The Context Advantage

A sales VP we know bought her whole team AI seats last year. Ninety days in, the usage charts looked great. Everyone logged in, prompting away.

The forecast looked the same as it had before she spent the money. The deals took just as long to close. And her best rep had gone back to writing her own emails because, in her words, the drafts sounded like a brochure for a company she'd never worked at.

Nobody screwed up. The tool did exactly what it promised. That's the uncomfortable part, because it means the next tool won't fix it either. This MIT number gets quoted a lot, and for good reason: their August 2025 study looked at 300 of these rollouts and 95% showed no return.¹ The handful that worked had AI wired into the actual job. The rest got it bolted on the side.

Everybody is spending. Fewer are winning.

For all of that money, 88% of organizations are using AI somewhere and fewer than one in ten have it running at any real scale.² Lots of pilots, almost no production. The agent wave behind it already looks shaky too. Gartner figures 40% of those projects get cancelled inside two years, and estimates that only about 130 of the thousands of vendors waving the word "agent" around are the real thing.³

The question nobody asks at the demo

Here's what never comes up while everyone's watching the demo: if the models are this good, why isn't more getting done?

Think about the best person you ever hired. Sharp, fast, knew the field cold. Day one, they were still pretty useless on the things that mattered, because they didn't know your accounts. They had no idea what you'd promised the customer in Tulsa, or that you'd killed that discount two years back. They had to learn from scratch.

The four things it needs to be useful

Building a context layer is mostly that walking-through. Four kinds of knowledge go in, and they are not equally easy to get.

Notice the order. The first two you can mostly buy or scrape. The last two you have to earn, because they live in people's heads and in records nobody has read. Hold onto that - it turns out to be the whole game.

Why the genius still gets it wrong

Underneath the magic, the model is guessing the next word. Hand it nothing solid and it'll guess something that reads beautifully and happens to be wrong - like a renewal email quoting a price you stopped offering in 2023. Feed even the best model the wrong context and you get the wrong answer back.⁴

It isn't a rare hiccup you can train out, either. Stanford ran 26 of the top models and the best one still missed about one time in five.² That's why "it's just wrong sometimes" is now the single biggest worry executives report about AI, ahead of security.² A tool that's confidently wrong every fifth try is a tool somebody has to check behind. So they check, and the time you were supposed to save walks right back out the door.

Wiring in your data is only half the job

You'll hear that retrieval handles all this. Point the model at your documents, let it look things up, done. It helps. It's also half the job, and the missing half is the half that bites. Retrieval pulls whatever's sitting there, and what's sitting there is usually a junk drawer. We once watched a model grab the wrong price sheet because two of them lived in two folders, both named "final," one a year out of date.

That's not the model's fault. Pointing AI at your data isn't the same as pointing it at your truth. Getting to truth takes two things companies tend to skip. One is ownership, so every piece has a person responsible for keeping it current and there's a single version that wins.⁶ The other is delivery, so the right thing shows up at the moment of the work instead of sitting in a system nobody opens.

The work behind the walls

There's a name for this now - context engineering - which mostly means deciding on purpose what the model sees each time it runs, so it behaves like it knows where it is.⁴⁵ The plumbing is standardizing, too. A protocol called MCP went from nobody's-heard-of-it to the default way to connect models to real systems inside a year, with more than 10,000 connectors built for it already.⁶ None of it is glamorous. It's the wiring behind the walls. Nobody points at it on the house tour, and nothing works without it.

A Monday morning, handed back

Take one rep on a Monday. Her first hour goes to reconstructing where things stand: scrolling chat, opening four email threads, checking the CRM, trying to remember what the account last heard before she gets on the phone. She's normal. Microsoft tracked this and found the average worker gets interrupted every couple of minutes, up to 275 times a day, on top of 153 chat messages and 117 emails.⁷

Give her the layer and the Monday changes. One screen. The promises, the last few conversations, the current price, the open issues, sitting there in plain language, all of it current. She gets the hour back, plus the second one she didn't know she was losing.

Here's the part that surprised us: the people who gain the most are the ones with the least experience. The research lands the same way - newer and lower-skilled workers improve most, because the tool finally hands them the information instead of making them grind for it over years.⁸ One controlled study this year had AI closing roughly three-quarters of the gap between weaker and stronger performers.⁹ Stanford's everyday figures are 14 to 15% more cases cleared an hour in support, 26% faster in engineering.² We've watched a brand-new rep sound like a five-year veteran by their second week, because the system was quietly feeding them moves the veterans had already worked out the hard way.

From minutes saved to job done

Move up to a team and the layer stops saving minutes and starts doing the job. A global professional-education company came to us with a number already in hand: a two-point lift in lead conversion was worth £6.5 million a year to them. They had no lead scoring to speak of. What they did have was tens of thousands of recorded sales calls that nobody had ever mined - a goldmine of history and language sitting in cold storage. We built call-center intelligence on top of those calls, and a scoring system that sorts every lead into one of four tiers, so a rep opening the queue sees who to call and what to do about each one.

That build is the four kinds of knowledge in miniature. The leads were facts. The calls were history and language. The tiers were rules, pulled out of conversations nobody had listened back to at scale. The model didn't supply any of it. We did, from their own material.

A different build got a different shape. A mobile-security company wanted more pipeline without hiring. We sat in on their sales calls for a week, wrote down the qualifying rules and the lines they never cross, loaded in the account history, and pointed the whole thing at one number: a new qualified opportunity per rep, per week. Not a dashboard, not a chatbot. You know it's working when the reps stop rewriting the drafts and just hit send.

None of this is edge-case stuff. The clearest, best-measured wins so far are in support and engineering - the high-volume work where the payoff is easy to see.²⁸ Picture a support desk whose virtual assistant has read all 40,000 of your closed tickets and knows the real fix, not a decent-sounding guess. That desk gets it right the first time.

One version of the truth, and it stays current

Zoom out to the whole company and the problem changes shape. Sales tells a customer one thing, support tells them another, both are sure they're right, and the customer decides neither of them knows what's going on. Or a deal falls apart, nobody records why, and a year later someone new walks into the same wall.

One governed layer means everyone works off the same story, and the history sticks around instead of leaving when people do. McKinsey's figure for companies that rebuild around this is 10 to 30% better operating profit, with one large bank now booking close to $2 billion a year from it.¹⁰ A competitor can buy your tools by Friday. They can't buy ten years of what you promised, what you chose, and why.

Pull an AI project apart and ask which pieces a rival could copy, and how long it would take them.

The model is a rental

The same models are sold to everyone, on the same terms. Nothing about the one you use is yours alone. Gartner files them under "strategic commodities" now, which is the analyst way of saying they don't set you apart.¹¹ The off-the-shelf tools aren't far behind, with AI writing and summarizing baked into Office, Workspace, and Notion already. Handy. Not an advantage. Anything a rival can copy over a weekend was never a moat.

Your context is the part they can't copy

Warren Buffett's old idea was the moat, the water around the castle.¹² But the moat's moved. It isn't the algorithm, and here's the part the whole "data is the new moat" chorus tends to skip: it isn't really your data either, at least not the way people mean it. Facts are the cheap part - everybody's CRM holds the same shape of records, and a sharp competitor can quickly approximate how you talk. What no one can copy is the following two: the judgment your best people carry in their heads, and the memory of what you've actually done and why.

That education company's edge was never that it owned call recordings. Plenty of companies own call recordings, and they're doing nothing with them. The edge was the knowledge trapped inside them, which had never been written down, and which only showed up once somebody did the work of pulling it out. That's the moat. Not the data - the rules and the history.

The market is already pricing this. The companies most exposed to having that advantage copied have trailed the more resilient ones by around 26 points this year.¹¹ Your context is either an asset or a liability, and there's no third setting.

A warehouse full of data is not an advantage

Here's where we part ways with many consultancies. Raw data isn't a moat. Insight is. Data sitting in silos is just a cost, not an edge, and stockpiling logs gets you a storage unit, not a castle. Messy, scattered data sinks whatever AI you build on top of it; clean, connected, governed data is the only kind that holds up.¹¹ That's the whole difference between a data lake, which holds everything and explains nothing, and a context layer, which you build to use.¹³

The honest caveat: a dataset on its own isn't safe forever. The big models have trained on nearly everything, so they can now do work that used to need your special data. Volume was never the point. What stays defensible is context you've structured, you own, you've connected, and you put to use. 

The slow leak

There's a quieter way this goes wrong, and it has earned a name: context debt. Context rots if you leave it alone. Facts go stale, rules start contradicting each other, the reasoning behind old calls evaporates. Teams will stand up a clean layer and then never touch it, and inside a few months it's giving worse answers than nothing would. So we put review dates on every piece now. People ignore them - right up until the AI quotes a dead price to a live customer, and then everyone cares about freshness.⁶

You've seen the human version. Your best operations person retires and a decade of "oh, here's how we actually handle that" walks out with her. Tend the context and it compounds for you. Ignore it and it compounds against you. You're feeding one or the other whether you mean to or not.

The cost of waiting

Run it forward eighteen months. Two companies in the same market, same vendors, same models on tap. One of them started building its context layer last year. Their new hires are useful in two weeks, their support desk closes on the first touch, and every quarter the layer knows a little more than it did the quarter before. The other company is on its third pilot, still impressed by demos, still asking why the numbers won't move.

The gap between them didn't open overnight. It compounded, the way these things do, because the first company's advantage feeds on itself and a stack of AI seats never will. By the time the second company works out what it's actually missing, the first one has a two-year head start on the one asset you can't go out and buy. That's the real risk of waiting - not that you fall behind, but that you fall behind on the thing that gets harder to catch every quarter.

When this clicks, the urge is to go big: a company-wide data program, the whole knowledge base, everything mapped at once. Don't. On one of our early builds we burned three weeks trying to document every rule before we shipped a thing, and it taught us the obvious lesson - nobody can recite the rules cold. They only remember a rule when somebody breaks it.

One circuit at a time

So we work the way an electrician does. You don't open every wall on day one. You start at the panel, wire one circuit, prove it carries load, then move to the next room.

One team, one workflow, ship it, expand. That's the same pattern MIT found in the projects that worked: wired deep into one real process, tuned as they went, owned by the manager who runs that work rather than a lab off in a corner.¹ It's also why "buy more tools" keeps failing. 88% of companies already own the tools.² The whole problem is the layer underneath them.

Owned, and connected

Two things separate a layer from a heap.

Days, weeks, and forever

The timeline is friendlier than people expect, because the pieces fill in at different speeds. Facts, days. Language and rules, weeks - and that stretch is tedious, no way around it, because most of it is locked in people's heads and only comes out when somebody trips over a rule nobody wrote down. That tedious stretch is also the whole point: the line between an AI you trust and one you keep checking behind. History compounds from day one, which is the real reason to start now and not next quarter.¹³

Don't just buy another tool

Last thing, because it's the wrong turn we watch people take most. When AI underwhelms, the reflex is to buy more of it - another model, more seats, a flashier agent. But dropping a tool onto a process with no context doesn't fix the process. It runs the broken version faster. Microsoft's own researchers put it flatly: the risk is using AI to speed up a broken system.⁷ The fix doesn't start with the tool. It starts with what the tool needs to know.

The model question is settled, for all practical purposes. Frontier capability is cheap, everywhere, and sitting in your competitor's account on the same terms as yours. The scarce thing is the context that makes it yours - and the scarcest part of that is the judgment and the history only your people hold.

Get it wired in and the day looks different. People get their mornings back. New hires earn their keep in weeks, not quarters. The team trusts what comes out without re-reading every line. Everyone tells the customer the same true story. And underneath it you're building something that compounds while your competitors keep buying tools that don't stick.

Remember the rep who went back to writing her own emails? Give her the right context and the drafts sound like her again, because they're built out of her best work and everyone else's.