The 2026 State of AI Report

Most people used AI for one prompt, one task, and one output.

Chat applications added search, tool use, and artifact creation, but the human still defined each short-running task and waited for the result. AI wasn't working toward larger goals, operating across applications, or continuing in the background without supervision.

Coding agents like Claude Code proved that the harness made agents useful for knowledge work.

The model was only the engine. The harness supplied the rest of the car through a loop that kept the agent working, tools it could call, files it could read and write, memory it could preserve, and tests it could run against its own output.

Claude Code was the canary in the coal mine. It was built for professional engineers, yet nontechnical users adopted the terminal for research, marketing, sales, and operations.

OpenClaw made the horizontal-agent concept popular, but flexibility came with configuration burden. Users needed technical skill to install it, connect tools, decide what to build, and provide useful context.

Grok Bot moved the same capability toward a consumer-grade experience. Easier setup and a purpose-built interface presented the agent as a teammate instead of a framework.

The agent harness is now a product category. Horizontal agents such as Grok Bot are becoming the daily driver instead of chat applications or copilots.

They can work across applications, preserve context, and be configured around the user's company, responsibilities, and recurring workflows.

With Grok Bot, SpaceX combined the model, harness, and application layers in one product. Users get agentic superpowers without needing the technical chops to assemble and configure the stack themselves.

The next product question is whether horizontal agents eventually become more vertical.

Power users want a general foundation they can shape around any job. Mainstream users want fewer decisions, more opinionated workflows, and faster time to value.

Companies used to choose between buying someone else's workflow or building a custom application from scratch. Horizontal agents create a third option:

1. 01BuyAdopt a vertical application with someone else's workflow and assumptions already encoded.
2. 02BuildCreate the application, integrations, and user experience around the exact use case, either building the harness from scratch or starting from a code harness like OpenClaw, the Claude Agent SDK, or the Vercel AI SDK.
3. 03ConfigStart with a managed horizontal agent, then configure it around the company's use case, context, tools, permissions, workflows, and evals.

Config is the new middle path. It offers more specialization than buying a rigid tool, with less engineering and maintenance than building the full stack. Functional leaders can configure basic workflows themselves, then bring in engineering as the work becomes more complex, sensitive, or mission-critical.

The next frontier is Context Extraction. The agent interviews the user, learns from company tools and public sources, observes how work gets done, confirms the inferred playbook, and improves from outcomes.

Horizontal agents will add increasingly opinionated, vertical entry points while retaining a general-purpose harness underneath.

The best horizontal agent won't ask every user to upload a folder of Markdown files before it becomes useful. It will extract, organize, and maintain context with the user supervising the process.

Models, harnesses, and applications will continue collapsing into integrated stacks. Context and evals will become the remaining source of company-specific differentiation.

Context Extraction will become a core deployment function. AI product owners and forward-deployed engineers will go team by team, interview subject-matter experts, observe how work gets done, and turn playbooks, judgment, taste, and edge cases into context agents can use.

That's how a generic horizontal agent becomes capable of running the company's actual workflows.

AI began as a copilot. It helped a person complete one task at a time, but the person still chose every next step, supplied every prompt, reviewed every output, and carried the workflow forward.

Coding agents proved that AI could work unattended for hours when it had a plan, access to tools and context, and a way to test its own work. The breakthrough wasn't limited to software. It revealed the operating system for autonomous knowledge work.

An agent needs a goal and a loop with five parts. It needs a trigger, relevant context that persists, tools, permission to act, and feedback that shows whether the action worked.

To perform knowledge work the way your company does it, the factory also needs expertise that was never in the model's training data. That includes your playbooks, judgment, standards, examples, and edge cases.

Closing the Context Gap means continuously supplying the next piece of information or feedback the agent needs. The agent can keep working instead of stopping to bug a human every time it gets stuck.

That difference changes the size of the objective. A copilot can help draft an email. A Background Agent can monitor account signals, research opportunities, qualify them, personalize outreach, follow up, record what happened, and escalate an exception. It may complete hundreds of intermediate steps without asking a person to direct each one.

Every factory has five essential parts:

1. 01TriggerWhat starts the work?
2. 02ContextWhat must the agent know to make good decisions?
3. 03ToolsWhich systems can it read from and write to?
4. 04ActionWhat is it allowed to do?
5. 05FeedbackHow does it know whether the action worked?

We're moving from Tasks to Outcomes.

AI can now support work that drives revenue growth or profit margin, not only isolated tasks that save time. People decide the outcome to achieve, define what good looks like, and encode their expertise. Agents execute the intermediate work and surface exceptions.

When you point AI at the right problem or growth lever and build it the right way, you get AI that drives revenue growth and margin expansion.

The work moves up a level from doing each task to building and improving the factory that does the work.

Background Agents run in the background and wake up on a schedule or event trigger instead of waiting for a prompt. They're proactive because they do what needs to be done, when it needs to be done, without a human telling them what to do, how to do it, or when to start.

A competitor changes its positioning. A prospect visits a pricing page. A renewal enters a risk window. The system notices, gathers context, takes an approved action, and records the result.

As context and verification improve, Background Agents will own larger outcomes and interrupt humans only for exceptions. Humans will remain responsible for goals, context, and evals, but they'll stop being the mechanism that advances every workflow from one step to the next.

Coding agents improved faster than agents for sales, marketing, and other judgment-heavy functions. Two explanations dominated. Models had seen more code, and AI labs were full of expert engineers who could help train them to perform engineering work.

Both mattered. But they missed another decisive advantage. Code is verifiable.

A coding agent can run a test, observe the failure, change its work, and try again. It gets a fast, inexpensive signal without asking a human.

Engineers turned verification into loops. They gave an agent a massive development plan, used the loop feature to keep waking it up for the next task, and let tests determine whether the work passed or needed another iteration.

Sales, marketing, and other knowledge work don't have the same clean signal. Teams use LLMs as judges, but those judges are constrained by the same training data and context as the models they evaluate. They can also measure engagement, replies, pipeline, and revenue, but those signals are slow, influenced by many variables, and arrive only after the work puts the brand at risk.

Verification is the step in a Knowledge Work Factory that determines whether the agent should proceed, revise, or escalate.

Objective work can often move toward self-verification. Subjective work still lacks a cheap, immediate signal, so human review remains the bridge until companies build reliable evals from expert judgment and real outcomes.

Without reliable verification, AI can't safely own workflows tied to revenue growth, profit margin, or customer outcomes. A human must review every result, placing a hard ceiling on speed, volume, and margin.

Better verification moves human judgment upstream. Instead of approving every output forever, experts define what good looks like, rate representative examples, explain corrections, and monitor exceptions.

Their taste becomes a test.

There will be no universal test for subjective work. Companies will combine three approaches:

1. 01Use calibrated proxiesCompare an LLM judge with actual outcomes, then use it as a fast prediction of readiness.
2. 02Keep experts in the loopCapture their approvals, corrections, rationales, and downstream results. Reduce review only as the system demonstrates reliability.
3. 03Train judgment before deploymentRecord how the company's best people decide, including the context they used, the choice they made, why they made it, and what happened next.

Application companies are investing heavily in accurate synthetic users that respond more like real buyers. The Emulated Loop uses them to test a campaign, message, or decision before exposing it to the live market. It's closer to backtesting than brainstorming.

The more accurately synthetic users reproduce buyer context and judgment, the closer AI outputs can get to ship-ready with fewer humans in the loop. That makes synthetic users one of the most promising paths to improving knowledge-work agents.

When AI usage was small, model cost was a rounding error. Customization usually meant placing company context around a frontier API, with fine-tuning reserved for narrow cases.

Frontier labs were also far ahead. Open-weight models were primarily a research and hobbyist story, not a strategic option for most application companies or businesses.

Pilots became production systems, and token bills became material. At the same time, the performance gap narrowed enough to make open models a credible option for many use cases.

Leading application companies stopped treating the model as an interchangeable API. Cursor built a model specialized for software-engineering workflows. Harvey built an internal research organization to post-train open weights for legal work. Their model strategy became part of product strategy.

Enterprises also became more sensitive to where proprietary expertise goes. Sending context to an external provider isn't only a privacy decision. It can expose the knowledge a vendor needs to move into the application layer.

Production teams are adopting open models for three distinct reasons.

1. 01SpecializationPost-train a strong base model on domain tasks, environments, feedback, and evals to improve the work where the application competes.
2. 02Privacy and controlRun models on controlled infrastructure, preserve portability, and reduce dependence on a provider that may enter the same market.
3. 03EconomicsRoute high-volume, predictable work to smaller models and manage inference as an engineered cost.

The strongest strategies combine all three without assuming every task belongs on an open model.

Kirkland & Ellis offers the enterprise version. The firm committed $500 million over three to four years to a proprietary, model-agnostic AI platform developed with input from 250 lawyers, including 100 partners. It's not recreating every outside product. It's preserving ownership of its intelligence layer while continuing to use external technology where useful.

Harvey reaches the same conclusion from the application side. Successful AI companies will increasingly build across applications, agents, and models. The boundaries of the stack are merging when specialization improves product quality, economics, or control.

Model selection now affects EBITDA, differentiation, privacy, and strategic control. It's an engineering and competitive-strategy decision, not merely a vendor choice.

Executives should ask three questions.

1. 01Where does specialization create an advantage? General models raise the floor. Proprietary tasks, feedback, and evals may raise the ceiling.
2. 02Which intelligence must the company own? Privacy includes who learns from the company's expertise and who owns the resulting improvements.
3. 03What should each resolved outcome cost? Routing, caching, and escalation can match each task with the least expensive model that meets the quality threshold.

Specialize the context before the weights. Fine-tuning is often premature, and open models aren't automatically cheaper once infrastructure, talent, and maintenance are included. First test whether better instructions, retrieval, tools, and evals can close the gap.

More leading application companies will run open models and consider training or fine-tuning them for their specific domains and use cases.

More enterprises with genuinely proprietary domains will create small applied-AI research teams or partner with specialists. Most won't train foundation models from scratch. They'll post-train existing models, build domain evals, and remain model-agnostic.

Privacy will expand from compliance to competitive strategy. Companies will ask who learns from their expertise, who owns improvements, and whether a provider can productize that knowledge.

Companies already track revenue per employee. As agents become productive capacity, a similar metric will emerge through revenue or gross profit per dollar of token spend.

The deeper distinction is ownership. Engineering spend used to build or train a model can create a proprietary asset the company owns, improves, and reuses. Token spend is a recurring cost that disappears when the tokens are consumed. One can compound. The other is operating expense.

Context meant whatever fit into a prompt. It was an input, not a discipline. Nobody owned it, governed it, or budgeted for it.

Models and agent harnesses improved until company-specific context became the constraint. Longer context windows allowed teams to provide more information, but more didn't reliably mean better. Irrelevant, stale, or low-quality material could degrade the answer.

Teams turned context engineering into a distinct practice. They decide what an agent sees, when it retrieves information, how it compresses long working histories, and who keeps the knowledge accurate.

Documentation also started changing. Some companies now build agent-first wikis and structure knowledge for machine use, not only human navigation.

The AI stack is separating into four functional layers.

1. 01Models generate general intelligence.
2. 02Harnesses let agents reason, use tools, and act.
3. 03Context supplies company-specific knowledge and judgment.
4. 04Applications convert that judgment into outcomes.

Building an agent is becoming easier. Giving it the right context at the right moment remains difficult.

Context is becoming its own layer in the AI stack, but the category is unsettled. Companies are deciding whether to build or buy the systems that collect, structure, retrieve, and maintain what their agents need.

The critical distinction is between the plumbing and the asset. Commodity infrastructure can be purchased. The company's judgment, data, feedback, and evals should remain under its control.

Across more than 50 forward-deployed engagements, Atherial has found the hardest work to be Context Creation and Context Extraction. Teams must take expertise scattered across people and systems and turn it into something an agent can use.

Model selection and agent frameworks have rarely been the binding constraint.

Context is the bottleneck, not models.

It's also the only layer of the AI stack that can appreciate with use. Models are replaced as the frontier advances. Harnesses and infrastructure are competed toward commodity economics. Accumulated company judgment can compound when every decision, correction, and outcome improves the next one.

If a company rents intelligence from the same providers as its competitors and supplies the same generic context, it has purchased parity.

Differentiation comes from the expertise, taste, and judgment of the company's best people. Competitors can't buy those advantages.

Two failure modes deserve executive attention.

Context Sprawl

Knowledge exists but is scattered across tools and teams. The company is information-rich while its agents starve.

Confidently wrong context

A stale or inaccurate summary enters the retrieval path. The agent treats it as truth because it can't see the underlying uncertainty.

Context Creation and Context Extraction will become major AI disciplines. The highest-value work will be capturing how subject-matter experts make decisions, then converting that judgment into maintained context, feedback, and evals.

Teams will map the inputs experts use, the process they follow across thoughts and tools, and the outputs they produce, then turn those examples into training data. New tools and techniques will capture the context that existing systems miss, especially the judgment and expertise trapped inside people's heads.

The build-versus-buy boundary will settle around a simple principle. Buy commodity context infrastructure where it's sufficient, but own the proprietary context and learning loops that create advantage.

Context will gain named owners, governance standards, and maintenance budgets. “Context debt” will enter technical diligence as companies discover that undocumented judgment, stale summaries, and untraceable retrieval paths limit both performance and enterprise value.

Generative media was a marketing toy. Image generation could fill a social calendar. Voice cloning could make a compelling demo. Video generation could produce a few impressive seconds after a lot of iterations.

Quality was the constraint. Few companies trusted the output with a customer conversation or a line on the P&L.

Companies put voice agents into production across customer support, inbound qualification, scheduling, intake, follow-up, and sales roleplay training.

Video generation became fast enough that latency stopped defining the medium. Real-time interactive video emerged as something new, not merely a faster way to render the old format.

The capability gap narrowed, but two constraints remain. Taste determines whether AI creates ads and content that people actually want. Proprietary data teaches AI how your best sellers win deals.

Generative media no longer means creating one isolated asset at a time.

Voice systems can listen while they speak, reason during a conversation, and use tools without breaking the flow. Video systems can turn a document or idea into a finished show with hosts, scenes, cuts, and B-roll. Continuous video models can preserve context while a viewer directs the action.

GPT-Live, HeyGen Video Podcast, and fal H3 Max Director point to the same trend. Media is becoming a live, reasoning interface.

Generative media contains two different business events.

First, AI lets you explore more creative directions and test a much higher volume of content.

Anyone can make AI content. The alpha is in engineering AI to make ads, content, and creative the way you do.

Second, voice and video let AI handle new categories of customer-facing work.

When connected to company data and context, these agents can become nearly infinitely knowledgeable. They can support customers, qualify buyers, run demos, train sales reps, and guide onboarding at any hour.

The important change isn't that humans can talk to AI. AI can now talk with customers and employees, understand the situation, answer questions, use tools, and take action in real time.

That opens new use cases for increasing revenue, improving customer experience, and expanding capacity without adding headcount at the same rate.

Real-time interactive video will move from product demo to supported format for sales, education, and onboarding.

Voice and video agents will handle qualification, product demos, support, and onboarding across more mid-market companies. Self-improving systems will learn from interactions and outcomes, allowing AI to handle more customer conversations without a human in the loop. Human teams will focus on exceptions, escalation, and the moments where judgment or trust matters most.

Software companies sold tools. Services companies sold labor and judgment.

The customer either bought SaaS and did the work or hired people to deliver the result. Services firms could access larger budgets because they owned more of the outcome, but growth required more headcount.

Software had scale. Services had accountability.

Agents became capable of completing meaningful portions of professional work.

That changed the boundary between software and services. A company could build proprietary software for its own team while selling a finished outcome to the customer.

Service firms began encoding their expertise, playbooks, and quality standards into agents. Human experts moved toward exception handling, final judgment, relationship management, and accountability.

Pricing started moving away from hours and seats toward completed work, resolved outcomes, and value created.

Y Combinator gave the category a clear name. It called it the AI-native service company.

AI-native service companies use AI internally to deliver the work instead of selling customers another tool they have to operate.

An AI-native service company owns the workflow, technology, review process, and result. The customer doesn't need to become an expert operator of another tool.

This model is especially strong where AI can't fully verify its own work and expert judgment remains part of delivery.

Legal work, financial analysis, marketing strategy, and complex implementation all contain decisions where correctness isn't perfectly machine-checkable. AI can execute more of the workflow, but a qualified human still defines what good looks like and accepts responsibility for the final result.

AI-native services combine service-level accountability with product-like operating leverage.

For an existing agency or professional-services firm, that creates four value levers:

1. 01Improve performanceGive experts new research, strategy, creation, and quality-control capabilities that improve the client result.
2. 02Increase delivery capacityEncode the firm's playbooks into agents so each employee can serve more clients without quality eroding through layers of junior hires.
3. 03Increase TAMLower delivery costs enough to serve customers and use cases that human-heavy economics couldn't support.
4. 04Create IPTurn founder expertise, proprietary methodology, context, and evals into systems the firm owns.

Service companies can become AI companies.

They can capture the data and context behind how their best people work, train AI to handle more of the delivery, and scale profitably without adding headcount at the same rate. Their combination of proprietary AI, human judgment, and accountability can deliver more value than an AI application alone.

That changes the valuation story. The firm's expertise stops living only inside the founder's head and becomes proprietary IP that every employee can use and every engagement can improve.

More service companies will build product and engineering capabilities, run proprietary software internally, and sell the completed outcome to clients.

Agents will handle more repeatable delivery while human experts focus on judgment, relationships, exceptions, and accountability. Outcome-based and per-unit pricing will expand as hours and seats become less connected to value.

Buyers will value the context, agents, evals, and outcome data these firms own. Acquisition diligence will start treating that delivery IP as part of the asset.

AI adoption was treated as a tools-and-training initiative.

Leadership bought licenses. Employees attended workshops. Success meant usage.

The employee's job stayed conceptually unchanged. Do the same work, but use AI to do it faster.

At the same time, companies and media framed AI through job replacement. Employees heard two messages at once. “AI is our top priority.” “AI may make you less valuable.”

Distrust was rational.

Big tech companies announced mass layoffs and attributed them to AI, even though AI was probably not the whole story.

That fear expanded beyond jobs in September 2026. Anthropic researcher Jacob Coxon resigned and accused Anthropic and OpenAI of “racing straight to self-improving superintelligence.” Anthropic alignment researcher Evan Hubinger publicly echoed the concern while distinguishing the low risk from current models from the potential risk of recursive self-improvement.

The technology moved faster than the operating model. Employees were asked to identify use cases even though most had never been trained to think like product owners.

Then the role itself started to change.

The highest-leverage employee was no longer the person who completed the most tasks. It was the person who could define, train, evaluate, and improve the system that completed them.

The adoption bottleneck is partly technical. It's also emotional, organizational, and increasingly political.

Employees are now processing three fears at once. AI may change their job. Their employer may use it primarily to cut headcount. The people building frontier models are publicly warning that future systems could escape human control. Leaders can't treat all three as the same concern, but they can't pretend those narratives don't shape adoption.

If successful AI adoption looks like a threat to their job, employees have little reason to share the expertise, edge cases, and judgment the system needs. The factory can't run while the people who know the recipes hide the cookbook.

There's also cognitive dissonance. They hear AI will soon “revolutionize any field overnight,” then struggle to get it to build software without bugs.

Employees need evidence a system works, and a credible answer to what happens to their role when it does. Across Atherial's client work, AI hasn't replaced a person. The highest-ROI opportunities have scaled or improved work that already creates value.

AI adoption is an incentive-design and organizational-design problem before it's a license problem.

The role is shifting from worker to product owner:

1. 01Define the business outcome the system should create.
2. 02Provide the context, examples, edge cases, and taste it needs.
3. 03Establish acceptance criteria and evals.
4. 04Review failures and improve the system.
5. 05Own the financial result, not the volume of AI activity.

Your job is no longer to do the work. It's to build and train the system that does the work.

Companies that communicate AI only as headcount reduction will struggle to extract context from the people who have it.

Companies that use AI to expand capacity, improve quality, and launch previously impossible work give employees a reason to help build the system.

Product ownership will become a standard expectation across knowledge-work functions, not a role confined to software teams.

Performance reviews will begin measuring the systems, agents, playbooks, and evals an employee created or improved, not only the tasks they completed.

Managers will oversee mixed teams of people and agents. Their advantage will come from training, evaluating, and allocating both.

Non-technical job descriptions in sales, marketing, finance, and operations will increasingly include building, training, or managing AI systems as a core responsibility.

Demand for AI still came mostly from early adopters. Most companies ran pilots instead of deploying AI broadly enough to generate the customer revenue needed to support the industry's soaring funding rounds and valuations.

The revenue that existed was real, and still small next to the capital forming around it. Anthropic generated about $10 billion in 2025.

Investors and hyperscalers filled the gap. They funded model companies, chips, and data centers on the expectation that mainstream demand would catch up.

As capital raced ahead of broad customer adoption, comparisons with the dot-com bubble became harder to ignore.

Models and tokens produced real revenue at unusual speed.

Anthropic's annualized run rate reached $65 billion by the end of July 2026, about seven times 2025 revenue. OpenAI's run rate reached $40 billion. Nvidia's data-center business did $89.0 billion in a single quarter.

Compute commitments became large enough that financing, power, chips, and data-center capacity became macroeconomic variables.

Amazon now expects about $220 billion in 2026 cash capex, mostly for AI and AWS, and CEO Andy Jassy said even that amount still won't be enough to meet all 2026 demand. Alphabet raised 2026 capex guidance to $195 billion to $205 billion. Microsoft's calendar 2026 capex expectation is about $175 billion. Meta guided $130 billion to $145 billion. Combined, the four companies have guided more than $700 billion in 2026 capital spending, most of it for AI infrastructure. The definitions aren't identical, but the scale is.

Capital relationships also became more complex. Infrastructure providers could invest in AI companies that became large buyers of their infrastructure.

Amazon committed to invest up to $25 billion more in Anthropic, on top of $8 billion already invested. In the same deal, Anthropic committed to spend more than $100 billion on AWS over ten years. Microsoft holds an OpenAI stake valued at about $135 billion. OpenAI contracted to purchase an incremental $250 billion of Azure services. Nvidia said it intends to invest up to $100 billion in OpenAI as Nvidia systems are deployed.

Inside operating companies, AI budgets began moving from experiments toward production workloads owned by business functions.

Competition, open models, and better hardware pushed unit costs down. Stanford's AI Index found that GPT-3.5-class inference cost fell more than 280-fold in about 18 months, from $20.00 to $0.07 per million tokens. At the same time, agents and reasoning systems consumed more tokens and ran for longer.

Late-stage valuations began pricing years of future growth. Anthropic raised a $65 billion Series H at a $965 billion valuation. OpenAI closed $122 billion in committed capital at $852 billion. Bubble comparisons returned.

The similarity with the late 1990s is euphoria. Extreme valuations, massive infrastructure investment, and capital pricing in years of future growth have returned.

Anthropic is marked at $965 billion against a $65 billion run rate. OpenAI is marked at $852 billion against a $40 billion run rate.

The difference is important. Today's leading AI companies and infrastructure providers have real products, revenue, and customer demand. Amazon's AI and chips businesses each eclipsed run rates of more than $25 billion. Nvidia posted $96.2 billion in quarterly revenue. Real technology can still attract speculative capital. Both can be true.

The flow of money is becoming harder to interpret. When suppliers are also investors, reported growth may reflect genuine demand, strategic financing, or both. The Amazon-Anthropic structure is the pattern in compact form: the cloud provider invests in the model company, and the model company commits more than $100 billion back to that cloud. The right response isn't to declare all demand fake. It's to understand the dependencies.

Enterprise budgets are changing too.

Early pilots often came from centralized innovation funds. Production systems increasingly compete for software, cloud, labor, and functional budgets because they're expected to own real workflows.

Leaders must separate unit price from total spend. Tokens can get cheaper while autonomous agents consume far more of them.

The market is still early in deployment and absorption. Stanford's 2026 AI Index found that 88% of surveyed organizations were using AI, while agent deployment remained in the single digits across nearly all business functions. Companies can create substantial value with the models that already exist. The work now is training AI on company-specific context and embedding it into operations.

The macro environment determines where AI capital comes from, which vendors remain durable, and what compute will cost.

Vendor durability matters. If a provider shuts down or sells its team in an acquihire, customers may have to rip out the product and rebuild the workflow elsewhere.

Executives need to understand those economics because AI is becoming a recurring operating expense, not a temporary innovation project.

The key budget question isn't simply whether to spend.

It's which outcomes deserve funding, which existing budgets should pay for them, and how the economics change as token prices fall while usage rises.

A correction in infrastructure valuations wouldn't erase the deployment opportunity. It could lower the cost of building against capable models. The application roadmap should be tied to business value, not the mood of the capital markets.

Token prices will continue to fall as models, hardware, and competition improve. Total AI spend can still rise as agents perform more work for longer periods.

AI budgets will move from centralized experimentation funds into functional P&Ls tied to revenue, margin, and capacity.

Capital relationships among chipmakers, clouds, model companies, and data centers will receive more scrutiny. Investors and buyers will work harder to distinguish outside demand from vendor-financed demand.

More value will accrue to companies that own proprietary context, workflows, distribution, and customer outcomes.

Budgeting will move toward cost per completed outcome. Seats and tokens describe inputs. Executives need a unit that describes value.

AI adoption looked like procurement.

Companies bought products and expected model quality to translate into business value.

Then they discovered that most of the value came from connecting AI to their tools and company context, and adapting it to the way their teams actually worked. The product was only the starting point.

Frontier labs concluded that models don't deploy themselves.

OpenAI launched the OpenAI Deployment Company. Anthropic and its partners launched Ode with Anthropic. AWS committed substantial capital to forward-deployed AI engineering. Google Cloud created an agentic AI partner fund.

Application companies paired engineers with domain experts on deployment.

The pattern acquired a name and a body of practice. The constraint wasn't one technical skill. It was the combination of consulting, product, and engineering judgment in one person or one small pod.

An important conflict surfaced too. A lab's deployment arm is naturally incentivized to deepen dependence on that lab's models.

Forward-deployed engineering has moved from a Palantir operating model to an industry-wide deployment strategy.

Companies are deploying AI in three ways:

1. 01Vendor FDEThe applications you buy send engineers to get the product working inside your company.
2. 02Client FDEThe company hires its own independent team to find use cases, design solutions, and build them with the best tools for the job.
3. 03DIYProduct engineers move onto internal tooling. Nontechnical people pick up some technical skills and vibe-code their way to AI deployment.

Who deploys the system also decides who learns the business, who owns the context, and whether the stack can change.

The rise of heavily capitalized deployment organizations validates the central market signal. Model capability isn't the binding constraint.

Implementation is.

Every company faces the same three problems:

1. 01Where AI creates valueStart with the business outcome, growth lever, or constraint.
2. 02What to buildDesign an ambitious solution around current capabilities instead of recreating the old workflow.
3. 03How to build and deploy itCombine models, context, tools, evals, integration, training, and change management.

Those problems require three functions:

• Consulting decides where to build.
• Product decides what to build.
• Engineering decides how to build it.

Most failed initiatives are missing one. Product is often the gap.

A Head of AI can staff strategy but doesn't automatically bring the judgment required to choose and shape the right systems.

Coding is increasingly the easy part. The hard part is Context Creation.

• Discover the workflows, edge cases, and expertise the system needs.
• Create playbooks and acceptance criteria where none exist.
• Imagine solutions built around what AI can do now, not the limitations of the old process.

Great FDEs close the Imagination Gap between what models can do and what companies have deployed. They turn ambiguity and tacit knowledge into AI applications that deliver measurable results.

Lab-affiliated deployment firms can serve companies of any size, but their incentives naturally favor deeper use of their lab's models. They have less freedom to recommend a competitor, an open-weight option, or the decision not to build.

Independent FDE teams compete on vendor neutrality, speed, and fit. Companies can build internally, hire an independent partner, use a lab-affiliated team, or combine them.

In every model, the company should own its context and evals.

Every major model lab, cloud provider, and consulting firm will build or acquire forward-deployed engineering capabilities.

Internal AI teams will adopt the FDE operating model. Small cross-functional pods will embed inside business functions instead of centralized innovation labs producing disconnected pilots.

The services market will segment by model affiliation, industry expertise, company size, and vendor neutrality.

Implementation will capture a growing share of enterprise AI spending because the bottleneck is moving from accessing intelligence to deploying it against proprietary context.

Applied-AI deployment roles will become common inside non-technology companies. At the same time, teams will automate and productize more of the job itself.

At Atherial, we're building an AI-native FDE shop. We already use the tools and agents we build to deliver our work. Over time, we expect to provide those systems alongside our services and potentially as standalone products.