What your firm gains when its AI works from one shared context

In most of the founder-led firms I work with, AI arrived from the bottom up. People taught themselves, each found a different use for it, and each is often saving something like 5 to 15 percent of their time, whatever the exact figure turns out to be. The founder picks up tips along the way and puts some of them to use, as needed. It is a sensible way for a small firm to adopt a new tool, and it is fine.

The next step is not a new idea. Firms have wanted their knowledge written down in one place for decades, from operations manuals to intranets and wikis. The value was rarely in doubt. What was in doubt was whether it justified the work: the record took real effort to build, more to keep current, and it only paid off if people remembered to read it.

AI changes that on both sides. It makes a written record of how the firm works, from its philosophy and ideal customers to its processes and individual work, much cheaper to build and maintain. It also means the record actually gets used, because every person's AI draws on it in their daily work. That record is what I will call the firm's AI context, and the benefits it was always supposed to deliver may, for the first time, be within reach of a small firm.

Individual gains do not add up to a firm that works better

The gains from bottom-up use are real, and they are private. They live in one person's habits, prompts and chat history. When that person is out, the work slows back to its old pace, and when they leave, what they taught their AI leaves with them. Two people doing the same task can get noticeably different results, because each has explained the firm to their AI differently, or not at all.

Ethan Mollick put the underlying problem plainly in Making AI Work: Leadership, Lab, and Crowd (One Useful Thing, May 2025): "AI use that boosts individual performance does not naturally translate to improving organizational performance." Something has to change in how the firm works before the individual gains become the firm's.

Microsoft's research points the same way. Writing on the 2026 Work Trend Index, Shashi Bellamkonda notes that organizational factors account for about twice the reported AI impact of individual effort alone. One caveat matters: the index surveys only people who already use AI, so it describes the engaged part of the workforce rather than everyone. Even so, it suggests that what the organization does around AI may matter more than how hard individuals try.

In a founder-led firm this lands on a familiar problem. The context that made the work good, meaning who the firm serves, how it decides, and what it will and will not do, has usually lived in the founder's head. Bottom-up AI does not change that. It spreads pieces of that context across more chat windows without writing it down anywhere. The firm becomes a little faster in many places and no more transferable in any of them.

The firm becomes a little faster in many places and no more transferable in any of them.

The value was never the problem. The upkeep was.

When the idea comes up today, the argument for it is usually defensive. The tool you rely on may change, or be overtaken by a better one. The cost of using AI may change. What the firm knows sits in chat histories and tool memory it does not own and cannot easily move.

Each point is true, and none of it feels urgent when today's use is working. A defensive case asks a pragmatic founder to pay for insurance against something that has not happened yet, and many reasonably decline.

The older argument, that the firm would be better off with its knowledge written down, was usually right. It failed on economics rather than logic. Writing the record down was slow, keeping it current was slower, and nothing made people use it. A wiki paid off only for the people who remembered to open it, so the value stayed theoretical while the upkeep stayed real, and the record was often the first thing to slip in a busy quarter.

AI changes both sides of that trade. On the cost side, much of the labor is now cheap: collecting what people already know, drafting pages from a conversation, spotting where two documents disagree, and flagging what has gone stale. On the value side, the record no longer depends on anyone remembering to read it. Every person's AI draws on it in the work it helps with, so the benefits arrive by default rather than by discipline.

The firm's AI context runs from philosophy to individual work

Tobi Lütke, Shopify's CEO, described context engineering as "the art of providing all the context for the task to be plausibly solvable by the LLM," as quoted by Simon Willison in June 2025. Most people practice that as an individual skill, one prompt at a time. The firm's AI context treats it as a firm asset instead. The context is written once, maintained, and available to everyone's AI.

It has four layers, and they differ mainly in who owns them and how often they change.

Philosophy and direction

Why the firm exists, how it makes decisions, what it will not do, and where it is heading over the short, mid and long term, including the market changes and trends it is responding to. This layer changes rarely and usually belongs to the founder. It is also the layer most often missing from any written record, because the founder has never needed to explain it to anyone in full.

Ideal customers and offers

Who the firm serves and who it does not, what it sells, what its terms mean, and how it sounds. Much of this already exists somewhere, in a proposal template or a page on the website. It is rarely in a form an AI can draw on consistently.

Cross-functional processes

The handoffs between functions: sales to delivery, delivery to account management, operations to finance. This is where context is most often lost in a growing firm, and it is the layer bottom-up AI use never reaches. Each person's AI helps with their own side of a handoff. Nobody's AI owns the space in between.

Function and individual work

How one process runs inside one function, with its inputs, steps and examples of good output. Below that sits each person's own way of working, which can stay personal. The problem is not personal context. The problem is when the first three layers exist only there.

The layers live in plain documents the firm owns, outside any single tool. That is also what turns the defensive argument around. If the context belongs to the firm, a change of tool is a change of tool and nothing more. The goal is not a better AI tool. It is a better written record of how the firm works.

The same pattern holds in my own work at a much smaller scale. The more cleanly and concisely I document data structures, people, philosophies and direction, the easier AI becomes to use in a shared environment, and the less each new piece of work depends on explaining things from scratch.

Create it from both ends, then make it the default

Done as a documentation project, this still tends to stall, as it always has. What has changed is that the work splits into three parts, create, use and maintain, and AI makes each of them lighter.

Create it from both ends

Start bottom-up. Each person who uses AI well has already had to explain something about the firm to get a good result: who the client is, what good looks like, how the firm talks, how a task runs. Collect those explanations one workflow at a time, along with whatever documentation already exists. Existing documents are useful raw material, but they are often stale or contradictory and written for a binder rather than for AI, so treat them as a first draft.

Then the founder adds what only the founder can: the philosophy, the ideal customer, and where the business is heading over the short, mid and long term. AI can draft most of this from a recorded conversation, so the founder's job is correcting and deciding rather than writing from scratch. Where the bottom-up versions disagree with each other or with the founder's pages, for example when three people describe the ideal client three different ways, that disagreement is the founder's to-do list.

Pick one workflow, ideally one that crosses two functions, and write only the firm-wide basics it depends on. Use it on real work for a few weeks, then move sideways to the next workflow and reuse what is already written.

Make it the default

This is where AI changes the old problem most. A record used to depend on people remembering to open it. Now it can be set up inside the tools people already use, so every person's AI draws on it without anyone pasting it in. A complete, current record that nobody's AI points at is just another wiki.

Two decisions belong here. The first is access: what is firm-wide, what is limited to one function, what stays with leadership (often parts of long-term direction), and what is client-confidential and stays out altogether. Make those calls before the record is wired in. The second is a way back in. People need an easy route to flag something wrong or missing, or they quietly work around the record and it decays.

Keeping it current is the part that decides whether it lasts

Every document gets an owner, a review date and a trigger. The owner is one named person, not the team. The review date is a fixed cadence, at least quarterly for direction and trends, which go stale fastest. The trigger is an event that forces an update outside the cadence: a new offer, a process change, a lost-deal pattern, something a new hire could not find.

AI helps here too. It can flag pages past their review date, spot where two documents disagree, and check whether its own output on a live workflow still matches the record. What it cannot do is decide what the firm now believes. When something works better, it is written back, so the record improves with use. A firm learns only as fast as it writes down what it learned.

This is also the real risk. An AI working from an outdated record repeats it confidently, to everyone, every day. Maintenance is not the dull last step. It is what keeps the whole thing honest.

Maintenance is not the dull last step. It is what keeps the whole thing honest.

The people side decides whether any of this holds. Mollick's model pairs leadership direction with a central lab and the crowd of employees experimenting in their own work, and the crowd is what bottom-up adoption already looks like. Without a written context, though, the crowd tends to spread habits rather than knowledge. Mollick also observes that "many workers are hiding their AI use, often for good reason," including the worry that a visible gain simply raises the target. If reclaimed time is quietly turned into a higher quota, people stop sharing what works, so decide in advance where that time goes.

Finally, someone has to own the system as a whole. That is a role rather than necessarily a full-time job, and it is the one most often missing. Without it, the record is a project that ends rather than a system that runs.

The real case is what the firm can do once the context exists

The defensive case is about what the firm might lose. The stronger case is what it can do once the context exists. Five things change, and none of them depends on which AI tool the firm uses.

Fewer judgment calls go back to the founder. In most founder-led firms, a steady stream of small decisions routes back to the founder, because only the founder knows how the firm would decide. With philosophy, terms and boundaries written down, people can check a decision against the record, often with their AI's help, before they escalate it. The founder still makes the calls that matter, and fewer of the routine ones arrive on their desk.

Every improvement reaches everyone at once. Bottom-up gains stay where they started. When the firm's context improves, whether that is a sharper description of the ideal customer or a better version of a process, every person's AI draws on the change the same day. This is the direct answer to individual gains not adding up: the firm improves in one place, and the improvement spreads by default.

Handoffs get cleaner. When sales and delivery draw on the same written description of the client, the offer and the terms, what was promised is more likely to be what gets planned. The client meets the same firm whoever they talk to, which is hard to achieve when each function works from its own memory.

New people become useful faster. Onboarding is often the first benefit founders recognize. A new hire, contractor or fractional executive whose AI works from the firm's context learns how the firm works by doing the work, rather than by finding the person who knows. The firm's knowledge stops depending on who happens to be available that week.

Strategy can be tested against the whole business. When the founder considers a new customer segment, a new market, a new offer or a new product, the question is usually answered against what they remember of the business that day. With the firm's context written down, the same hypothesis can be tested against a fuller picture of how the firm actually works: what it would change, who it would affect, and where it conflicts with how the firm already decides. That does not make the decision for the founder. It makes it a better-informed one.

What this means for you

If your team's AI use is working, you already have the raw material. The next time someone shares a tip, ask where it gets written down so the rest of the firm benefits. Then pick one workflow that crosses two functions, write the firm-wide basics it depends on, give each page an owner, a review date and a trigger, and judge the result by whether a new person's AI could do that work well in their first week.