What is Headless AI? | Mid-Market AI | Episode 108

Parker Harris, co-founder of Salesforce, stood on stage at TDX this week and asked publicly: "Why should you ever log into Salesforce again?" That question is what this episode is about.

Headless AI means your software is about to start working whether a human touches it or not. But this is not about building autonomous agent fleets or ripping out your stack. It is about AI-driven actions delivered to the systems already doing the acting - one workflow at a time, with a dial that lets you decide how much human judgment stays in each loop.

This episode covers the three untenable extremes facing mid-market companies today, what headless looks like across six real-world patterns, the tariff demand planning crisis hitting distributors right now, how home health operators are recovering 3-4% of net patient revenue lost to claim denials, what Salesforce Headless 360 and the OpenAI Agents SDK mean for your stack, and why this wave is different from the failed headless ecommerce wave of 2019-2022.

For five audiences: the portco CEO navigating margin pressure, the PE operating partner building AI into the value creation plan, the MSP channel partner watching seat revenue compress, the vertical software product manager whose roadmap changes when the UI becomes optional, and the operator already building this without knowing what to call it.

Mid-Market AI is produced by Paragon Technology Solutions.

Paragon - Managed Intelligence Provider (MIP™)

Transcript

SPEAKER_00 0:05

Parker Harris, co-founder of Salesforce, 27 years building enterprise software. This week at TDX in San Francisco, he stood on the stage and asked publicly, why should you ever log into Salesforce again? Today's topic, headless AI. What it actually means, why it matters right now, and what to do about it. You're listening to Midmarket AI. I'm Ariel Jalali, CEO of Paragon, a managed intelligence provider. We put chief AI officer-led data and AI engineering capabilities inside PE-backed and mid-market companies. So who is this episode for? The Portco CEO navigating margin pressure and planning cycles moving faster than the business. This episode is about finding where that pain lives and doing something about it. For the PE operating partner who learned the ERP lesson the hard way, there's a new version of that mistake forming right now around AI. For the MSP channel partner watching seat revenue compress, this is the layer that's yours to own before someone else takes it. For the vertical software product manager sitting on years of domain data and expertise, this is what your roadmap looks like when the UI becomes optional. And for the operator already building this without knowing what to call it, well, here's the vocabulary to scale it and tell the story to your stakeholders and potential acquirers. Last episode recovered the harness. If you haven't heard it, start there. This builds directly on top of it, but is not mandatory in terms of understanding this episode. Bid market companies are already living these buzzwords. They built claims routing workflows and demand planning signals without calling anything fancy. They just called it Tuesday. This episode gives that work a name, pattern, and a way to scale it. Before I tell you what headless AI is, I want to name what it isn't because the two most common framings are both wrong for mid-market companies. And there's a third mistake that's just as dangerous. Humans is the integration layer. Right now in most mid-market companies, a human being sits between a system producing information and a system that needs to act on it. They read the report, they make the decision, they log into the other system and they take action. Data entry and screens. Slow, expensive, inconsistent business hours only. When they're good at their job, the knowledge lives in their head. When they leave it, it walks out with them. 84% of supply chain leaders say foreign trade policy changes are directly affecting their planning right now. 73% of healthcare providers say that claim denials are increasing. RCM teams spend 51 to 75 hours a week on denial work, losing 3-4% of net patient revenue in the process. These are not challenges that you hire your way out of. The human as the integration layer is breaking. Pure agent orchestration. A designed fleet of specialized agents handing off a complex multi-step workflow, autonomously deciding their next action at runtime, compelling in a demo, fragile in production, expensive to govern, hard to debug, and the design patterns are still evolving fast enough that betting your operations on one architecture today is a real risk. The AI space is simply moving too fast to put all your chips on one agentic design pattern. Then there's the option of doing nothing. Also untenable. The tariff cycle isn't waiting, the claims denial aren't waiting, the competitor who figured out the demand planning workflow isn't waiting. So how do you answer all three? AI-driven action is delivered to the systems already doing the acting, one workflow at a time, with a dial that lets you decide exactly how much human judgment stays in each loop per workflow independently based on the risk and confidence of the model or models that you're using. The ERP already fires purchase orders, you don't replace it. You just give it an AI-driven instruction set so it fires the right one at the right time without a human initiating it. Your scheduling tool already manages the care assignments or staffing assignments, you feed it an AI-driven recommendation so it reroutes automatically when a caregiver calls out. The production platform already controls the firing parameters. You ship it optimized recipes it didn't have before. The agent isn't the actor always. The existing system can still be the actor. The AI tells the system what to do and when without a human in the middle. That's headless AI in plain language, available today on the stack that you already have mostly. Developers call it headless. Analysts call it services software. The financial press called it what happened in February as the SAS Apocalypse. Gosh, that's a hard word to say. What it means for your company is that your software is about to start working, whether a human touches it or not. The interface isn't disappearing, it's moving. From a screen a human clicks to an API an agent calls. The software still exists, the data still live somewhere, the workflows still run, the login screen, the dashboard, the UI, that's becoming optional or changing a lot. So we've seen this movie before. Around 2019 to 2022, as we came out of the big bump in e-commerce from the pandemic, Headless Commerce was touted to be the future. Every agency was selling it. For most mid-market companies, though, it was a nightmare. More complexity, higher cost, longer timelines, and most quietly just back went back to their e-commerce platform of choice, be it Shopify or whatnot. The lesson learned there was headless is powerful with the right governance underneath it. In the wrong hands, it's just more surface area for things to break. I'll come back why this wave might actually be different, because I think it is, but I want to earn that claim specifically. Six pictures of what headless looks like in practice. Each one sits at a different position on the dial. Number one is browsers for agents. Most websites don't have an API, they have a UI. When an agent needs to interact with one, it uses a headless browser or an actual browser, in the case of personal productivity agents. And uh, you know, there's no screen, there's no clicking, the agent navigates invisibly for the most part, extract what it needs, takes the action, highly autonomous, works on anything with a UI. Then there's GitHub for agents. For your Azure shops, Microsoft owns GitHub, and that's not a coincidence. The agents open pull requests, run tests, review code, ship changes. Developers already live in GitHub. The agents are just moving in, semi-autonomous, and humans review uh before they do any code merges. Then there's data notebooks for agents. Where your machine learning engineers live, a good data scientist costs a lot and takes six months to find at least, and by month 20 are probably a flight risk. So agents now run those notebooks autonomously, build the models, test the hypothesis, iterate, surface the findings, autonomous on execution, but human on the interpretation. Then you have this notion of BI with arms, or as I call it, a report with arms. You have reports, someone runs them, someone reads them, someone takes action. Maybe. BI with arms means that the report detects the pattern and triggers the action inside whatever system already exists. The BI layer catches a supplier lead, drifting outside tolerance. Instead of flagging it in a dashboard that nobody checks until Thursday, it writes a new purchase order review task in the ERP, pushes a Slack or Teams message to the right person with the data already attached. It doesn't just sit there in terms of insight, it moves. So this is your starting point. Human reviews, system acts. This is what we would call the crawl phase. Then there's the concept of accessibility. The people in your organization that are locked out because the software was too complex, the warehouse manager who never learned all of the tricks in the ERP or the WMS, the field rep who never figured out Salesforce or refuses to update it. When the interface is a conversation instead of a dashboard or can be pulled magically from other sources, the whole company gets access to the data. For a PE operator, that's an eBit Dust story. So in terms of Salesforce going headless, why do we care? When the largest CRM company in the world said that the UI is optional and everything is an API, agents can run the whole platform without a human ever opening a browser. That's the incumbent validating the entire dial. And it has profound ramifications for anyone building software or is in the software space. Start with BI as arms. Every company has a report someone reads and manually acts on. Find that report this week. Let's talk a little bit about industrial distribution and the tariff demand planning problem. So demand planning has always been hard. What tariffs did was make it impossible to do manually at speed. Supplier costs shift overnight. The freight lanes that made sense last quarter don't make sense this quarter. 84% of supply chain leaders say that trade policy changes are directly affecting their planning right now. Traditional demand planning was built for a world where inputs were stable for months. Well, that world is gone. A schedule agent pulls supplier pricing feeds, freight indices, current inventory levels, and it runs a scenario model. Above the confidence threshold, it writes a draft purchase order into the ERP for review. Below it escalates the human with analysis already attached. Your demand planner stops spending three days building the model and starts spending three hours making the call on the scenarios that actually need the human judgment. Monday morning, identify two to three data feeds driving your demand planning decisions. Are they being monitored weekly on a human cycle? Well, that's a gap. You can reduce that from weeks to days to hours to minutes. That's where the agent goes. Another example, specialty manufacturer, quality routing and compliance. A food and beverage manufacturer, for example, with multiple SKUs and a lot line of traceability requirements. Every shift change, every batch run, every QC flag, historically a human reads a report, decides what to escalate, routes the exceptions, and at scale, that's a full-time job just managing the queue. Agents monitor production line data, detect the anomaly, cross-reference against the spec tolerances, and route the exception to the right person before it becomes a line stoppage or a recall. The quality manager only sees what's actually needed for a human decision. Where the model isn't confident, it doesn't act. It escalates. The audit trail is automatic. Another example is in the area of home health, claims and care staffing. 73% of healthcare providers say claim denials are increasing. RCM teams are spending 51 to 75 hours a week in denial work. Nearly half losing 3-4% of net patient revenue. That's not a staffing problem, it's more of a rules-based workflow problem. And agents were kind of born to solve this type of a problem. You can have agents monitor payer portals headlessly, catch the denial the moment that it hits, pull the patient record, identify the fix, resubmit within hours instead of days. That's cash flow. That's DSO. That's the metric your PE sponsor tracks weekly. In terms of care staffing, a caregiver calls out, a patient needs change, a certification lapses, the system reroutes automatically, surfaces the exceptions that need a human, keeps the schedule intact without a coordinator spending three hours on the phone. The pattern across all three. Find the workflow where the human is spending most of their time managing data rather than making decisions. Put the AI-driven instruction directly into the system that was going to take the action to begin with. Here are some examples from our own work here at Paragon. The first is in manufacturing, specifically a manufacturing control software company. Their platform manages high-temperature metal firing processes. Decades of firing recipes accumulated inside of it. We built a model on top of that recipe data, optimized it, and shipped the improvement parameters back through the platform's existing APIs. The controls operators didn't get new software, their system just started making better decisions, smarter. Defects down, energy consumption down, quality up. The platform became the delivery vehicle for intelligence it didn't have before, without a single customer changing how they operated it. Another example is in the area of CX. We worked with a CX platform that was at least 15 years in production. It's a good example of software that was built before the era of AI, as most software has been in terms of enterprise space. Nobody was really touching it or upgrading it. It was in that sort of precarious phase where you could apply patches, but it was a little bit too brittle to do a massive rewrite. So we built machine learning models to find the shortest path between customer intent and resolution across all the major intent types. We called this distance over classification. And the optimization shipback as a scoring layer. Every incoming interaction gets routed by the model before a human or legacy rule ever sees it. Time to resolution drop. Nothing got ripped out. It just got smarter. Another example, we're working with a top healthcare software company installed on-prem. It's a heavy regularly regulated space. The staff is doing manual data entry one screen at a time. We helped them build a model processing orders at over 96% accuracy and ship the result as a pre-fill layer. Staff review and confirm rather than enter from scratch, skipping multiple screens. The software didn't change, the workflow did. So three industries, three legacy systems that nobody was brave enough to wholesale replace, same pattern. Build the intelligence on top of the data that already exists, ship the output back through whatever integration surface is already there. The platform gets superpowers. Nobody migrates. Product managers at vertical software companies. That's the question for your next roadmap session. What data has accumulated in your platform and what would be worthwhile for your customers if it started firing headlessly into their existing workflows? We call this moving from data in the business to data as the business. Let's talk a little bit about what we're learning about the Salesforce Headless 360 announcement. Because it's not just about Salesforce. Every capability, CRM service, sales, workflows, Slack, Teams now exposed as an API or an MCP tool. We'll get to this buzzword in a second. Stands for Model Context Protocol. It's basically how agents talk to each other. It's like the Agentic API or Command Line Interface Command. 60 new MCP tools, more than 100 tools immediately available from Salesforce and their ecosystem. The AI agents in Salesforce's Slack tool alone are up 300% since January of 2026. And that's already happening inside of their platform. Salesforce drew a useful distinction. Customer-facing agents need to be tight, deterministic, brand controlled. And they have something called the Ralph Wiggham loop, which is uh named after a beloved character from The Simpsons that just keeps iterating uh dumbly on tasks until he succeeds. It's a dynamic agent that autonomously decides its next step at runtime. That distinction is the dial made visible. Some workflows need a fence, some need freedom. The harness determines which is which. Our last episode, incidentally, was about the AI harness and the reference infrastructure required to run it. For your Salesforce portcos, if you're a PE partner, the CRM, your sales team logs in is becoming a platform that agents run autonomously. The question for your Salesforce admin this week is which workflows still require a human to initiate them, and which of these are just rule-following workflows. A bigger question about using a verticalized tool like Salesforce, look, you have to imagine that every enterprise software is doing a land grab right now. Every single one wants to become your agenc platform of record. We've talked in prior episodes about making two-door decisions. I'm a firm believer in right tool or right platform for the job, as opposed to putting all of your chips on one table. So my recommendation is to get as much headroom as you can out of Salesforce, but don't make it another tenpole in your enterprise for every single agent that you need to build. Have it focused specifically on the CRM-related use cases. The other announcement that we want to talk about is this recent April 15th announcement from OpenAI, where they announced OpenAI agents SDK, which is a software development kit. This is how at a deeper layer software applications talk to each other. The shallower layer is an API, the heavier layer is an SDK integration. One model native harness with a native sandbox execution. Agents work on long horizon tasks, multi-step, running over minutes or hours in controlled isolated environments. The key architectural decision, they separated the control harness from the compute layer. Your credentials never touch the environment where an agent executes. Any ejected malicious command cannot access the central control pane. The agent operates in a sandbox, the sandbox is disposable, the harness persists. By the way, this is a similar architecture to what NVIDIA open sourced at their recent uh GTC conference with NemoClaw. It was a secure governed harness for frameworks like OpenClaw. We're going to see a lot of these frameworks for orchestrating agents in a secure enterprise governance-centric mindset. And we're going to be keeping an eye on them at the podcast and at Paragon. If you're a PE PortCo CTO, this is what your production grade agentic infrastructure should look like. Scoped, auditable, revocable access. Think about these as company badges or security credentials. If your current agent deployment doesn't have those three properties, that's a gap to close before you scale. All these announcements are saying the same thing. The model is not the product anymore. The software is not even the product anymore. The governed infrastructure around the model that interacts with the software is the product. For our MSP channel partners, the headless wave creates the biggest opportunity and the most direct threat to your current business model at the same time. The thread first. Your seat-based revenue is compressing from both ends. Agents don't need more licenses, typically. When a client automates their claim coordination workflow, they don't renew those seats. When their CX platform gets AI embedded, the agents don't show up in the license dashboard, and the vendors are moving to capture the new pricing layer themselves. Microsoft, for example, is already positioning agents as licensable entities within their own identity and access model. The vendor who owned the seat layer is trying to own the agent layer too. Forty eight percent of MSPs rank AI and automation at the top client need in 2026 ahead of security and backup. Only 13%, though, are generating meaningful revenue from those services. That gap is your window. It won't stay open for long. The opportunity, the layer the vendors cannot own is the governed integration between AI driving decisions and your client's specific operational systems. That requires knowing the client's operations deeply, which you do, which are workflow and rules-based, and they need human judgment. This is where data lives. What the failure modes look like, that's your advantage. You've been inside these businesses and in these trenches for years. But here's the thing: your clients are not just calling up and asking for headless AI. Hey, I want an extra order of headless AI with pepperoni and mushrooms. It's not like that. They're calling because their demand planning partner is underwater, their best billing coordinator just quit, their PE sponsor is asking hard questions about efficiency and costs. The MSP who can hear those calls as AI-driven action opportunities wins the decade. The capability gap, AI engineering, model building, business architecture, harnesses, the optimization API pattern. That's where Paragon comes in. The MSP brings the client relationship and operational context. We bring the integration and intelligence layer. If you're an MSP partner thinking about moving to this space, that conversation is worth having. Ask us about our managed intelligence provider um solutions. Little plug for Paragon. So this is a trend that's also very specifically for private equity operating partners. And it's an argument I haven't heard anyone making yet. You learned something expensive in the 2000s and the 2010s. Skipping ERP modernization at acquisition was a false economy. You'd buy the company, inherit the spaghetti code, and three years later you're doing a $5 million ERP implementation under time pressure. That lesson got learned. It's in the playbook now. You understand the value of consolidating and upgrading ERP. But just when you learn that lesson, we're going to move your cheese a little bit. There's a new version of the mistake forming right now around AI, and it runs in both directions. Some operating partners are saying, do the infrastructure upgrade first, then we'll be ready for the AI. Others are saying skip the upgrade and just layer on AI on top of what we have. And both are actually the wrong decision. Headless AI is the bridge between those two difficult and impossible decisions. It doesn't require 100% clean infrastructure. You ship AI-driven decisions and actions into the legacy ERP today. You upgrade that as time permits. You get operational value now. You build the institutional knowledge layer, the decision logic, the proprietary workflows encoded as models. That doesn't get thrown away when the modernization happens. It gets ported, it gets better, it compounds towards exit. The intelligence layer you build today survives the infrastructure transition. It gets more valuable on the other side because it's running on cleaner data, on a more modern system, and you're doing this incrementally. You're not choosing between AI and the ERP upgrade. You're sequencing them so both investments compound together. Good today, better tomorrow. And the portfolio play, let's say that three portcos with the same demand planning problem uh present themselves. Two healthcare assets with the same claims DSO issue, one workflow built once, deployed across the portfolio in terms of knowledge. That's an operating partner playbook, not a company-by-company initiative. The Futurum Group's 2026 survey of 830 IT decision makers found that enterprises have moved past the productivity argument for AI. Direct financial impact, revenue growth, and margin improvement nearly doubled as the primary ROI metric. Sales teams leading with, hey, save four hours per week are kind of in a backwards conversation. The board wants AI connected to the PL. Headless AI deployed on the systems you're already doing the acting is the most direct line from AI initiative to margin improvement. So this time it's different. We've seen headless fail before. So let me be specific about this wave and why it might actually be different. Because this time it's different, and I'm you can't see me, but I'm doing heavier quotes, deserves to be earned. The last time we tried to go headless, the primitives weren't there. Chatbots were rule-based and brittle. And nobody really wanted a chat with a chatbot before ChatGPT, let's be honest. Models couldn't reason across steps, agents couldn't plan, check their work, and try again. The vision was right directionally, but the infrastructure was a bit of plywood. What changed is the whole stack? Transformers gave us language understanding that actually multimodal AI means that agents can see, read, interpret the world in a way that a person does, and agentic frameworks mean that a model can decompose a goal, execute, evaluate, and course correct. The sandbox execution, open AI, and others are shipping just mean that the loop runs safely in production. The plywood just got replaced with structural steel. What's also emerging is a healthy separation between understanding, planning, and doing. And over the course of these episodes, I encourage you to think about those buckets separately and pick the right tool for the job and your reference architecture, as opposed to just going all in on one, on one vendor. The model understands the context, the agent plans the steps, the integration executes. That maps onto how enterprises, how they're actually structured. And it means that AI slots into that structure rather than replacing it. You don't need to be on the frontier. You need to find a flexible agentic design pattern framework that works for you. You don't even need to bet on one specific agenc design pattern. Um, there are many different patterns that can play together that can iterate. The cool thing about this set of tools is that you can iterate much faster than than you could in the past. So you, you know, in the course of shipping one thing in the past, you can ship a hundred things and see which ones win. So you plug into the layer that makes sense for your way your business is today. Your 2008 ERP isn't a liability. It actually can become the foundation. You can make the most important things better, one automation at a time. That's not a consolation prize. That's the right strategy for a mid-market company compounding towards exit and greater value creation. Before you move any workflow off human-initiated, you should really ask yourself three questions. One, does the agent have scoped access, only what it needs, nothing else? Is every other action logged, which is an audit trail that you could hand to a regulator or an acquirer? Is there a confidence threshold? Low certainty decisions escalate to human rather than execute? If any of those is no, close the gap first. That's what tells you to move the dial without breaking everything. On the protocol layer, a quick note, because you're going to hear a lot about this. MCP or the model context protocol is the current standard for agent-to-tool communication. 97 million SDK downloads, adopted by every major AI platform. And as Salesforce has announced, already 60 new tools from Salesforce this week alone. But it has real security vulnerabilities that are being actively worked on as of the time of this podcast. 30 CVEs filed in the last two months of 2026, including a near critical remote code execution flaw. Not a reason to avoid MCP, just a reason for your harness or reference architecture to sit in front of it. Let's talk a little bit about agent-to-agent. So A2A or agent-to-agent protocol from Google handles how agents talk to each other across vendor boundaries. Both MCP and A2A now sit under the Linux Foundation's Agentic AI Foundation, co-founded by all the big companies, including OpenAI, Anthropic, Google, Microsoft, AWS, and Block, providing neutral governance, convergence pressure. And let's assume these protocols will evolve. So don't architect around either one specifically, but build your harness so that your operations don't feel the pain when they do. There's a lot of shifting sounds in terms of the protocols, and things are moving very quickly there. For Azure shops, Headless 360 and the OpenAI SDK are arguments for building your own governance layer inside of your Azure environment before the headless wave hits your stack. Your vendors are going headless one by one, own the layer between their APIs and your operations. At Paragon, incidentally, this is what MIP delivers: the harness, the optimization API, the architecture you own, not Salesforces, not OpenAIs, not anthropics, yours, built for the two-door world. So do you know which systems in your stack are already AI accessible or API accessible, meaning that an agent could use them today without a human logging in? If no, that's your first homework assignment. Is there a report in your business that someone reads and manually acts on? That's BI without arms. That's your first use case. Name it. Start there. When your software vendors go headless, and they will, one by one, do you own the governance layer between their APIs and your operations? If no, that's the conversation to have with your operating partner and your CTO this week. So, like I keep saying, this is a fast-moving space. And so I'm going to be looking for five signals that would cause me to update my worldview around anything that we talked about today. The first is agent reliably crossing a production threshold. The human in the loop assumptions in this episode are real constraints, not design choices. If reliability crosses the point where you can trust a long horizon task to run unsupervised, staffing model assumptions, for example, shift materially, but we're not there yet. Secondly, if Azure ships turnkey governance, if Microsoft ships native agent governance in terms of product that plugs into existing enterprise agreement and it just works, the harness mode shrinks. Watching the Azure AI Foundry roadmap here. The third is vertical software MA premiums emerge explicitly. When deal announcements start calling out headless architecture and proprietary data layer in the valuation rationale, that's when this becomes table stakes. I expect for first clear examples within the next 18 to 24 months. And finally, protocol consolidation is accelerating. So if MCP and A2A converge faster than expected, the protocol layer gets simpler and the timeline on everything in this episode pulls forward. Final point, there's always regulatory movement, specifically in highly regulated spaces like healthcare and finance. So if autonomous agent actions in healthcare billing or financial workflows attract a lot of federal or state regulation, that raises the governance bar, which actually favors operators who built the harnesses early. So if you're in a health or financial services space, make sure that your harness is already producing an audit trail that a regulator would ask for. None of these actually break the thesis, but they're just something that I'm going to be having a lot of AI-configured alerts on and watching the space closely. Give me some feedback and tell me which part of the playbook to adjust and when as we go along. The manufacturing control software company whose firing recipes now ship optimized into every customer's production environment. The contact center platform that got smarter without anyone touching it. The healthcare software where staff can skip screens because a model processed the order first. None of that came from a keynote. It came from operators who understood their data well enough to make it work for them. You don't really need to rip and replace. You need to make your data actionable. One use case, one optimization, one automation at a time. BI with arms is your crawl. Governed agents running headlessly inside your existing stack is your walk. Your data working for you around the clock inside your environment and under your governance, compounding towards exit, that's your run. Crawl, walk, run. To the portco CEOs, find one workflow where the human is a bottleneck between data and the action. Start there. To the PE operating partner, build your intelligence layer now. It survives the infrastructure transition and it gets more valuable on the other side. To the MSP channel partner, the governed integration layer is yours to own. Move before other vendors take it. To the vertical software product manager, your moat is the data, not the UI. Start shipping it headlessly. And to the operator who's been building this without even knowing what to call it, I salute you and celebrate you. You've been right all along. Now that you have the vocabulary and the architecture to scale what you started, go forth with confidence. Send this episode to your operating partner, send it to your Portco CTO, send it to your channel partner friends, send it to your product team. If you haven't heard the last episode on AI Harness, I'd recommend going back and starting there. The episode is the market. This episode is the answer. I hope you're outside on a walk right now. Go enjoy it. I guarantee you the AI space will have changed by the time you get back. Cheers