If you're running an AI services firm in 2026, the benchmarks have never looked better and the pipeline has never felt stranger. Every few weeks, another model crosses another threshold - reasoning, coding, math - and the technology press treats each one like a moon landing. Meanwhile, your clients are still six months into a "pilot."

Bret Taylor, who as founder of Sierra AI and Chairman of OpenAI has a closer view of frontier model development than almost anyone, said something on a recent Cheeky Pint conversation with John Collison that deserves more attention than it got:

"I think if we paused model development, we'd still have trillions of dollars of economic value that have yet to be realized. I think if we had a mature applied AI market where the CFO could go buy that agent to onboard new supply chain vendors that just worked, we could actually accelerate that trillions of dollars of economic value."

That framing should recalibrate how AI services firms think about where they sit and what their job to be done really is. 

The interview is well worth listening to in full. Building at the intersection of applied AI and investing, we decided to focus on the main take-aways for AI services firms, adding our humble two cents to Bret's views.

The models are extraordinary. Reading about GPT-o3 outperforming expert humans on graduate-level science questions, or Gemini Ultra surpassing specialists on medical licensing exams, feels like science fiction rendered in a press release. It's genuinely exciting.  But almost entirely irrelevant to whether a mid-sized insurance company gets any value from AI this year.

Taylor is direct about why: "I actually think one of the main things impeding adoption of AI is the lack of existence of all those other companies."

He isn't talking about the Anthropics and OpenAIs. He's talking about the applied layer of AI Enablement: the firms that take model capability and convert it into a working accounts-payable workflow, a compliant claims-handling process or a functional onboarding agent for supply chain vendors. That market, Taylor argues, barely exists at scale.

The insight is worth sitting with. Say a mid-market manufacturing company has 40 core operational processes with meaningful digital surface area: procurement, contract review, supplier onboarding, demand forecasting, compliance reporting. Even if current standard models can handle 70% of the cognitive work in each, converting that capability into a deployed, governed, integrated agent requires expertise the company doesn't have internally and a typical SaaS vendor can’t deliver. Someone has to understand the organisational context, build the harness, configure the goals and guardrails, connect the data systems, and manage the change. That someone is an applied AI service firm. You might know them as an AI implementation or development company too.

About ten years ago, Morgan Stanley published a research paper on “the next $100 billion software company.” It’s been so long that I can’t even find the original paper to link to, but I remember one of its core ideas very clearly: there’s a long tail of businesses with limited user bases that haven’t yet been automated by packaged applications.

If a company could solve all of these at once, it could become that elusive $100 billion business - the opportunity left behind by packaged software is that large. My best guess is that we’ve made some progress in automating this long tail over the past decade, but the opportunity is still very much there and up for grabs.

However, rather than creating a magical, chameleon-like piece of software to solve these “niche” problems, the solution is more likely to be an AI-enabled and highly customized service. In many cases, the competition is still Excel spreadsheets and interns, rather than well-funded, aggressive SaaS giants trying to defend their dominance, creating favorable market dynamics for AI service firms.

Here is where most observers are getting it wrong, and where the real service opportunity sits.

Taylor's diagnosis is precise: "I'm not sure companies are set up to essentially absorb the benefits of AI efficiently right now. We ship our org charts as companies naturally. There's not usually a person responsible for that process."

Most large organisations are structured around departments: legal, finance, procurement, IT, marketing. Each department has a head, a budget, and a mandate to optimise its own function. AI tools are being deployed into those departments the same way: give the legal team a contract redlining tool, give the finance team a forecasting copilot, give HR an onboarding assistant.

But value doesn't live in departments. It lives in processes that cross departments. Taylor uses a specific example: onboarding a new supplier. Legal handles the contract, procurement negotiates the relationship, IT integrates the vendor into systems, a business unit sponsors it. "If you tracked the median amount of time it takes to onboard a new supplier and it was 17 days," he says, "I bet you could make it 17 hours." But the only way to do that is to optimise the end-to-end process – which means someone has to own it horizontally, across all the silos that currently fragment it.

This is a completely novel way to think about and organise a company.

For the past century, the dominant trend in enterprise organisation has been specialisation. We built depth over breadth. A career meant choosing between managing people or becoming a highly specialised individual contributor. In medicine, after 15 to 20 years of subspecialty training, a clinician's reference frame can shrink to a remarkably narrow domain - occasionally at the cost of seeing where their work fits in the full patient journey. Corporate organisations aren't fundamentally different. Optimising the local maximum became a profession.

AI inverts this logic. The atomic unit of AI productivity isn't a role, it's a workflow. The question isn't "how do we make the legal team more efficient?" The question is "which process does legal participate in, and how do we redesign that process around AI?" Those are different questions with different owners and different answers.

What the company of the future looks like. The most forward-thinking firms are already beginning to organise around workflows rather than functions. They encourage employees to build fluency in the entire process, not just their node within it. Career development shifts from "become the best contract lawyer in the firm" toward "become someone who understands the full commercial relationship lifecycle and can design AI-assisted processes across it." The generalist, long sidelined by the specialisation trend, returns as an organisational asset rather than an awkward edge case.

What the service firm serving that company looks like. Two distinct service capabilities become essential. The first is workflow design: the ability to map a company's core operational processes, understand which carry the highest value and the highest AI surface area, and redesign them from first principles rather than patching department tools together. The second is process optimisation: the ongoing work of taking each redesigned process and improving how it functions as AI capability evolves, the data changes, and the business context shifts.

This maps directly to what Carlotta Perez has described as the fundamental dynamic of major technological transitions. In her framework, innovation doesn't happen in a vacuum. It's a tug-of-war between the Techno-Economic system - fast-moving, driven by the logic of new technology - and the Socio-Institutional system: slower-moving structures like organisational design, legal frameworks, and education. Historically, the full productivity gains of a new general-purpose technology only materialise when the institutional layer catches up and reconfigures itself around the new paradigm.

The model capability is the Techno-Economic system moving fast. The departmental org chart, the siloed KPIs, the specialisation-as-career-path - that's the Socio-Institutional system lagging behind. AI services firms that develop frameworks for how to bridge that gap have a unique opportunity to help organisations. But it requires a high level of trust and executive buy-in.

Sierra, Taylor's AI customer experience company, reached $150 million in ARR in eight quarters. It does not charge per token, per seat, or per API call. It charges per resolved case. Escalations to a human agent are free.

Taylor's framing of why this matters goes beyond clever positioning: "Reducing your token utilisation for the same outcomes is your problem, not your customer's. That's a great incentive to drive more efficiencies over time."

The analogy he reaches for is the shift from CPM advertising to cost-per-click. The CPC model didn't just re-price the same thing. It changed what the vendor was accountable for. Publishers could no longer sell impressions and walk away. They had to care whether the ad worked.

The same shift is increasingly expected in SaaS, and, I suspect, AI Enablement services.

But what if you are running an AI consulting or implementation firm? The traditional model charges for time and materials: hours of discovery, weeks of implementation, a fixed project fee, maybe an ongoing retainer. The client carries the outcome risk. If the deployed agent underperforms, the services firm has been paid regardless.

Outcomes-based pricing flips the accountability structure. A workflow design firm that charges a percentage of the reduction in supplier onboarding time has a direct financial interest in making that process actually work – not just in delivering a project on schedule. A process optimisation firm that earns a share of the efficiency gains it unlocks can't coast on a beautiful implementation deck.

There are real complications to this model in a services context. Outcomes require measurement infrastructure that many clients don't have. Attribution is contested: was the improvement from the process redesign, from the model upgrade, or from the change management program delivered by another firm? And the capital structure of most services firms isn't built for deferred revenue tied to performance.

But the direction of travel seems clear. Taylor is explicit that this model "creates a strong incentive for the software company to have skin in the game to just help you navigate that last mile" and that "so many of the problems in the software industry are due to that lack of accountability." If you're positioning your firm as a genuine partner in a client's AI integration, rather than a vendor executing a scope of work, the pricing model is one of the clearest signals of which one you actually are.

It’s hard to argue with anything Taylor says in that interview. Most of the points raised will almost certainly prove true over a 10-year horizon. The real question is: what’s the best next step for people on the ground who are embedding AI-driven processes within real organisations?

The art lies in turning a client’s request for “Microsoft Copilot training” into permission to ask deeper, more challenging questions about their core organisational processes. And the science is in finding people who are not only well-versed in emerging technologies, but also understand the centuries-old organisational dynamics that protect the status quo. We’re going to need more engineers with empathy.

— Daria