AI: The Half-Time Report
What I said in January. What actually happened. What I'm adjusting.
In January, I put a slide in front of my company with a surfer riding a massive wave.
The surfer was us. The wave was AI. My message: this wave is coming for our customers, our business, our society. Existential threat. Existential opportunity. You either drown or you ride it. There is no standing on the beach watching.
The room was energized. The timing felt right. AI labs were rewriting the rules every few weeks. AGI was around the corner. White-collar jobs were going to be eliminated in one to two years. Trillions of dollars flowing from the smartest investors in the world. Every conference, every newsletter, every board conversation carried the same message: move fast or fall behind.
We leaned in.
February and March: enabling our teams. Tools, training, AI-assisted workflows across every function. Goal: internal productivity — get everyone using this before it uses us.
February through May, simultaneously: rearchitecting our products to embed AI capabilities that would create real customer value. A parallel track. H2 rollout target. Full company mobilization.
April: the electricity hit. Teams were visibly more productive. Engineers shipping faster than we’d seen in years. The surfer was upright. It felt like the wave was exactly what I said it would be.
May: the first budget meeting that felt wrong.
API bills. Token costs. Inference at scale. None of it behaves like SaaS seats — it runs while you’re not watching, compounds while you sleep, surprises you on the 15th. We weren’t alone. FinOps Foundation found 73% of enterprises reported AI costs exceeding original projections in 2026. Uber burned through their entire AI coding budget in four months. Expensive company to be in.
June brought more budget surprises. Model limitations I hadn’t fully anticipated. And the quieter disappointment — the gap between what these models can actually do and what the keynotes suggested.
Six months of emotional roller coaster. Not as terrifying as the predictions. Not as exhilarating as the demos. But here’s what that middle place gives you: clarity.
So what actually happened?
1. Quick productivity gains are easy. Sustainable ROI is hard.
Only 18% of professional services firms actually track AI ROI. Firms getting real returns moved to value-based pricing before deploying AI — the business model determines the return, not the tool. Same technology. Completely different economics.
2. The AI budget is a new animal. Most CFOs haven’t met it yet.
Token costs, agent costs, inference at scale — none of it behaves like traditional software spend. One enterprise spent half a billion dollars in a single month after failing to set usage limits. Our May was a smaller version of that lesson. Treat AI spend as a separate P&L line or it will find you.
3. Your model providers are not your partners.
Alex Karp said it publicly last week: enterprise CEOs are privately livid about paying for tokens while AI labs absorb their intellectual property. We’ve lived a quieter version of that frustration. Models we’d built workflows around — proven, affordable, working — started getting deprecated. Migrating to newer, more powerful models means higher costs, revalidating everything we built, and no guarantee the next generation won’t be deprecated in six months. There is no end in sight. For professional service firms, your client data is the alpha. “Controlling your weights is controlling your fate.”
4. The entry-level pipeline is quietly narrowing.
Lawyers and accountants aren’t being replaced. But workers age 22 to 25 in the most AI-exposed roles have seen a 13% employment decline since 2022 — not because they’re fired, but because firms stopped hiring them. Future partners and senior advisors are a shrinking cohort. Nobody’s saying this at conferences yet.
5. Adoption is table stakes. Quality of output is the differentiator.
Thomson Reuters’ 2026 report: 79% of legal professionals use AI, 69% of accounting professionals. Results across the board: “marginal productivity improvements.” Everyone has the tools. Whether what comes out is good enough to put in front of a client — that’s the actual race now.
6. Governance failure is a client trust event.
KPMG quietly withdrew a major client report after discovering AI-generated hallucinations. Gartner warned in May that uniform AI agent governance leads to enterprise failure. Most firms invested in tools. Almost none wrote a policy for what gets reviewed before it leaves the building. Not caution — liability.
7. AI-native competitors are forming around your commodity work.
Not a large firm with a bigger AI budget. A small shop with the same foundation models, lower overhead, and no legacy process to defend — targeting structured, document-heavy, junior-dependent work. They’re already bidding against you.
8. Client relationships and the data from them — that’s the moat.
Every professional services firm now has access to the same Claude, GPT, Gemini. Firms that pull ahead will build proprietary datasets — client feedback, engagement history, relationship signals — that compound over time and can’t be copied by a competitor who just signed up for the same API.
9. The pricing model is breaking.
EY data: AI-augmented audits now complete 35% faster. Clients know this. Firms still billing by the hour for AI-assisted work are heading toward a conversation they don’t want to have. Every PS firm has a billable-hour problem hiding inside a productivity story. It just hasn’t arrived as an invoice dispute yet.
10. AI is raising what clients expect — without raising what they’ll pay.
Clients know you have the tools. Faster delivery, deeper insight, more strategic counsel — that’s the baseline now, not the premium. Firms still delivering at 2022 speed with 2026 pricing find out at renewal. Not in a confrontation. In a silence.
January was: ride the wave at all costs. H2 is: read it first.
Three adjustments. We consolidated our internal AI tools from two to one — simpler to govern, easier to track. We set an upper limit on token spend and have a designated leader watching usage like a hawk. And we shifted the filter for every new AI investment: does this create measurably better outcomes for customers? If not, it doesn’t justify the cost. Every good investment in H1 had a specific problem, a clear owner, and a human review before anything reached a customer.
Not as scary as January suggested. Not as simple as the demos made it look.
The wave is real. I just know more about how to read it now.