the next step with approved tools

                  This is a business investigation, not a generated answer.

                  When the Agent Acts

                  Consider this request: "Prepare a renewal offer with a 20% discount and send it to the customer." Company policy says account managers may approve up to 10%, and anything above 10% requires VP approval.

                  The agent drafted the renewal offer and prepared the approval request, but it did not send anything to the customer. That is autonomy within defined boundaries.

                  One interaction used many Databricks capabilities, such as investigation, recommendation, action preparation, and escalation. AI brought the data to an outcome much more quickly, with the guardrails intact.

                  Four Takeaways 

                  Dr. Oates closed with four ideas to carry forward:

                  1. AI agents act; they don't just answer. They bring some degree of governable autonomy to enterprise work.

                  2. An AI agent is a system, not a model. An LLM is a key part of an agent, providing natural language interaction and the ability to reason. But an agent also needs data, tools, and a control loop, and the amount of autonomy allowed is up to you.

                  3. The challenge is trust, not intelligence. Models are now so capable that intelligence comes more easily. Trust is the hard part. AI agents can't be relied on to behave on their own, so trust has to be enforced through mechanisms like Databricks Unity Catalog.

                  4. Put the AI agent where governed data already lives. Build the AI stack on top of your data rather than moving data over to the AI stack for the best results.

                  The AI ecosystem is complicated and evolving quickly. An experienced partner can help your organization move fast while building the trust that enterprise agents require. 

                  Interested in how your organization can put enterprise AI agents to work? Reach out to Synaptiq to learn how building AI agents on governed data in Databricks can drive measurable business outcomes.

                  Additional Reading:

                  From Answers to Action with Enterprise AI Agents: Webinar Recap

                  Chatbots can tell you your churn rate. In contrast, enterprise AI agents can find the accounts at risk, work out why,...

                  How to get ROI out of AI in the Construction Industry

                  Construction finance professionals face the ongoing challenge of managing tight profit margins, evaluating project...

                  What's on the Shelf? Machine Vision, LLMs, and Modern Shelf Analytics

                  Walk down any grocery aisle, and you're looking at one of retail's most valuable pieces of real estate, and one of its...