Consulting firms spent five decades assembling hierarchies of people around each new wave of technology, calling the result digital transformation and billing by the hour. In the AI era, as a Forbes essay by former CIO Lenin Gali observed this week, that same transformation is being turned on its authors. The firms that restructured everyone else are now restructuring themselves first, and the shape of what they are building is remarkably consistent across companies that would normally agree on nothing.
The common unit is small. IBM calls it a Forward Deployed Unit: three or four specialists, someone who owns the business result, an engineer, a data expert, an industry hand, working alongside ten to twenty AI agents that write requirements, generate code, build test cases and hunt for defects. IBM's Javier Olaizola described a data migration for an Indian bank that would once have taken about 300 people twelve to fifteen months, delivered instead in under nine months with a fraction of the team, because the agents absorbed the repetitive conversion and testing. The humans stayed at the ends of the process: defining the problem and judging whether the output was good enough to ship.
That template is spreading fast enough to move markets. Writing in The Business Times, a Vertex Ventures investor traced how, in a single month in mid-2026, OpenAI launched a $4 billion majority-owned deployment subsidiary seeded with embedded engineers, Microsoft committed $2.5 billion and 6,000 experts to a "Frontier" unit, and Amazon put $1 billion into its own forward-deployed engineering group. On the day OpenAI announced its deployment company, shares in Accenture, Cognizant and Infosys fell. The market read it correctly: the model vendors were no longer selling access, they were coming for the implementation business.
Why does the "last mile" suddenly command the margin? Because the bottleneck in enterprise AI was never the model. Deloitte's 2026 survey of more than 3,000 leaders found only a quarter of organisations had moved 40 percent or more of their pilots into production. What stalls the rest is the unglamorous work: mapping workflows built over decades, cleaning scattered data, satisfying regulators. Enterprises already have better models than they can deploy. What they want, the investor argued, is a vendor willing to be accountable for the outcome end to end, and HFS Research finds three-quarters of them ready to renegotiate contracts to get it.
The economics is the real story. Traditional managed services grow headcount in lockstep with revenue, gross margins pinned at 20 to 35 percent. A platform that automates a workflow gets cheaper with every engagement, and margins can climb to 60 or 70 percent as the same system is reused. Palantir has run this playbook for a decade while being dismissed as a consultancy; its US commercial revenue grew 149 percent in the second quarter. The tell that a firm is genuinely building a platform rather than rebranding a body shop is simple: revenue per engineer rising over time, not headcount rising with revenue.
The change reaches all the way to the invoice. McKinsey has committed to shifting about a quarter of its global fees to outcome-based pricing, a tacit admission that selling hours while running delivery on agents means charging for a cost structure you no longer carry. "If you are pricing by the hour," Olaizola put it, "the incentive for anybody is basically to maximize the hours, not the outcome." The open question is what happens to the people who used to be those hours. Firms insist junior staff will still be hired, but developed differently, learning judgment from day one rather than grinding through the routine tasks agents now handle. Whether that promise survives contact with the new margins is the part worth watching.