Quick answer: The tools were never the constraint. Most companies bought the same generative AI everyone else bought, handed it to smart people, and stopped there – so the AI has no memory of how the business actually runs and produces generic output.
The fix is not a better model. It is connecting every meeting, message, and platform your company generates into one place the AI can draw on, then governing what happens once it can act on that context on its own.
Why isn't the AI you already bought changing anything?
AI is a business expander. Given the right tools, automations, and dashboards, it lets your most senior, most experienced people operate at multiples of their previous capacity.
Every other person on the team gets more independent too. For us, that meant we stopped needing a dedicated specialist for things like measurement, because people across the company can now build a dashboard that pulls from several platforms and produce an insightful takeaway themselves – not a chart, an actual read on what's happening and why.
None of that shows up automatically. The whole trick is consistently putting data in. You have to connect every piece of the organization – meetings, messages, documentation, platform data – or the AI is reasoning from nothing.
I have believed this for a while. What changed recently isn't the belief. It's that the tooling finally caught up to it. Early AI tools had hard limits on shared memory and context – you could not feed them enough, and what you fed them did not persist. That has been substantially solved. The only limitation left is how much of your own data you can get in front of the model.
What did we actually connect, and why did it take this long?
We made meeting recordings mandatory, internal and external, across the company. Every meeting, every phone call. That single decision produced more usable business context than any tool we bought.
The recordings get transcribed and put into a form we can extract insight from – not just "what was said," but patterns across meetings that feed marketing decisions, sales strategy, where we invest, and how we give people feedback on their own performance.
That meeting context doesn't sit alone. It combines with Slack conversations, email, internal documentation, and the performance data sitting inside our platforms. Together, that combined layer is what decides what content we publish on our own blogs, which clients we go after, and how a manager evaluates someone's actual output over a quarter instead of their most memorable week.
Getting all of that in was the hard part, and it's still the hard part for most companies, because the off-the-shelf integrations don't reach far enough. We built connectors for nearly every tool we run day to day. Where the standard integration stopped short, we built our own – a custom MCP for Microsoft 365 is one concrete example – to pull in more than the vendor's default connector allows.
This matters more for a company like ours than it would for a single-product business, because we are not one operating model. We run a performance network, a managed agency practice, a pay-per-call marketplace, and a media operation, spanning finance, insurance, home services, lead generation, mobile apps, and subscription services. The managed-program practice alone carries seven years of aggregate history, and a context layer has to absorb all of it or it isn't doing the job.
A context layer that only works for one of those business lines isn't a context layer at all. It's the same narrow tool with a bigger label on it.
Fourteen years of running a company across that many different models is what makes the gap between "AI as a tool people use" and "AI wired into how the business runs" so visible when you're standing where I stand.
Is the market right that AI isn't paying off?
The market is mostly right, and that's worth taking seriously rather than waving off.
MIT's NANDA initiative published research in 2025, The GenAI Divide: State of AI in Business 2025, that Fortune reported was based on 150 leader interviews, a 350-employee survey, and analysis of 300 public AI deployments, and found 95% of generative AI pilots delivered no measurable impact on the P&L. Only about 5% of integrated deployments created significant value, per the same coverage. That is not a small-sample anomaly. That is the market.
The important part of that research is the diagnosis, not just the number. The report attributes the failure to a learning and context gap, not model quality. Generic tools stall inside companies because they don't retain feedback and don't adapt to how that specific business works. The model is fine. The company never told it anything.
Half of that market view is right: AI is creating more work, not less, for anyone trying to make it functional. Standing up a connected context layer, governing it, and maintaining it is ongoing effort, not a one-time project. Anyone telling you this is a plug-and-play productivity gain is selling something.
The other belief embedded in "AI isn't making people money" is that AI output is inherently slop – generic, interchangeable, not worth paying for. That's true only if you haven't done the work above. If the AI doesn't know how your company writes, sells, and operates, and your first-party data isn't stitched into it, of course the output reads like everyone else's output. It was built from everyone else's data.
Solve the context problem and the path forward stops being about efficiency for its own sake and becomes directly tied to personalization and revenue creation – the AI's output starts sounding like your company because it's finally working from your company's material, not the internet's average of it.
The smart operator uses that to reinforce the business lines they already run, with more automation and more personalization layered on. The more ambitious move is going from an idea to an executed, live solution largely on your own, where the work product is recognizably a reflection of you, not of a vendor's template. That's the same in-house, build-it-yourself instinct that put our AgentFinder property together end to end without an outside firm.
What happens when a connected system breaks?
We run interconnected agents that operate independently – discovering content ideas and producing content across our own web properties without a person in the loop for every step. It works. It has also broken. How it broke is the part worth telling in detail.
On our AgentFinder property, a routine that ran a gap analysis and replenished the content queue got disabled. Separately, a rule fires content discovery automatically whenever the queue depth drops below a set threshold. With nothing refilling the queue, that rule did exactly what it was built to do – it just did it constantly. It fired 32 discovery runs in a single day instead of the one it should have run, some of them concurrently through a duplicate scheduler path we didn't know still existed.
That produced 237 model web-search calls in a day, some at very high input-token volumes. There was no spend ceiling on that function, only a reporting threshold, so it ran to completion and we found out the next morning. A single day ran at roughly thirty times the established baseline for that function.
We fixed what caused it: removed the duplicate scheduler, added a daily cost report with an alert triggered at three times the rolling average, and put manual review above a set daily threshold.
What happened two weeks later is what changes the diagnosis. A separate failure – this time a publish-failure loop, an unrelated mechanism – caused a second, smaller spike. Different bug, same shape. That told us this is a category of failure, not a one-off bug we patched. Any system where agents can trigger their own subsequent work needs a hard ceiling on spend, not just a better alert. We still don't have a hard per-project spend cap, and I'd call that the most important open item in this whole effort.
If you are wiring agents together and you have not asked what happens when one of them is wrong and nothing is watching the meter, that is the question to answer before you scale it, not after.
How much access is too much access?
Connecting everything creates a genuine trust question, and I don't think most companies have thought it through before they start.
We have become far more open with sharing and access to data than we used to be. A lot of what happens inside a company like ours happens inside business-line silos that not everyone can see into. That was always a drawback for leadership – people running one part of the business making decisions without visibility into how the others were performing.
AI let us fix that problem: a single dashboard, built from connected data, that gives the right level of visibility into how every part of the company is doing, for the people who need to see across lines, without hand-building and maintaining that view.
But the same connectivity that solves the visibility problem creates a permissions problem the moment you turn on something like mandatory meeting recording. When we made every meeting and call recorded and transcribed, we made those transcripts available to everyone who attended that meeting. That boundary mattered to us. If you weren't in the room, you don't get the transcript. Broad internal visibility into how the business performs and automatic broad access to what was said in a specific conversation are two different decisions, and treating them as the same one is how a context layer turns into a liability instead of an asset.
Who owns the cost of all this?
Somebody has to own it in arrears, on a schedule, or it drifts.
Our IT support function and our finance director own cost governance together. Every month, they look back across spend on API keys and agent usage, and at who moved into extra usage beyond their assigned plan. That review is what lets us move people onto higher or lower-tier plans and catch optimizations before they become a pattern.
We also created a dedicated AI revenue-operations role. That person's job is building automations and agentic workflows, both independently and in support of other teams, and developing a group of internal AI champions who keep pushing what's possible rather than settling at whatever the tools do out of the box. Part of that role is constantly testing new models and finding ways to bring them in – so we can reach frontier models for the hard reasoning problems while routing repetitive, automatable work to lower-cost models, including open-source ones.
McKinsey's State of AI research, published in November 2025, found that companies getting value from AI are about three times more likely to have strong senior-leadership engagement in how it's deployed, and are far more likely to have a defined process for when model outputs need human validation. Cost governance and validation are not an IT ticket you assign and forget. If leadership isn't in the review monthly, the review stops happening honestly.
Isn't this just an expensive way to rebuild reporting you could have bought?
"Context layer" is a rebrand of knowledge management and enterprise data warehousing, categories companies have poured money into and quietly failed at for two decades. You're describing an expensive, bespoke effort to build reporting and workflow automation you could have simply bought from a vendor who already solved it, and it is a lot of engineering effort for something that, on paper, looks like a dashboard.
That's fair, and it is an expensive way to build reporting. I won't argue that part away.
Here's why I'd still make the trade. We used to pay an annual subscription for a business intelligence tool – Tableau – that we no longer need. Even with that subscription, there were limits on how many data sources we could push into it at once, and getting value out of it required a dedicated analyst function whose job was building and maintaining the dashboards themselves. That's a standing headcount cost layered on top of a standing license cost, for a tool that still couldn't reach everything we now feed into a connected AI system.
Now that stitching happens automatically across data sources that used to require a person gluing them together by hand. Yes, we spend a meaningful amount on tokens to run it. That spend is nowhere near what the earlier setup cost us in license fees plus the dedicated analyst function required to make it useful. And the more important shift is who gets to use it: building a cross-platform report used to belong to one specialist, and everyone else waited in that person's queue. Now people at every seniority level can put a report like that together themselves, the same day, without filing a request.
So the honest answer to the objection is: it's not free, and it's not nothing – but it's a smaller, more distributed cost replacing a larger, more centralized one, and it hands the capability to build with it to the whole company instead of to one function inside it.
Does this mean the company needs fewer people?
This is a capability story, not a headcount story.
We are not going to stop hiring. If anything, AI has created more work for us, not less, and we are actively hiring for roles we weren't looking at a year ago.
The skill we're recruiting for now is a more native, working understanding of AI – people who can manage the additional work AI has created, using AI tools themselves, in new, hybridized roles that didn't exist inside this company before. Someone who can own an agentic workflow end to end, watch its outputs critically, and know when to intervene is a different job than the one it partially replaced, and it's a job we're hiring for, not eliminating.
The reason this holds is the same reason the thesis holds at the top of this piece: AI expands what a business can attempt. Every new thing we can now attempt – more content, more automation, more personalization, more internal reporting built on demand – is new work for a person to own, direct, and improve. The tools didn't shrink the job list. They lengthened it, and changed what belongs on it.
Is adoption really a technology problem?
No. It's a people problem, and it's the biggest part of this entire transformation.
You cannot hand someone a tool and tell them to use it. You have to actively teach people how to become an AI-empowered knowledge worker, and you have to oversee how they work that into their day, because different people need meaningfully different levels of hands-on support to get there.
Most people are already busy with a full plate of existing responsibilities. Telling them to "just learn this, figure it out" isn't realistic, and it isn't fair. So we built training around it, and we created a dedicated role responsible for both adoption and enablement – meaning that person works alongside people as they adopt these tools, not just provisions a license and moves on. We also built feedback loops so we can see where the support is working and where it isn't, and adjust.
McKinsey's research backs the instinct behind this: in its March 2025 State of AI survey, only 21% of respondents at organizations using generative AI said they had fundamentally redesigned even some of their workflows to fit how AI works, rather than bolting a tool onto the process that was already there. The companies capturing value from AI are the ones with structured validation and strong leadership engagement in the rollout, not the ones that distributed licenses, skipped the workflow redesign, and hoped adoption would follow. Change management isn't the soft part of this project. It's the part that determines whether the rest of it does anything at all.
What does a context layer actually produce, in practice?
Three things come out of a working context layer, and they're the ones that matter most in practice.
Demand generation. Automated content creation across social channels, feeding a pipeline of new prospects into sales, at a volume that would have required a materially larger content team to sustain manually.
Client assets, produced near-instantly. Landing pages and creative built from templates and rule sets we've already taught the AI, generated in a fraction of the time – work that used to take multiple weeks of production now happens in a sitting.
An automated monthly close. The close used to run through a single manual gatekeeper – one person who reviewed and approved every stat adjustment that individual account managers submitted before the numbers were considered final. That step was a bottleneck by design, and it was also inconsistent in a way nobody talked about much: what got questioned and what got waved through depended on which adjustments that one person happened to catch that month. We replaced the gatekeeper with an automated review, built on the same context layer, that applies the same adjustment rules to every account manager's submission every month instead of relying on one person's memory of the edge cases. The hard part wasn't wiring up the automation. It was forcing the judgment calls that used to live only in that person's head into rules explicit enough for a system to apply consistently. It saved time, and it improved the accuracy and consistency of the close itself, because the review no longer depends on a single pass squeezed in under month-end time pressure.
None of those three required a smarter model than the one everyone else has access to. All three required the context layer underneath them – the connected meetings, documents, platform data, and rules that let the AI know what a correct landing page, a correct stat adjustment, or a correct piece of content looks like for us.
Where would I start, if I were building this today?
One step, ahead of everything else: connect every data source you have and pull context from all of it into a single AI system before you do anything else.
That's not the interesting part of the project, and it won't produce a demo anyone wants to show a room. It's also the step that makes every subsequent step possible. You cannot identify efficiencies or build workflow automation across a team using a system that only knows a fraction of what that team actually does. Sequence it any other way and you'll spend the next year building automations on top of a foundation that has to be redone.
What's the one question that actually separates the companies getting value from AI from the ones that aren't?
It isn't "which model are you using." It's "who reviews what your agents did, and how often."
Every company evaluating AI right now can buy the same tools we can. The difference shows up in whether someone connected the company's own material into those tools, and whether someone is accountable for what the resulting system does when it's wrong. We found out the hard way that a rule which is correct in isolation can still run thirty times its normal volume in a day if nothing owns the ceiling above it.
The one thing I'd do differently from where we started: put a hard spend cap on every agentic project before it goes live, not after the first incident teaches you why you needed one. Everything else about connecting your data in is worth the effort. That part isn't optional, and it shouldn't wait for a governance review to catch up to what your agents are already doing on their own.
FAQ
What is a "context layer," and how is it different from a data warehouse or a BI dashboard? A context layer is the connected body of a company's own material – meetings, messages, documents, and platform data – made retrievable by an AI system so its output reflects how that specific company operates. A data warehouse stores data for reporting; a context layer feeds an AI that can reason across that data and act on it, and it needs far less dedicated specialist maintenance to stay current.
Why do most AI pilots fail to show up in a company's financial results? Research from MIT's NANDA initiative, as reported by Fortune, found that 95% of generative AI pilots produced no measurable P&L impact, with the cause traced to a learning and context gap rather than model quality – generic tools don't retain feedback or adapt to a specific company's workflows.
Does connecting internal data into AI create new risks? New risks come with it, and they're worth naming rather than assuming away. Broader internal visibility and automatic distribution of sensitive content, like meeting transcripts, are different decisions that need different boundaries. We limit meeting transcripts to attendees only, while using a shared dashboard for cross-company performance visibility.
What happens when autonomous agents are given the ability to trigger their own work? They will eventually find a way to run far more often than intended if there is no hard limit on spend, only a reporting threshold. We saw a single day run at roughly thirty times an established baseline before anyone reviewed it, from an agent rule interacting with a disabled process it depended on.
Who should own AI cost governance inside a company? IT and finance in arrears every month, reviewing spend against plan by function and by person, paired with a dedicated internal role focused on building and improving agentic workflows across the company. Leadership needs to be genuinely engaged in that review, not just informed of it after the fact.
Will AI adoption reduce headcount over time? Not in our experience. AI has created more work, not less, and we are actively hiring for new, more hybridized roles that require a native understanding of AI to manage that additional work. This is a capability expansion, not a headcount reduction.