The State of AI for Operators

Palantir popularized FDE: send your engineers to live inside the customer, learn the domain, ship the solutions hand-in-hand. It was needed because Palantir’s platform was opaque without a human translator. Now Anthropic, OpenAI and Google are doing the same thing, for the same reason. The models are powerful, but the affordances aren’t there, solutions don’t install themselves.

As someone who ran a solutions engineering team, this is all very validating. As an AI consulting firm we could see it as a threat, but we also think there’s plenty of pie.

• Applied AI teams, Solutions Architects, AI consulting wings are all variants of the same playbook.
• Joint ventures are forming up with Blackstone & Goldman Sachs getting hands-on with Anthropic, and more private equity getting in on the game with TPG & Advent contributing to “The OpenAI Deployment Company”.
• What about the big 4, I hear you ask? Well, there’s ‘partnerships’ that have been announced over the year with Deloitte, McKinsey, BCG et al. Basically AI is hard enought to figure out, they’ll grab anyone’s hand. We’re here if you need us chaps: hello@intentional.team
• The job is essentially translation. What does this business actually do? Which workflows have effort worth capturing? Then an army of fresh compsci grads looking to shave some yaks.
• For many orgs, this is going to be required. You have data. You have appetite. You have AI budget. What you don’t have (or can’t afford or entice) is the engineer who can bridge between the GTM team and a functioning agentic workflow.

It’s not going to be cheap, and as with most scaled consulting models, it’s going to look like a claws-in model. We don’t imagine there will be much discussion of model-agnosticism, token optimization, or avoiding vendor lock-in.

For at least a decade “design system” has meant a Figma library nobody enforced and a Storybook nobody read. While everyone else has been off at AI sleepaway camp, design has been staying at home, retaking its credits in summer school.

Design folks we speak to have, not unreasonably, been waiting on Figma to get their shit together. Unlike product (words!) and engineering (<words!>), the visual/creative component of design isn’t well served even by the multimodal capabilites. The delta between a nano-banana output and real graphic design remains pretty large. But hope is at hand!

• Claude Design and equivalents are landing. Foundation labs are productizing design tooling that can consume myriad inputs, and output something on-brand, on-grid, and ship-ready. But it’s pretty pricey. Holy token consumption batman.
• Figma is finally shipping useful things. The “design as final artifact” era is over. Make-systems are converging with build-systems and Figma’s finally getting itself onto the right side of that line.
• Design systems become AI lego bricks. Tokens, components, layout grammar are what the model needs to do useful work. Treat them as instructions for an agent, not a style bible for humans.

Leave it to Anthropic (who else!) to stir things up with Claude Design. The interface is ironically terrible, the affordances weak, and the workflow opaque, but when it works it’s pretty sweet. The productivity gains of a central digital system are starting to show up in a way that doesn’t required the type of tasteful UI engineers that were the preserve of the few. Models can now consume design tokens, component libraries, and layout grammars, and produce production-faithful output, without it all looking like purple-gradient-Claude-slop.

The design system isn’t documentation any more, it’s a tool.^†

Anthropic and OpenAI are heading toward IPOs in 26–27. Profitable companies IPO well. Unprofitable ones don’t. The next few months are going to feature increasingly creative ways for them to extract revenue without renegotiating the contract.

• Enterprise token allocations are going away. Renewals are landing as metered-only. The subsidized seats model is dying for Enterprise, will SMB follow?
• Uber spent their annual AI budget in 4 months. Token-based pricing is hard to track, harder to control. Your CFO is panicking
• The tokenizer changed. Claude 4.7’s tokenizer counts ~27% more tokens than 4.6 for identical content. Yay invisible price-hikes.
• Enterprise security is now paywalled. Want domain capture, audit logs, RBAC, Compliance API? Welcome to the metered tier.
• AI FinOps is about to be a function. Someone in your org needs to own the meter. If it’s nobody, it’s your CFO’s surprise in Q3.

It’s not all doom & gloom. Open-weights models are trailing the frontier by 12-18 months in capability. Good enough for many tasks, trending to good enough for agentic coding. Combine OpenCode and DeepSeek-V4-Pro and you have a fallback for when you hit your Claude limits.

While we don’t know the true cost of inference at Anthropic et al, the cost of raw compute is visible through inference marketplaces like Openrouter. Data suggests there’s a hefty markup coming from the closed shops. For fixed capability, inference costs are dropping, maybe even ~5-10x annually due to techniques like sparse MOE, specialized hardware, etc.

But we’ve still not hit the goldilocks zone yet, the ‘good enough’ cost:performance, and the goalposts keep moving. Tool usage, Reasoning, these all increase demands on inference, with the tradeoff of increased utility.