Claude In The Conference Room, AWS In The Walls, And Procurement Holding The Check

I asked Excel to stop breaking the budget, it called a friend named Claude, and now my PowerPoint is whispering to an AWS server like it is planning a surprise party for procurement, cute, until you realize the cake is billed by the millisecond. Microsoft just put Anthropic’s Claude inside Office next to OpenAI, not because logos look pretty together, because Claude currently does two unsexy jobs better, it lays out slides that do not look like a freshman group project, and it wrangles spreadsheets without setting off the sprinkler system in finance. Internal bake-offs said Claude Sonnet 4 wins the precision chores in Office, so Microsoft is buying outcomes, not vibes, and Claude quietly excelled at the boring bits, which is the part that actually saves money. The part that makes accountants reach for chamomile, Microsoft will pay Anthropic through Amazon’s cloud. Your Windows laptop edits a deck, the request routes to Anthropic on AWS, while Microsoft keeps smiling about its very expensive alignment with OpenAI, calling it strategic and very cost effective, sure, and meanwhile those invoices all meet each other in a triangle with shared wallets and cloud credits. It is messy, rational, and totally about capacity. GPUs are scarce, latency is a tax, and nobody with survival instincts bets the company on a single supplier when the chips sit in buildings they do not own. This is not romance, it is electricity and floor space, and someone else holds the keys at the data center. Timing writes itself. OpenAI gave Microsoft the sprint start, Copilot shipped into everything, sometimes with all the grace of a Roomba meeting a shag rug, and then two currents hit at once, big customers tried the tools and did not see clear productivity gains in at least one government study, and OpenAI started hedging, getting friendly with Google Cloud, teasing a jobs platform that stares at LinkedIn, and eyeing custom chips à la carte. When your star partner diversifies, you diversify faster, not out of spite, because you can read a room, and the room says capacity and control are the only real love languages in enterprise software. So what is actually happening, sans confetti. Office is going multi model, which shifts the contest from who has the prettiest model to who routes the right task to the right model with the least friction and the best result. Office becomes a switchboard more than a soulmate, and the operator owns the relationship. Whoever runs that switchboard holds the feedback exhaust, the telemetry that becomes tomorrow’s features, and next quarter’s sales pitch, parce que bien sûr, and the leverage at renewal time. OpenAI stops being the default, starts being a line item, which is corporate for we can still hang out, just maybe not every weekend. Winners and losers, with receipts, Microsoft gets leverage and resilience, it can tell OpenAI to sharpen pricing because Claude is already wired in, and DeepSeek sits in Azure like a spare tire you hope never touches asphalt. Anthropic gets validation where careers are made or stalled, inside Excel and PowerPoint. Amazon gets paid as the toll taker for a rival’s flagship, which is delicious if you enjoy irony, and they do. OpenAI loses some glow and some negotiating power, not fatal, just gravity reasserting itself. The rest of us maybe get cleaner decks and fewer flaky formulas, while employers funnel more data into systems they barely vetted beyond a showroom demo, that part should make your eyelid twitch. Principle first, then the bruise. People deserve to know when automated systems shape their work, to challenge those systems when they mess up, and not get quietly profiled by telemetry they never agreed to. In practice, once Office normalizes mixing models behind the scenes, every workplace becomes a quiet lab for model behavior, and the feedback loop is invisible to the folks affected. Pricing unchanged is a cute headline, the real cost shows up as lock in, today’s open routing can harden into proprietary flows tomorrow, where your macros, prompts, and custom automations only run well if you stay inside the bundle. Also, when models disagree, whose output wins, and who is responsible when a fabricated number slides into a board packet, the model vendor or the platform that picked it. Someone eats that, and if the answer is the user, congrats, that means you. This is the “then what happens” moment, the quiet disaster that shows up in a quarterly review, not a breach headline. If you run IT, push now for transparent routing policies, auditability that crosses vendors, and a clean exit path, contract language that says you can take prompts, logs, and custom workflows elsewhere without breaking everything, plus a real answer on red-teaming across providers, not glossy PDFs. Ask where data sits, where it travels, which models touch it, how deterministic the routing is, how you override it when you must, and what gets logged so you can reconstruct the play when someone shouts foul. If you are a worker, ask for disclosure when your documents cross model boundaries and a process to challenge outputs that affect pay or evaluation, yes it is a mouthful, yes it matters, and yes, they will squirm, which is how you know you touched the live wire. The hopeful bit, because cynicism is just laziness in a nice jacket. Competition on task quality can make the tools less annoying and more useful, and routing done right can reduce errors, not multiply them. Actual solutions look different from security theater, they are boring and specific and testable, and you can walk away with your data if the deal goes sideways. Our job is to demand the boring safeguards while the market is still fluid, remember that convenience is not consent, and keep our eyes on who runs the switchboard, because that is where the power sits. Think of model routing as plumbing for ideas, invisible until a pipe bursts over finance, then suddenly all you care about is where the valve is. Which brings us to the main event, how this routing game will actually work, what it means for your stack, and how to keep your team from becoming unpaid QA for someone else’s roadmap. --Doc (yeah I said it!) find me anywhere under @DocAtCDI