AI is changing work by moving people out of production and into supervision. The parts of your job that happen on a computer (writing the draft, building the deck, pulling the report, formatting the data) are increasingly handled by a model. The parts that stay human are accountability, relationships, and judgment.
At TJ Digital, we run AI through every workflow across roughly 42 client campaigns. That’s what lets us deliver about four times the work at the same rates a traditional agency charges.
The models are already good enough to do most knowledge work. What they’re missing is the information about your business that would let them do it without you standing over their shoulder.
So the useful question right now is what you should be collecting and organizing, starting today.
What Work Will Humans Still Do as AI Gets Better?
Current frontier models can already do essentially anything a person can do on a computer. Human experts are still better at plenty of specific tasks, and that gap keeps closing.
I think there are three ways humans keep adding real value for the foreseeable future.
Why Accountability Stays Human
Someone has to hold authority, accept risk, and answer for the outcome. A model can produce the recommendation and even execute the action. The responsibility still sits with a person or a legal entity.
Regulation makes this explicit in some cases. The EU AI Act requires deployers of high-risk systems to assign human oversight to natural persons with the necessary competence, training, and authority.
That requirement is tied to one regulatory category, so don’t read it as “every AI decision legally requires a human.” The broader point holds anyway. Automating the work does not automate away the responsibility for it.
Why Human Relationships Get More Valuable
Clients and colleagues still want a person involved in the things that matter to them. Stanford researchers surveyed 1,500 US workers across 104 occupations and found real discomfort about handing client and vendor communication to AI. 45.2% wanted an equal human and AI partnership, and another 35.6% wanted human oversight at critical points.
There’s real counter-evidence worth knowing about. A field experiment with 776 Procter & Gamble professionals found that individuals working with AI matched the performance of traditional teams and crossed functional expertise boundaries more easily.
AI can clearly handle social interaction. Human relationships carry things beyond information exchange: mutual obligation, reputation, trust built over years, and commitments someone can be held to.
Where Human Judgment Still Beats AI
AI judgment is already better than human judgment in some areas. But we have a perspective from living in the real world as a human that I think will be hard for a model to replicate for a while.
The research on this is sharper than “humans reason better.” A study of 758 BCG consultants found that on tasks inside the model’s capability range, AI users completed 12.2% more subtasks, worked about 25% faster, and scored about 32% higher on quality.
On a problem deliberately built to sit outside that range, the group without AI got it right 84.5% of the time. The AI-assisted groups got it right 70.6% and 60% of the time. The model handled the obvious numbers, missed a signal buried in interview data, and people went along with the polished wrong answer.
Knowing when to distrust a confident AI output is becoming a professional skill on its own.
| What humans keep doing | Why it stays valuable |
| Accountability | Someone has to approve consequential actions and answer for the result. Regulation sometimes requires a named person. |
| Relationships | Clients prefer human involvement in decisions that affect them, and commitments need someone who can be held to them. |
| Judgment | Model capability is uneven. Deciding what the real objective is, spotting the exception, and catching a wrong answer still falls to people. |
There’s also the fact that we have bodies and can move things around in the physical world. Robotics has further to go, so I’m focused on what happens on a computer. That’s where the change is happening right now.
What Will Day-to-Day Work Look Like at an AI-Driven Company?
My picture of my own company one to two years out is that we’re mostly having conversations. We review work an AI did, we talk to the AI about it, and we talk to each other while the AI listens in.
Eventually I want the AI participating in those conversations, keeping us on track and pushing us to actually make a decision. From that conversation, it would have everything it needs to go do the work. Things would just happen.
That’s closer than it sounds. In one human and AI collaboration experiment, 2,234 participants produced more than 11,000 real ads. Human and AI teams made about 50% more ads per worker, exchanged 25% more task-focused messages, delegated 17% more work to the AI, and performed 62% fewer direct text edits.
People stopped typing the artifact and started specifying, critiquing, and delegating it.
The caveat is that individual workflows change faster than organizations do. A six-month randomized study across 66 firms and 7,137 workers found that people who adopted AI saved roughly two hours of email per week. Meetings and coordination barely moved, because those require other people to change behavior at the same time.
Giving your team AI does not redesign your approval chains or your decision rights. You have to do that part on purpose.
Why Doesn’t AI Work Autonomously Inside My Company?
At this point you might be saying that this won’t work because AI doesn’t have the context it needs to operate inside your organization. That’s exactly right, and it’s the whole point.
A general-purpose model does not know which revenue definition your finance team treats as authoritative. It doesn’t know which customer is on a special contract, which policy was replaced last Tuesday, or who is allowed to approve a refund over a certain amount. All of that is context that lives inside your company.
The original retrieval-augmented generation research framed part of this as the difference between what a model stores in its weights and what it retrieves from an external source at the time of the request. Businesses need the second kind, and they need it organized.
Giving an agent access to all your documents isn’t enough either. Ask it which enterprise customers are at risk and what to do about them, and it needs to know what your company means by “enterprise,” which fields hold the authoritative numbers, which renewal playbook applies, and which recommendations require approval. Miss those and you get work that looks convincing and is wrong.
Google reports that Bloomberg Media saw 63% better SQL accuracy from its agent after unifying metadata and business context. That’s a vendor-reported number, so treat it as an implementation example rather than a general estimate. The direction is the useful part.
The model brings general intelligence. Your company has to bring local reality.
What Three Types of Information Should You Collect for AI?
The only missing piece is that context. So the thing to be working on right now is a system for collecting, maintaining, and organizing it. I break it into three types.
| Type of information | What goes in it | Question it answers | Where it lives |
| Knowledge base | Approved definitions, policies, product facts, current decisions | What is true and authoritative right now? | An LLM wiki, ideally in a portable format |
| Collections and logs | Email, chat threads, tickets, call notes, CRM events | What happened, and what was said? | SQL, a vector database, or platforms like Gmail and Slack |
| Processes | SOPs, checklists, decision trees, templates | How do we do this task? | Skills, versioned like code |
How Do You Set Up an LLM Wiki?
This is the canonical version of each piece of information. One concept per document, curated, opinionated, and small enough that a person will actually maintain it.
We use Google’s Open Knowledge Format for this. It’s a directory of Markdown files with structured metadata at the top of each one. That means it lives in version control and any agent can read it without a proprietary runtime.
If you looked at this earlier in the year, the spec has moved. OKF v0.2 added trust signals in July 2026, including provenance, verification, freshness, a status field that runs from draft to stable to deprecated, and a date after which a concept should be treated as stale.
That matters more than it sounds. An agent can then prefer a verified current definition over something stale or generated.
What to keep out of the wiki: every Slack message, email, and support ticket. That recreates the mess the wiki exists to solve.
Where Should You Store Emails, Chats, and Logs?
The second type is the large, messy pile. Emails, chat histories, tickets, transcripts, system events.
This goes in a SQL database, a vector database, or stays in the platform it already lives in, and you retrieve it through search or RAG. SQL is better when the question has precise constraints, like all open enterprise deals in EMEA renewing in 60 days. Vector search is better when the relationship is about meaning, like finding complaints similar to the one that just came in.
Most real business questions need both.
Two rules I’d give anyone building this. Keep the source record, its timestamp, and its permissions available separately, because embeddings are a derived index and may need rebuilding. And carry permissions through retrieval, so indexing a document never quietly makes private HR files readable by the whole company.
Don’t let an old chat message override current policy just because its wording happens to match the question. Raw history tells you what happened. The wiki tells you what the company currently believes.
How Should You Store Your Processes for AI?
The third type is instructions, or descriptions of how you do things. These are best stored as skills.
A skill is a folder with a required instruction file plus any scripts, reference documents, and templates the process needs. Anthropic released it as an open standard in December 2025, and the specification is public, so the same procedural package can move between compatible agents.
The design principle is worth stealing even if you never adopt the standard. Store processes as modular, discoverable, versioned packages instead of one enormous prompt. The agent loads the full instructions only when a task matches, so your entire operations manual isn’t sitting in every context window.
For anything consequential, a good skill covers more than the steps:
- When the process applies and what has to be true before it starts
- Which tools, systems, or APIs it may use
- The normal sequence, and what changes it
- How to tell the work succeeded
- What runs automatically and what needs approval
- When to stop and escalate to a person
Writing “never approve over $10,000” in an SOP tells the model what you want. If an action must be blocked, block it in the permission layer.
How Should a Company Start Preparing for AI?
Pick the process your best employee runs that nobody has ever written down. Write it down. That’s the whole first step.
In most companies, nobody has ever clearly documented how the good employees actually do the work. That’s the real limit on what an agent can do for you, more than model capability.
We do a version of this for every client through what we call a Brand Ambassador, an AI project that holds everything there is to know about a business and how it talks. Building it and maintaining it is most of the value we deliver, and it’s a big part of how we operate. The same logic applies inside any company trying to get real work out of AI.
Will AI Replace Knowledge Workers?
Not on the current evidence. The share of time spent producing routine work is falling, while the share spent specifying, reviewing, coordinating, and handling exceptions is rising.
The BCG study is a good reminder of what happens when the human review step disappears. Skilled professionals got the hard case wrong more often with AI than without it.
Do You Need a Vector Database?
Only if your questions are about meaning rather than exact values. If your team mostly asks questions that map to fields and filters, a relational database and good search will take you further than embeddings will.
Most companies end up running both, with structured filters narrowing the set and semantic search finding the relevant material inside it.
What Should You Document First?
Start with the definitions people argue about. Your metric definitions, your customer segments, and your policies.
Those are the things an agent will quietly get wrong, and they’re the cheapest to fix. Write each one as its own short document with an owner and a date.
The companies that will get the most out of more capable agents are the ones that have made their facts, processes, and decision rights readable by both machines and people. That work is available to you right now, and most of your competitors haven’t started it.
If you want help getting your brand’s information organized so AI systems can use it, that’s what we do. We’re at capacity at the moment, so join our waiting list and I’ll reach out personally when we have room.